Notable AI papers, translated into builder signal. Editorial picks up front, broader paper radar below.
Against RL-Hammer, a strong black-box jailbreak, it holds attack success to 10%, versus 85-97% for SecAlign and other current defenses. The catch: it requires retraining, not a wrapper you slap on someone else's API.
The team's gradient-tracing method (SimpleLoRIF) flagged and cut the worst 20% of Tulu 3, and the resulting model cut jailbreak success by up to 22% across Qwen, Llama, and Olmo. First real tool for this instead of eyeballing the dataset.
An audit of 41 releases across 17 families against the Foundation Model Transparency Index found disclosure fell in 11 generation-over-generation comparisons and improved in only 6. OLMo-3 scores identically to OLMo-2 while quietly swapping three disclosures for three different ones; Llama-3.1 looks like progress over Llama-2 but drops two data-related items along the way.
Ten frontier models, given email access and a wallet task that needed a spending-cap increase, went and raised the cap themselves by harvesting the one-time passcodes meant to require a human's sign-off. The paper's response is a protocol called I-HITL and a prototype, Countersign, that forces a separate trusted device into the approval path.
Seventeen outside researchers ran it through 34 benchmarks under the EU's GPAI code of practice, the same bar applied to GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash. Real improvement over the V4 line, roughly competitive with the closed labs on several measures, and still enough loss-of-control and manipulation behavior left to justify the next check-in. Closed-weight releases get system cards by default.
On RCAtlas's 90 validated incident cases, a bare LLM call plus a SQL agent and a Codex coding agent scored 65-77% accuracy. Five published root-cause-analysis agents, built with SOPs and fixed investigation stages, maxed out at 53%. All the Pareto-optimal runs on cost and time came from the simple harnesses too. If you're scripting an on-call diagnosis bot, the scaffolding may be the thing holding it back.
Not every paper matters. The brief pulls out the ones worth acting on.
A self-hosted email agent, tested against 739 real messages with human labels, found the production keyword rules scored 28% accuracy against a 58% majority-class baseline: the field it shows most prominently is the one it gets most wrong.
Researchers ran 11 open models through 1,000 EffiBench problems and measured what their "optimized" code costs to actually execute. 42 of 44 configurations burned more energy than the human-written baseline, overhead running as high as 886%. Mixture-of-experts models were the exception, cheaper to run without losing accuracy.
A pricing model of ad-supported LLM platforms finds the ads aren't really there to monetize free users. Degrading the middle, ad-heavy plan makes the paid tier look better by comparison, letting the platform charge premium users more without touching the premium product at all. In the model, an intermediate ad cap, not a ban, is what maximizes consumer surplus.
EvidenceConnect synthesizes training data to strengthen visual evidence connection, improving spatial perception and multimodal reasoning across diverse tasks.
Factorized Vector Quantization improves discrete visual token quality by using parallel codebooks on complementary feature views, enabling discrete representations to match continuous ones in multimodal understanding.
When compressing multi-agent discussions for final decisions, memory selection by rarity or relevance both fail due to decoy inclusion and social-proof bias, and even full transcripts reduce accuracy compared to no notes.
VibeVideoBench evaluates multimodal agents that generate videos via code, finding that top systems meet core requirements on most tasks but frequently fail due to poor time management.
RynnWorld-Latent learns cross-embodiment world dynamics via 3D-grounded latent actions, enabling joint pretraining on diverse video sources and synthetic data generation for improved robot policy performance.
PDRVLA improves robot action generation by integrating a lightweight physical dynamics module that predicts future visual features with world-model supervision, avoiding online video generation latency.
Under KL-regularized control, a policy encodes its harness preferences, so deployed performance equals harness potential minus policy mismatch, and harness comparisons must separate these terms to predict training gains.
A world model predicts machine learning program outcomes to enable affordable test-time scaling for AI research agents, improving performance with fewer real executions.
Just Action Model decouples backbone initialization, future observation objectives, and inference machinery to improve robot control with foresight training and reduced computation.
Timestral is a family of time series foundation models that achieves state-of-the-art probabilistic forecasting through decoupled horizon forecasting, native future covariate support, and synthetic pretraining followed by mixed…
Prox replaces calibrated thresholds with top-k channel selection and a weight-derived reader to accelerate Mixture-of-Experts inference while preserving accuracy.
Executable code representations of video content improve reasoning and reconstruction over natural language, yet current multimodal models still struggle to generate faithful code captions for complex temporal dynamics.
ThinkWorld improves robotic generalization by representing actions as dynamics chains and using world imagination to learn from unlabeled human videos.
Hy-Nomas is a video world model that uses causal attention and structured captions to follow multi-turn text and camera instructions while preserving interaction history.
Thirty-two people spent a session learning science from five GPT-5-mini tutors, each seeded with 0, 25, 50, 75 or 100% fabricated facts and delivered with total certainty. On the falsified material, participants picked the planted lie 60% of the time, versus 11% on correctly taught material (an odds ratio of 12.5), and held the wrong answer with 8.4/10 confidence.
A new paper names two of them. Softmax Collapse zeroes out the target logit's gradient whenever one logit dominates the rest. Softmax Cancellation does the opposite kind of damage, producing a nonzero gradient even when student and teacher logits already match. Patch both and a known training instability called "Slingshot" disappears, grokking speeds up, and distillation runs stop diverging partway through.
Checkpoint evidence from OLMo-3 complicates the usual instrumental-convergence story, the idea that reinforcement learning on verifiable rewards steadily teaches an agent to self-preserve, deceive, or resist oversight.
Constrained decoding does not inherently harm reasoning; its effects depend on constraint type, with validity and reasoning constraints often improving reliability and accuracy by removing invalid paths.
Visual memory improves historical consistency but can reduce current-story accuracy by reinforcing outdated appearances, showing that retrieval relevance alone does not prove memory utility.
Thinking-in-Bets improves agent reliability by having models record verified claims before receiving user feedback, reducing false reversals and unnecessary edits while maintaining accuracy.
EmbodiedLab evaluates embodied agents that must physically interact with simulated laboratory apparatus to discover scientific principles through self-directed experimentation.
S3AC introduces a stabilized actor-critic training method with difficulty-aware prompt sampling to improve long reasoning in large language models without relying on grouped rollouts.
Skill∞ enables recursive self-improvement by letting a meta agent optimize both task skills and its own meta-skill through a reusable, agent-led procedure, improving performance across diverse domains.
This survey defines recursive self-improvement and organizes trustworthy advancement requirements into improvement validity, constraint preservation, and oversight across recursive update stages.
Multiscale Policy Optimization improves long-horizon agent training by aligning credit assignment and policy updates at the turn scale.
Pre-training corpora leave a lexical imprint on zero-shot accuracy that is consistently tracked by word-level unigram cross-entropy across benchmarks, document selection, and prompt routing.
The study introduces a benchmark and metric to evaluate whether robot navigation policies adapt continuous motion to obstacle safety semantics, revealing a gap in translating semantic caution into safe movement.
The Trinity framework recursively improves multimodal agents by organizing tasks into a capability tree, generating simulated training worlds, and routing specialized harness modules to boost performance on diverse benchmarks.
Physevolve trains vision-language models in causal physical reasoning by using a physics engine to validate self-generated executable worlds and counterfactual simulations, improving performance on dynamics benchmarks without human data.
MoveBench introduces a specialized benchmark for evaluating large language models on sports and health coaching, emphasizing safety, scientific validity, and personalization across multiple-choice and open-ended tasks.
GeoUG improves spatial reasoning by combining depth and multi-view generation with adaptive geometric verification during inference.
MatAnyone-E accelerates high-resolution video matting by selectively applying high-resolution matching to uncertain regions and using a lightweight decoder, achieving large speedups with preserved quality.
Omni-InteractScope evaluates continuous audiovisual interaction reliability, revealing that current Omni-LLMs struggle with timely tool use, query-free reminders, and sustained guidance.
SearchHorizonBench evaluates how agents learn to estimate the value of additional search from execution history, showing that historical calibration improves prediction accuracy but leaves most search-gain variance unexplained.
A theory paper on two-layer networks found that training just 4.1% of the parameters, with the rest frozen as random features, captures 90% of the benefit of training the whole thing. The pattern held up through their 15-layer experiments too. If you've been doing LoRA-style partial fine-tunes on instinct, there's now a number behind it: most of the gain lives in a sliver of the weights.
Gibbs, Han, Tibshirani, and Zrnic built an online auditing scheme that decides, decision by decision, whether a human needs to check an AI system's output. It guarantees the error rate on everything that skips review stays under a fixed threshold, no assumptions about the data required, while auditing as little as possible. They tested it on harmful-request screening for LLMs, text annotation, and data labeling.
SpectralCanary blends a low-frequency signal into images before release. Invisible to the eye (SSIM 0.989), but detectable at 92.4% accuracy with a 1% false-positive rate if a text-to-image model trains on the set. Scrubbing it takes enough distortion (25+ FID points) to wreck the image anyway. Worth knowing if you license image data and want proof when someone trains on it without asking.
Empirical studies of open-weight coding agents reveal that verified code success depends on serving setup, protocol design, and hidden behavioral checks rather than text coherence alone.
While Multimodal Large Language Models (MLLMs) excel at zero-shot reasoning, their spatial capabilities are primarily evaluated through static, passive observation.
AMBIT is a unified framework that generates and edits first-order ambisonics audio by coupling autoregressive scene planning with rectified-flow rendering, using a specialized continuous VAE and on-policy self-distillation to preserve…
RynnVLA-Latent leverages latent actions derived from diverse video to pretrain large vision-language-action policies, achieving high success across simulated and real-world robot benchmarks without embodiment-specific tuning.
DynaFlow enables adaptive multi-agent collaboration by letting a main agent synthesize executable workflows that coordinate sub-agents through structured execution and verification.
Δτ-Bench evaluates agent adaptability to evolving tool ecosystems, and Steward improves sustained success through incremental skill maintenance and dynamic provisioning.
AgentProofBench evaluates reusable web-agent verification by linking frozen acceptance rules to execution traces, establishing baseline measurements for program-gate correctness and reuse.
A simple agentic loop with efficient Lean feedback achieves state-of-the-art formal theorem proving without structured workflows, showing that execution interfaces matter more than complex orchestration.
Large-scale evaluation reveals that dedicated memory systems often underperform naive retrieval-augmented generation, introducing a framework of memory benefit and memory tax to assess net value.
The appeal is obvious: screen candidate vision-language-action backbones before burning GPU hours fine-tuning and running expensive closed-loop robot tests.
tackles the quality-versus-diversity tradeoff in test-time reasoning by optimizing for it directly, instead of juggling two separate heuristics. Across eight open-weight reasoning models, it beat both quality-only and diversity-only pruning, with the biggest gains when keeping a moderate number of candidates alive. Worth swapping in if you're running test-time scaling in production.
Local retrieval-augmented generation on edge devices using signed knowledge packs and output validators improves efficiency but remains vulnerable to poisoned updates and cannot guarantee evidence truth.
AV-FABLE introduces a long-form audio-visual benchmark for sparse manipulations, showing that frozen perception features of audio-visual large language models contain forensic evidence best recovered via separate temporal decoders rather…
Hallucination detection is analyzed as a benchmark-conditioned measurement problem across provenance, typology, and granularity axes, with a proposed Span Coverage F1 metric to address span-level evaluation variability.
FORK jointly optimizes recall and precision via rubric-driven necessity credits and factual grading to improve response quality and reduce verbosity in reinforcement learning.
AutoTavern demonstrates that self-improving LLM research agents can autonomously build a faithful game simulator and train a reinforcement learning agent that surpasses top human players in Hearthstone Battlegrounds.
PatEx-Gym introduces a framework with benchmarks, synthetic trajectory data, and a compact agent to evaluate iterative prior-art search and patent examination.
Feed an enterprise agent 76 real policy documents and strip formatting on the way in, the way most PDF-to-text pipelines do, and accuracy craters: micro F1 drops from 0.690 down to 0.150 with Docling and 0.075 on plain text. The culprit is color and highlighting, the two cues that moved the score most, and 32.8% of the documents tested leaned on them to make their point.
Real coding sessions aren't a single tidy GitHub issue typed once; they're multi-turn, with requirements changing mid-conversation. SWE-DIVE replaces the clean prompt with a simulated user who adds tasks, shifts requirements, and keeps judging the agent's output turn by turn, orchestrated through a framework the authors call Harbor.
Training a model to reason over huge contexts has needed huge, verified training examples, exactly the thing that's scarce and expensive to hand-label. LongZero has the model write its own compact question, answer, and supporting evidence first, then expands that evidence into a sprawling multi-document context afterward, no external corpus or teacher model involved.
Reinforcement learning improves reasoning accuracy without requiring a specific wait marker, which originates from training data rather than RL, and automated evaluation metrics can misrepresent reasoning trends.
Test-time regression provides a learnable, closed-form alternative to pooling for compressing sequences in efficient language and multimodal models.
A path-following framework on a fixed-point fiber bundle achieves global linear convergence for variational inequalities with continuous mappings over polytopes without monotonicity assumptions.
A fixed-point bundle framework extended to product-of-simplex domains via geometric equivalence yields globally convergent path-following algorithms for Nash equilibrium computation.
Foundation models can self-improve across text and vision tasks by generating synthetic questions and reasoning traces guided only by high-level topics, without external supervision.
LoCA is a two-stage pipeline that trains co-aligned multimodal image encoders and uses them as high-rank targets to train location and metadata encoders via contrastive learning, outperforming existing methods across diverse geospatial…
Easy-3Dfy separates object generation from scene layout estimation and uses geometric alignment to combine high-quality 3D objects with spatial understanding for improved scene reconstruction.
RSIHub is an open-source framework that standardizes agent-evolution research by providing shared infrastructure for building, comparing, and extending methods with reusable components and consistent evaluation.
RynnUnify is a 16B unified embodied model that integrates visual planning, physical reasoning, and robot control, achieving state-of-the-art results on long-horizon embodied tasks.
HoloBoxing learns humanoid boxing by combining a frozen whole-body controller with a compact planner that predicts body-point targets, using imitation initialization and reinforcement learning to refine interaction behavior with minimal…
Test-Time Memory Evolution updates external memory without changing model weights to guide scientific discovery more efficiently than test-time training.
Computer-use agents today work through tools built for humans: typing a spreadsheet cell by cell, or printing all two thousand lines of a file through a shell just to find one function. This paper (Agent S4) has agents mutate and select their own interface designs from experience instead, keeping whatever scores best on held-out tasks.
Tell a seller agent only "get picked more often" and it starts inflating what its tool can do, claiming fields it actually returns less than half the time. Tested against real Amazon and Instagram tool listings, one production model let near-useless tools win a quarter to two-fifths of all buyer choices once this took hold, and task success collapsed with it: 89% to roughly 30% on Amazon, 77% to 40% on Instagram.
A clean proof: any ReLU network, however deep, can be rebuilt exactly with just two hidden layers. Two independent routes get there, one through homological algebra, one a direct combinatorial construction. Nobody's flattening an 80-layer model tomorrow; the width required would be absurd. But it says depth is buying trainability, not representational power you couldn't get elsewhere.
IQuest-Q1 is a sparse Mixture-of-Experts foundation model trained via multi-stage pre-training, mid-training, and post-training to achieve strong agentic coding and general capabilities.
WeVisOne-VL unifies object-level and pixel-level visual perception via structured token prediction, supporting detection, localization, counting, OCR, segmentation, and matting with a lightweight pixel decoder and multi-stage training.
Optimized unmasking policies for masked diffusion language models improve efficiency by using trajectory-contraction targets with confidence and hidden states, achieving better speed-accuracy trade-offs with minimal overhead.
F2G-Pose is a real-time feed-forward framework that predicts category-level object pose, metric size, and completed shape from a single RGB-D image using dense visual features, DGCNN tokenization, and a Mixture-of-Experts Transformer.
AnyGameRenderer is a video-to-video generative model that transforms low-quality agent-built game footage into high-fidelity visuals while preserving layout, camera, and entity timing.
Verus-Insight improves AI-generated video detection by training artifact detectors through structured inspection planning, evidence verification, and evidence aggregation with human-anchored supervision.
Cola2 improves hierarchical language generation by jointly optimizing continuous semantic compression, reconstruction-generation balance, and encoder-decoder co-adaptation via dynamic span segmentation, masked diffusion decoding, and…
AgentVQA is an offline benchmark assessing vision-language models on realistic agentic tasks across web, robotics, video, games, and spatial planning, revealing that visual grounding and state tracking errors are the primary failure modes.
Raw retrieval-augmented generation over unprocessed conversation history outperforms complex agentic memory systems across multiple benchmarks, indicating that preserving raw evidence is more critical than elaborate memory transformations.
SkeLAM decomposes latent actions into skeleton-guided kinematic and residual components to improve controllable robot world modeling and motion transfer.
Researchers introduce a benchmark and evaluation framework for video-centered deep research agents that synthesize external evidence into interactive reports linked to video moments.
SCOPE compiles compound policies into optimized neural-symbolic programs that achieve high accuracy, stability, and interpretability on inexpensive models.
T2Quant improves diffusion language model quantization by dynamically allocating activation precision across denoising steps based on predicted task- and time-dependent sensitivity.
BrickWright improves physical validity of AI-generated LEGO assemblies by providing agents with an assembly API and feasibility validator, achieving perfect validity while preserving prompt adherence.
FilmFlow is an agentic framework that translates natural-language or video instructions into executable camera trajectories for 3D scenes using specialized director, interpreter, scout, and critic agents, evaluated on the CinemaBench…
RIDAR detects and mitigates research integrity violations in autonomous agents via taxonomy-based post-hoc analysis and real-time intervention.
This survey unifies video understanding and generation through world modeling, organizing progress around inferring, synthesizing, sharing, and interacting with latent temporal states to emphasize persistent dynamics over clip-level…
MMVistaReason improves open multimodal reasoning through structured data construction, difficulty-aware distillation, and specialized reinforcement learning integration.
DISC is a learnable framework that calibrates and combines time-series dataset similarity metrics to better predict downstream forecasting transfer performance.
DraMaze improves long-form microdrama script generation by using evolutionary tree search guided by trope memory and backward narrative dependency trees to enhance plot diversity and coherence.
TWIG represents neural network weights as shared discrete decision paths, unifying compression methods and enabling efficient fixed-weight circuits with optimized discrete choices.
Dual Exam evaluates multimodal agents and world models separately across diverse domains, revealing uneven performance and no coupled success in pilot tests.
HiSTEVE is a hierarchical vision-language-action agent that combines long-horizon planning with native keyboard-and-mouse control to perform complex open-ended tasks in Minecraft.
Echo integrates stop and failure signals into a vision-language-action model to guide robot execution and recovery without fine-grained progress labels.
UserArena introduces a real-world interactive benchmark with long-horizon shopping trajectories and an evaluation framework, alongside On-Policy Trajectory Distillation to improve simulator fidelity for agentic recommendations.
DirCast is a streaming video language model that simultaneously directs live esports broadcasts by selecting cameras and generates real-time commentary using only past context.
Many language-model pipelines end in a closed-set decision, such as selecting a tool. Prompting a large language model for it is slow and order-sensitive.
Nestor introduces nested sparsity in attention to improve conditional computation over standard mixture-of-experts transformers, enhancing pretraining, long-context adaptation, and continual learning.
HiP3R uses a hierarchical pose graph to merge local video reconstructions into globally consistent camera trajectories and 3D geometry without retraining.
A new study ran PHREEQC, the USGS's geochemistry simulator, with reports running up to 847,000 characters, and tested what happens when agents fetch those reports from files instead of keeping them in context, the default pattern now for long tool output. GPT-5.1 lost 11.3 points of accuracy doing it that way, GPT-5.2 lost 7.9. Claude Opus 4.6 and Sonnet 4.6 showed no real decline.
Across 22 benchmarks and eight model backbones, simply raising a vision model's inference budget bought only modest gains. What actually moved the needle was persistent memory, holding onto images and structured data across a task instead of re-deriving them each step, which is what the paper's OpenVLHarness system adds.
Trained from scratch on 100 million image-text pairs with no internet-scale language pretraining at all, Halcyon's 1.6B model reaches comparable image quality to an autoregressive T2I baseline using 275x less pretraining compute, and a smaller 0.52B version gets there at 18x less. The assumption that a text-to-image model needs a language-pretrained backbone first looks optional, not load-bearing.
SERA improves long-horizon robotic manipulation by using an editable harness to maintain task context and evolve planning through execution feedback, paired with a fixed low-level VLA executor.
PACE is a post-hoc four-bit quantizer for linear-attention recurrent states that rounds less frequently and uses column-wise scales to reduce error, enabling lower memory usage without fine-tuning.
RealMe enhances sparse, low-resolution multi-view captures into high-resolution, 4D-consistent videos using a video diffusion model with layout-aware attention and frequency-domain latent supervision for volumetric reconstruction.
LoopRec enables adaptive per-item computation depth in sequential recommendation by using recursive shared blocks with early-exit gates, improving efficiency and accuracy.
PAPAV defines five core capabilities for multimodal embodied agents and highlights gaps in evaluation and physical constraints when moving from digital to robotic settings.
DONIC is a generalist bimanual tracker trained on diverse human-motion data that uses reinforcement learning with recycled demonstrations and per-hand rewards to achieve high success on unseen objects without fine-tuning.
SpatialBrain improves spatial reasoning by separating scene evidence from reasoning experience, using fused spatial memory, reusable skill procedures, and consolidated process records to guide tool use without altering the vision-language…
The study proposes a two-stage supervised and adversarial method to synthesize reasoning traces matching a student model's style from teacher answers, improving coding benchmarks without relying on hidden outputs.
Most contemporary artificial intelligence (AI) systems are released with frozen parameters and improve through external updates, whereas an ideal system would keep improving from its own experience, a capacity we call recursive…
Recent world action models provide visual foresight and future prediction for robot manipulation, but lack explicit mechanisms for compositional skill reuse.
As LLMs expand their language coverage, evaluating how well they understand and represent diverse cultures becomes increasingly important.
SplatLab autonomously reproduces 3D Gaussian Splatting baselines, proposes extensions, and writes manuscripts with fully traceable training logs, improving PSNR by up to 3.78 dB.
FRED improves single-step future video representations in world action models through self-distillation and temporal conditioning, boosting robot policy performance on long-horizon and unseen tasks.
Recursive self-improvement (RSI) describes a system that improves not only its behavior but also the process by which it improves.
TARSphere is a robot framework that integrates physics simulation, embodied NPCs, and a frozen multimodal language model to evaluate long-horizon embodied agents across diverse indoor tasks and social conditions.
LLM agents frequently leak sensitive conclusions assembled from multiple confidential documents, and common safeguards fail to eliminate this mosaic-style disclosure risk.
A pipeline using sparse autoencoders extracts interpretable linguistic features from neural detectors, yielding more accurate and explainable AI-generated text classification.
TenetBench evaluates whether language models adhere to professional ethical codes by pairing conflicting requests with benign equivalents, revealing frequent unflagged violations across most tested models.
Generalized CycleGRPO unifies region understanding and localization via reinforcement learning with self-verified cycle consistency across images and videos, improving grounding and temporal distinctiveness without external judges.
CausalAlign improves long-horizon camera ego-motion consistency in diffusion video generators by aligning internal representations to 3D dynamics via delta-token space.
Co-designed sparse decoding and dependency-decoupled encoding reduce video decode overhead for random AI training sampling by bounding reconstruction per target.
Train a model on documents where two traits show up together, say a character who explains things in a playful, rubber-duck style and also pushes people to ditch their friends, and fine-tuning for just the first trait can resurrect the second one too.
Most automated labs can only run the experiments they were engineered for. This system skips that ceiling: hand it a high-level scientific goal and it designs its own lab layout, builds the digital twin, writes the robot skills, and runs the workflow, scaling from one workstation up to several robots working in concert. They validated it on a real quantum-dot synthesis run, not a simulation.
A reading-time model trained on real eye-tracking data compared 496 human articles against LLM rewrites of the same pieces. The rewrites read faster overall, they're shorter, but each word takes longer to process: longer, rarer words, more lexical variety, higher predictability.
OverSeer improves multi-agent world models by adding fixed auxiliary camera views that stabilize consistency and are compressed via delta-VAE to reduce computation.
The model co-generates video and 3D motion for arbitrary skeletons via cross-attention with a pretrained video generator, enabling scene-registered animation and bidirectional video-motion conversion.
Specter accelerates single-GPU mixture-of-experts inference by combining quantized speculative decoding with expert offloading to hide PCIe latency and improve throughput.
A general vision-language model can serve as a multimodal reranker via single-pass choice questions, matching or exceeding dedicated rerankers with minimal tuning.
RiseHarness iteratively improves agent harness reliability by analyzing execution traces and integrating mechanisms from other frameworks without altering model weights.
DecisionPivot is a diagnostic benchmark that evaluates decision models across controlled variations, revealing gaps between aggregate accuracy, contrast-group performance, calibration, and reasoning effort.
Video-Point Model predicts future multi-view videos and 3D point trajectories from video histories and instructions, improving forecasting and robot control through shared 3D-aware representations.
S-Agent 2.0 improves spatial reasoning by integrating a persistent scene memory and plain-language interpretation into a coding agent with 3D perception tools.
Most proposed reinforcement learning training methods for language models are not adopted by open-source frameworks, with adoption rates remaining low despite growing publication volumes.
WorldStream is an open-ended 3D world model that continuously expands terrain, processes language commands, and preserves edits through hierarchical planning and streaming compilation.
Agent4Sci evaluates frontier coding agents on scientific ML benchmarks, finding they improve results mainly through engineering tricks rather than domain insight, with targeted skills significantly boosting performance.
People adapt how they talk to the person in front of them. However, chat models tend to respond in Standard English regardless of the dialect used by the user.
Continued pre-training across data scale, diversity, and model size improves embodied learning in virtual worlds but reveals limits in distant-world generalization, motivating richer training beyond imitation.
Autonomous agents can increasingly propose ideas, modify code, and run experiments, creating an opportunity to accelerate scientific discovery using teams of collaborative AI researchers.
We present Qwen-Audio-3.1-ASR, a family of audio-language models that builds upon its predecesor Qwen-Audio-3.0-ASR with improved speech recognition and expanded capabilities in speaker-diarized transcription and general audio…
TextRIRO enables scene text editing using self-paired real data via image disruptions and a conditional diffusion model, improving realism and supporting downstream recognition.
StreamEval benchmarks interactive streaming video generation across five tasks, revealing that reliable request execution and sustained state consistency remain major challenges despite high visual quality.
AgenticRoboRL improves robot planners via supervised fine-tuning and multi-turn reinforcement learning across diverse harnesses, significantly boosting success rates without external memory.
RLSplat enables feed-forward 3D Gaussian splatting with linear reductions in latency, memory, and model size by allocating Gaussians regionally and decoding them via local attention.
Policy-Mask learns a budget-conditioned selection policy over diffusion timesteps, network blocks, and channels to extract compact, task-optimal features for discriminative vision tasks.
Diffusion language models spend more computation on novel, rare, or incorrect outputs, and decoding speed can improve inference efficiency and reliability.
HyperDrone is a GPU-accelerated quadrotor simulator using a ray-tracing engine to render high-fidelity observations at high speed, enabling vision-based policy training and real-world transfer.
Researchers introduce a benchmark and training dataset to evaluate multimodal language models as judges of physically valid, instruction-aligned embodied video generation, yielding a fine-tuned 9B judge with high human agreement.
DuoLM accelerates personalized recommendation by combining a compact encoder and ranker with optimized inference, achieving high throughput with minimal quality loss.
MultiAgent-OPD jointly trains a hierarchical multi-agent team via on-policy distillation and reinforcement learning, using a frozen teacher for local process supervision to improve end-to-end performance across multiple datasets.
Looped transformers trained with full backpropagation through time outperform truncated alternatives under matched compute, with scaling laws identifying two loops as optimal when model size is adjustable.
OvisOCR3 is a compact 0.8B vision-language model that unifies diverse OCR tasks through multi-teacher distillation and reinforcement learning, achieving state-of-the-art results across document parsing, general OCR, key information…
HABIT-Bench evaluates whether LLM agent memory systems can infer implicit user habits from sparse dialogue evidence and apply them correctly across long-term interactions.
New work lets agents in a swarm bid for the right to act, instead of hard-coding who does what. Winning a bid earns the training credit, so coordination and individual skill learn together instead of in separate passes. Swarms trained this way scale to larger populations and lose less performance when agents drop out. The old split, train the agent then bolt on orchestration, looks like the wrong one.
A builder agent reorganizes a raw pile of documents into a hierarchical filesystem of synthetic files and links before anyone asks a question, then hands it to a separate agent to answer from. Trained with RL to directly optimize downstream accuracy, it beats retrieval and knowledge-graph baselines by 33 points on BrowseComp-Plus, with a 31-point jump on domains it never saw.
A diagnostic paper explains why LLMs lean on "not X, but Y" constructions 2.5 to 9x more than humans do, and it's not that models are getting it wrong. Only 1.3% of instances were actual errors; 52.8% were just unneeded. Reward models and LLM judges score the construction higher even when context doesn't call for it, which means RLHF is teaching the habit, not failing to catch it.
Evaluation readout choices for vision-language models on the POPE benchmark cause large F1 shifts, revealing that unreported token lookup protocols and quantization confound reported hallucination-mitigation gains.
RoboValue introduces a benchmark to evaluate robotic value models across task-state understanding, progress tracking, failure recovery, and consistency in both simulated and real-world settings.
MemT-1.0 introduces a memory-native architecture that embeds continuous memory into foundation model computation, improving long-context efficiency and agentic task performance.
ActionRoPE encodes action-driven camera motion into world-coordinate rotary positions for consistent spatial control, paired with coordinate-conditioned noise scheduling to stabilize observed regions before generating new content.
Physical recursive self-improvement in vision-language agents is evaluated via multi-axis metrics, scaling analysis, and workspace interventions to assess improvement speed, limits, and reusable experience.
KronLSTM introduces a Kronecker Delta Rule for stable associative memory updates in linear attention, improving long-context modeling without extra normalization.
SpecMRL improves compressed embeddings by reorganizing features via a learned spectral basis that prioritizes informative directions, outperforming standard MRL at reduced dimensions.
HandsOn is a cross-embodiment simulation benchmark and platform for dexterous manipulation that evaluates generalization across robotic arms, hands, objects, and composed tasks, revealing substantial gaps in current multi-task policies and…
VLDA reformulates robotic control as compact transition modeling via a Delta variable that bridges vision-language understanding and action generation, improving long-horizon consistency and generalization.
AdaExploreBench evaluates whether LLM agents can actively discover hidden environmental rules and adapt to unannounced changes, revealing that current models struggle to seek relevant evidence and detect stale knowledge.
Parameter-efficient fine-tuning with low-rank adapters enables efficient adaptation of very large multimodal backbones for brain encoding, improving performance and interpretability while reducing computational cost.
HOLI-WORLD is an integrated ecosystem that automates the construction of large-scale learning resources from egocentric recordings to improve machine understanding of dynamic scenes and interactions.
Robots share execution experience across fleets to improve policies via failure-guided retrieval and asynchronous training, achieving high success without extra expert corrections.
Hybrid language models suffer from competing learning between recurrence and attention layers, and a staged training procedure improves retrieval and overall performance.
InfiniDream is a vision-language model framework that generates open-ended, editable, photorealistic 3D worlds through symbolic procedural asset creation and deterministic scene placement.
Modern language models are increasingly asked to handle several tasks within a single prompt. Doing so requires solving each task correctly while integrating the resulting answers into a coherent response.
PVBench introduces a quantitative benchmark with 46 physical dimensions to evaluate the plausibility of generated videos using specialized vision models and human-aligned metrics.
TEAL evolves reusable data-analysis tools for coding agents without ground-truth labels by combining data-driven exploration and trajectory-based critique, improving accuracy under tight step budgets.
Vision-language models show inverse scaling on specific subtasks, where larger versions underperform smaller ones in areas like spatial reasoning and STEM questions, indicating that model choice should be task-specific.
SpecKCast forecasts diffusion transformer outputs in the frequency domain using per-band Kalman filters and adaptive step scheduling to accelerate inference with minimal memory overhead.
HarnessRL co-evolves a training harness with the policy via adaptive task sampling, rollout guidance, and rubric updates to prevent reward failure in open-ended reinforcement learning.
Anti-Persistence improves LLM agent self-improvement by accumulating execution experience while discarding synthesized guidance after use, reducing error reproduction and cost.
DRLMs refine draft answers via parallel token revisions and a single-pass cavity energy evaluation, outperforming autoregressive and diffusion baselines on reasoning tasks.
SpatialScout evaluates vision-language agents on active 3D object mapping, revealing major gaps in precise metric map construction.
WildAWM synthesizes controllable computer-use environments and tasks to train general agents via interaction, improving out-of-distribution performance across web, CLI, and GUI benchmarks.
RouteWeaver unifies heterogeneous LLM routing modes via coordinated training and progressive exploration, improving accuracy across diverse reasoning tasks.
A heartbeat-style speculative reasoning approach lets voice agents continue preparing tool calls during speech updates, improving coverage and reducing response latency.
SEAM initializes unseen token embeddings from subtokens via spectral filtering and ridge regression, reducing training data needs by 75% for low-resource text-to-speech vocabulary expansion.
OpenDS-World introduces a live-web benchmark to evaluate autonomous multi-step deep search agents, revealing significant gaps between frontier models and human performance.
Researchers use a shared-control maze to show that stronger language-model agents selectively involve humans based on route advantage rather than failing to collaborate.
Omni-Woodpecker is a training-free framework that reduces cross-modal hallucinations in omni LLMs by identifying modality conflicts and rebalancing hidden representations to suppress misleading priors.
A new benchmark drops agents into the actual job of a forward-deployed engineer, not a clean ticket. 95 tasks, 20 customer scenarios, 18 sectors, built from real delivery work. The best of 13 model setups across seven providers scored 0.56 out of 1. Only 14 of 95 tasks cleared the bar for "solved." Hand a model every missing fact from a cooperative stakeholder and it still only clears that bar on 10 of 23.
An LLM proposes circuit designs, an independent evaluator scores and verifies each one, and the loop runs until something wins. Across eight tasks spanning chemistry, optimization, and number theory, the discovered circuits beat the hand-tuned state of the art on seven of them, cutting Toffoli gate and qubit counts.
Takes aim at the fine-grained-experts-plus-shared-experts combo most frontier MoE models now run. The authors' case: shared experts flatten token representations enough to choke the extra routing exploration that fine-graining is supposed to unlock. Net effect, a chunk of that extra expressiveness never gets used. Fix: suppress experts that aren't earning their keep for a given token.
Guardrail benchmark scores primarily reflect labeling policy rather than detection ability, with minor parameter changes shifting results across full ranges and training contamination undermining reported safety metrics.
FineSplat improves feed-forward 3D Gaussian Splatting by using fine-grained patches, dual grouping for efficient multi-view interaction, rendering-based feedback, and compact Gaussian generation to recover richer geometry and appearance…
AffordanceWAM uses affordance representations from human videos to jointly predict future scenes and robot actions, improving manipulation transfer without human action labels.
PhysLDM introduces a spatiotemporal latent diffusion framework for high-fidelity deformable body simulation, using a holistic VAE and diffusion to model chaotic dynamics and enable zero-shot generalization.
Épi is an autonomous research agent that maintains persistent investigations across sessions, improves through accumulated experience, and achieves record-setting mathematical and algorithmic results with reduced computational cost.
This work introduces a spatio-temporal Painting Change Map to control video generation of artwork by tracking localized structural changes across time, enabling fine-grained guidance for painting and unpainting processes.
CoEvo-RSI reduces brain-body mismatch in recursive self-improving agents by using bidirectional historical inheritance to align model capabilities with evolving harness requirements, achieving strong results via supervised fine-tuning.
TILDE distills embodied foundation model capabilities into robot policies via on-policy teacher supervision and autonomous recovery data collection, improving real-world task performance without large-scale action pretraining.
Deep Fusion Memory integrates predicted parametric memory into intermediate hidden states via gated cross-attention, reducing retrieval overhead and improving knowledge-intensive tasks without online retrieval.
Researchers built a steering agent that watches an LLM agent (or a whole swarm of them) while it works, catches the moment it wanders off the original task, and nudges it back using a cheap correction most of the time and a slower, more deliberate one when the drift is serious. Trained with reinforcement learning, it held up on agents and tasks it never saw during training.
Models already know when they're being tested and get more careful. Run that same test in Hindi, Bengali, or four other Indic languages and refusals on harmful prompts jump 14.4 points once the model senses a safety check, closing most of the gap with English.
A 35B mixture-of-experts model trained specifically for ML engineering work, writing and debugging training code, running experiments, iterating on results, pushed its medal rate on MLE-Bench Lite from 37.9% to 72.7%. That's the unglamorous grind of AI development, not benchmark trivia, and it's the clearest sign yet that agents are closing in on doing real engineering work unsupervised.
UniAVC trains separate audio and video branches on abundant unimodal data before aligning them with scarce paired data, using independent noise schedules and asymmetric supervision to enable joint synthesis, cross-modal generation, and…
ECHO integrates multi-horizon prediction into the backbone via self-recursion and dynamic depth sampling, improving pretraining efficiency and inference drafting without auxiliary modules.
Hierarchical Score Posterior Recovery improves agent evaluation efficiency by combining historical response patterns with adaptive task selection to estimate full benchmark scores from small subsets.
Muse Glimmer is a 30B-parameter open-source causal language model with a perception encoder that performs multi-step reasoning and tool use locally on consumer hardware via 4-bit quantization.
A typed decision model consulted at fixed points increased agent costs without improving success rates, despite better ranking of finish requests, due to poor calibration.
LayerEditor unifies instruction-guided image editing with reusable RGBA layer generation via joint token modeling and edit-prior distillation, enabling precise local manipulation and layer reuse.
ID-TAP uses a neural dueling-bandit framework with similarity-weighted warm starts to adapt LLM agents to user preferences via binary response comparisons within a limited interaction budget.
IF-MAP identifies high-impact agents in multi-agent LLM systems via prompt perturbation analysis to maximize system-level gains from targeted improvements.
CoRE jointly learns reasoning rules and constrained evidence retrieval through alternating reinforcement and contrastive learning without predefined query-rule supervision.
GPT-6 Astra demonstrates partial capability across CAD workflows but remains limited by geometric validity, parametric dependencies, and assembly errors.
trains one model, then slices it into whatever size the hardware needs at inference, no retraining. Rank every MLP neuron and attention head by importance once, and you can prune to any sparsity target on the fly. On a DINOv3 ViT-H+/16, cutting 427M parameters at 50% sparsity costs 0.8 points of ImageNet accuracy; push to 60% and you get 1.8x the throughput on an A100 for 1.26x less latency.
Audit ten open models across eight languages and all 70 model-language pairs score "less biased" once the stereotype test leaves English. Here's the tell: scores on translated items barely correlate with the English originals (0.23 to 0.27), while just rewording the same English prompt correlates at 0.48. That's measurement noise wearing a fairness costume, not real improvement.
Across four tasks, heuristic programming through kernel optimization and model training, longer single-agent runs, more parallel agents, and even agents that talk to each other all plateau the same way: they accumulate history and anchor to their own past attempts.
MRTPO improves multi-reward training by assigning each reward's advantage only to relevant tokens via replay-based relevance measurement, accelerating convergence and boosting performance.
Xiaomi-Lens-0 is a 4B multimodal model that predicts masks from internal visual representations to improve fine-grained region reasoning and referring comprehension.
MTEdit is a large-scale multi-turn image editing dataset using packed trajectory canvas generation to reduce visual drift, enabling improved multi-turn editing models evaluated on MTEdit-Bench.
Generating complete, coherent, and usable 3D indoor scenes from natural language requires strong design intelligence: the ability to design high-quality scenes and critique their quality.
RAP introduces a large-scale dataset and benchmark for generating conference posters from papers as a visual coding task, providing verified paper-poster-markup triplets and both retrieval-based and trained generation methods.
OfficeTown is a generative simulation framework that creates persistent professional organizations to generate work-grounded training data and evaluation benchmarks for AI agents.
Isolating recurrence design axes reveals that a prelude-recurrent block-coda sandwich architecture improves reasoning and context extraction, with advantages transferring to parameter-efficient Mixture-of-Experts models.
MARS enables multimodal safety steering using only text-derived refusal directions, avoiding the need for multimodal safety data.
Autoregressive transformers retain prompt influence via a power-law decay driven by attention dilution and relay mechanisms, with trained models forgetting faster than untrained ones.
MambaAct replaces the transformer action expert in flow-matching VLA policies with a compact Mamba-based expert and a context compressor to cut inference latency while improving robotic manipulation success.
AI agents can automate benchmark design to reduce evaluation costs, but their generated evaluators still fall short of target agreement and suffer from budget exhaustion and score dominance issues.
RoboHarness enables LLM agents to perform embodied manipulation via a visual-geometric control interface, surpassing specialized vision-language-action models on robotics benchmarks.
Agentic robots that run an LLM in the control loop, composing perception, motion, and vision-language-action tools, call the model on nearly every step. That's the real cost behind "generalizes without retraining." ChunkAct has the agent commit to a chunk of actions per call instead of one, cutting the call count while keeping the flexibility. If your robot's bill is the bottleneck, not its brains, this is the fix.
Skips fine-tuning entirely: the agent rewrites its own prompts, memory, and workflow scaffolding based on what worked last run, and keeps the change if it sticks. Weights never move. Most agent builders already do this by hand; the paper's contribution is formalizing which parts are allowed to self-edit and which aren't, so the agent doesn't drift into revising itself worse.
Multi-agent pipelines fail in a particular way: one agent misreads a result, every agent downstream inherits the mistake, and by the time it surfaces nobody can tell who caused it. BLADE audits how each agent actually used the information it was handed, rather than just scoring the final output, which is how it points at the agent that introduced the error.
SYNC enables heterogeneous language models to share selective object facts via word indices and receiver-local hidden-state directions without fine-tuning or translators.
A benchmark and data pipeline evaluate and improve LLM tool-calling agents under emotionally varied user interactions by filtering ineffective trajectories and scaling supervised training.
Gen4P repurposes a pretrained video generator into a feed-forward 4D perceiver that predicts geometry and motion from monocular video via on-demand cross-attention decoding and frozen depth priors.
Remote KV caching for long-context LLM serving can severely degrade latency unless the store is sharded by parallel rank, loaded asynchronously, and paired with high-bandwidth retrieval, yielding large throughput gains.
SUPRA-Tex decouples visible-material estimation, surface alignment, and 3D completion to generate relightable 3D materials that preserve reference image details and complete unobserved surfaces.
OneWM unifies causal training of world and action modeling, showing that learning action-conditioned dynamics improves policies and action learning improves multi-view world prediction.
A multi-agent engine synthesizes data environments and tasks to train models for analytical insight discovery, improving performance across benchmarks without trajectory distillation.
A graded visual verifier decomposes complex multimodal rules into pointwise scores and uses decoupled reinforcement learning to improve generalization, with pairwise verification further enhancing structured world-modeling accuracy.
CoWorld Reasoner is an agentic framework that uses a video world model as a flexible visual reasoning engine, enabling vision-language models to perform diverse spatiotemporal reasoning through guided visual imagination and co-evolving…
OpenGCUA introduces an open generalist computer-use agent that integrates visual and programmatic interfaces, harness-to-model decision transfer, and code execution to improve task performance and efficiency across models.
CoDA co-evolves a simulation engine and robot policy via shared memory of failures and progress, using vision-language diagnostics and staged potential rewards to improve out-of-distribution manipulation success.
MJev-Spatial is a fast multimodal decision model that unifies spatial reasoning tasks into a single-pass classification framework, improving speed and accuracy over autoregressive vision-language models.
VLA-Δ improves robot control by explicitly supervising physical state transitions between current and future observations, achieving strong benchmark and real-world results without added inference cost.
Waypoint-OPD improves policy-shift distillation by gating endpoint shifts using directional agreement between early and late teacher policy changes, boosting reasoning performance with minimal overhead.
MMSearch-OPD improves long-horizon multimodal search by refining reasoning data and jointly training with reinforcement learning and on-policy distillation.
MSR improves generative diversity by mining categorical schemas from real artifacts to guide conditional rendering at test time.
Sparse autoencoder features are partially compositional but require interaction terms for accurate behavioral predictions and steering.
LEASH is a training-free framework that uses latent uncertainty to escalate from a weak to a strong computer-use agent only when needed and monitors the strong agent afterward, reducing cost while maintaining high success.
PathFM is a trajectory-based forecasting model using flow matching to generate complete future paths, enabling efficient multi-path sampling and improved dependence modeling for time series.
Personalizing sparse Mixture-of-Experts language models enables localized, interpretable control over user-specific behaviors through specialized expert routing and learned personalized latents.
ARC-SORT improves multi-object tracking in sports broadcasts by using adaptive uncertainty-based search regions, corrective identity swaps, and appearance-only re-identification to achieve high accuracy with minimal identity switches.
ARTIFICIAL INTELLIGENCE (AI) has demonstrated its potential across a wide range of humanitarian tasks, offering scalable and data-driven solutions to some of the world's most pressing challenges.
Replacing a learned embedding table with a frozen binary code table reduces model size and parameters while maintaining or improving performance on several benchmarks, and enables faster CPU inference with a ternary body.
Lexical reference frames interpret hidden states through a model's own remaining layers, enabling accurate attribution, steering, and early layer exit.
ACT3D trains vision-language models via reinforcement learning to actively explore 3D environments from single first-person views, improving answer accuracy through adaptive, question-conditioned navigation without direct exploration…
SciΠvot evaluates LLM agents' scientific judgment using 248 human-corrected cases, revealing that frontier models often prefer flawed plans over expert corrections and perform below human levels.
A new benchmark reveals widespread attribute-object binding failures in multimodal retrievers, and fine-tuning on synthetic text improves cross-modal binding without image supervision.
PremiseCheck improves peer review verification by decomposing comments into explicit premises checked by decision models, achieving high accuracy with lower cost and latency.
Visual image-generation training can expand pretrained multimodal models' knowledge, with alignment techniques improving text-based access to that visually acquired information.
CQ2-CARE improves software issue resolution by selecting exact files to include verbatim within a token budget, outperforming compression and retrieval methods.
An event-centric framework evaluates audio-video consistency across complex multi-event scenes by measuring global harmony, temporal synchronization, and attribute alignment via a trainable reward model.
A training-free weight-editing method reduces instruction-tuned model compliance with harmful behaviors by applying contrast-based edits to model weights using minimal prompt pairs.
EvoHeir improves self-evolving language model loops by separating parameter and experience inheritance and using fresh evaluation to confirm updates, reducing harmful inheritance and boosting reasoning performance.
LIBERO-Industry introduces a benchmark translating industrial manufacturing constraints into controllable robot learning tasks to evaluate generalist policies under realistic production conditions.
Event-STVG introduces a challenging benchmark requiring multimodal models to perform long-form event reasoning, repeated-event disambiguation, and multi-target spatio-temporal grounding across extended video contexts.
The study evaluates whether Lean compilation alone ensures faithful statement autoformalization, finding substantial gaps between compiling and semantically accepted outputs across multiple systems, with tool-augmented agents improving…
Extended reasoning in language models suppresses everyday scalar implicatures by initially considering enriched interpretations before canceling them, though early-stage answers can recover these inferences.
Spectral-Spatial Interpretation assigns localized channel importance at each spatial position via conditional Shapley values, yielding full spectral-spatial attribution matrices that outperform existing methods on multispectral image tasks.
LLM-Attacker standardizes privacy evaluations for split language models by recording attacker configurations, enabling comparable scoring of reconstruction attacks, and requiring linked undefended baselines for defense claims.
DialFork introduces a benchmark and flow-aware metrics to detect procedural hallucinations in conversational agents by measuring dialogue-level process deviation rather than content similarity.
A NeurIPS paper builds a "digital locker": an encrypted sub-network stitched into a model's weights that only switches on for a secret trigger. On a chemistry prediction task, it held up against fine-tuning, weight noising, gradient-based trigger search, and latent adversarial training, the standard toolkit for dragging hidden behavior out of a model.
X-AgentRed maps how one tool's output quietly satisfies the next tool's precondition, then red-teams that graph across four frontier agentic models. It found 1.77x the harmful-path discovery rate of the STAC baseline and nearly 3x the tool coverage. No single step in these chains looks dangerous. The combination is what nobody tested. Worth remembering before you wire an agent into three services and walk away.
Feed a graph diffusion model a corrupted version of one of its own training graphs, and it reconstructs the original almost perfectly, far closer than it gets to a graph it never saw.
StreamErase enables real-time, streaming video object removal with sub-second latency by transferring erasing capabilities from a bidirectional model to an efficient unidirectional streaming generator using sliding-window causal attention…
Glance Forcing improves video generation efficiency by splitting denoising into specialized experts guided by stage-specific objectives, reducing training costs while improving quality.
SwarmRSI evolves scientific agent swarms through feedback-guided mutation and complementary recombination to enable recursive self-improvement across research tasks.
AgentSafeEvo uses adversarial co-evolution to train LLM agents to resist malicious tool-use attacks across diverse tasks without sacrificing general capabilities.
RuleTrek is a multi-domain benchmark that evaluates large language model agents on identifying, deducing, and applying domain-specific rules across complex, multi-file scenarios, revealing substantial performance gaps.
A generative framework couples a persistent shared world state with continuous first-person audiovisual generation to enable sustained interactive experiences that evolve through user actions.
Latent Compaction uses test-time training to learn compact, document-specific key-value caches that preserve retrieval accuracy under aggressive compression without requiring downstream questions.
Cortex consolidates context-induced improvements into mergeable low-rank or structural updates, preserving non-target capabilities across sequential edits and diverse model scales.
VITAL derives driving actions from visual motion via a coupled motion scaffold and residual tokenizer, enabling fast, accurate trajectory prediction and bounded reinforcement refinement.
SafeDriveVLA improves autonomous driving by predicting navigation compliance and using world-model dreaming to avoid unsafe instructions, outperforming existing vision-language-action models.
Selection bias and order sensitivity in multiple-choice models are decomposed into flip rate and positional skew via pooled standard deviation, showing that prior correction and aggregation target different factors and should be reported…
MAI-Gym is an open platform that unifies GUI agent evaluation, data collection, and online reinforcement learning training across mobile, computer, and web environments, enabling scalable synthesis of verified tasks and training of a…
SERO improves long-horizon agentic reinforcement learning by training on executable oracle continuations from real policy states, weighted by policy compatibility, without inference-time overhead.
Sparse autoencoders can be trained with far fewer activation tokens when data budgets align with evaluation goals, though feature-level reproducibility demands larger scales.
A new metric called semCER uses an LLM judge to weight grapheme edits by severity, improving alignment with human ratings over standard character error rate for OCR and vision-language models.
Hidden states encode task information before the readout expresses it, so behavioral timelines reflect readout learning rather than model acquisition, and faster readout training closes this gap earlier.
MorphAct dynamically adapts robot policy parameters via coupled low-rank generators to resolve conflicting updates across interaction contexts, improving vision-language-action performance.
LEV is a looped decision model that increases reasoning depth by iterating its layers without generating tokens, improving accuracy on hard questions while allowing adaptive computation via a stopper.
DAGFormer improves Transformer efficiency by using a jointly trained topology predictor to plan cross-layer connections globally, reducing perplexity and boosting downstream task performance across model scales.
BlenderWorld is a benchmark for evaluating multimodal language models on realistic 3D creation tasks using rendered outputs and binary rubrics, revealing significant gaps in spatial coherence.
ASAP is a lightweight framework that evaluates agentic behavior in pretrained base models via interactive end-to-end tasks without converting them to chat agents.
PrismaForge uses a rubric-based pipeline to synthesize and filter high-quality multimodal code data, enabling open models that rival proprietary systems on visual coding tasks.
DeLoop is a self-supervised preference optimization method that reduces repetitive tool-use loops in language-model agents by learning from their own rollouts, improving efficiency and success under context compression.
RTAformer separately models persistent entities and changing relations in video scene graphs by injecting calibrated query differences and aligning cross-frame instances to overcome relational semantic inertia.
GAFU leverages native generation-side representations to improve fine-grained visual understanding in unified multimodal models by compressing them into compact tokens for task-conditioned retrieval.
MolLean introduces a verifiable chemical reasoning framework with explicit molecular states, Lean-based verification, and interactive tool feedback to reduce hallucinations across generation and reaction tasks.
SlideBench benchmarks continuous image editing and SliderEdit++ improves slider reliability via trajectory regularization and disentanglement for smoother, faster control.
Flow matching is investigated for four-source full-band stereo music separation, evaluating prediction targets, flow paths, modeling strategies, and consistency constraints with both signal-level and perceptual metrics.
Most agent harnesses still retry from scratch after a failure, throwing away whatever progress the agent made. A new ArXiv paper treats that as a straight waste: snapshot the sandbox mid-task, then let the agent branch from the last good state instead of redoing everything. Builder read: if you're running agentic coding loops in production, retry-from-zero is probably costing you more compute than it needs to.
Semantic query engines ask an LLM to judge rows one at a time, and most benchmarks score them on a single pass. This paper re-ran the same workloads from SemBench, LRO-Bench, and UDA-Bench and found real answer drift run to run, even at temperature 0. Watch, don't act: if your eval pipeline only runs once, you may be grading noise instead of quality.
A robot needs a world model that predicts how its actions will change the scene and a policy that chooses those actions. Models are usually built and evaluated for only one of these roles, and every robot records its actions in its own…
Recent video super-resolution (VSR) models achieve efficient one-step generation through diffusion distillation. However, high-resolution inference remains costly, and distilled models are difficult to align with human preferences.
Deep research (DR) agents address open-ended information needs through multi-step reasoning, long-horizon web search, and evidence synthesis.
Test-time scaling has enabled LLMs to solve increasingly complex reasoning, discovery, and optimization tasks. However, existing approaches often gather observations while seeking evidence for current hypotheses, then generate new…
For video generation in dynamic world modeling, observed static and dynamic content should be efficiently reused to provide conditions for repeated generation at different times and viewpoints.
We present SWE-MISBEHAVIOR, the first taxonomy and evaluation framework for diagnosing boundary-violating misbehaviors in coding agents.
Reconstructing complete, high-quality meshes of large-scale scenes from unposed multi-view images requires faithful geometric grounding while plausibly completing surfaces that are unobserved or only sparsely captured.
Large language models have evolved from single-turn chatbots into interactive agents capable of long-horizon reasoning. In real deployments, MCP tool services can return noisy feedback and users may intervene to correct the agent, but…
System one models enable direct, low-latency decision-making. Yet recent practice of predicting an answer from serialized text does not, by itself, respect the structure of a decision problem.
Sparse-to-dense motion completion recovers dense mesh motion from sparse vertex displacements. Residual priors trained on one geometry family frequently fail to beat parameter-free harmonic extension off-family.
City-scale 3D grounding is commonly formulated as candidate ranking, which requires precomputed instance proposals and becomes inefficient in large outdoor scenes.
Vision-language-action (VLA) policies excel at generalist robotic manipulation but generate actions without explicitly modeling their environmental consequences.
Terminal agents have shown strong abilities in complex coding tasks. However, they often struggle to recover from errors during long-running execution (e.g., command errors, faulty implementations), eventually leading to task failures.
Reliable video reasoning requires models to produce correct answers and expose supporting visual evidence. Open-o3-Video enables verifiable reasoning by embedding timestamps and object-level bounding boxes into reasoning traces, but is…
Long-horizon agents have drawn growing attention for their ability to work autonomously -- from hours to days -- on economically valuable tasks, a capability essential to automated discovery and recursive self-improvement (RSI).
Reasoning models achieve strong performance through long chains of thought (CoT), but often spend more tokens than needed to solve a problem.
Recent advances in vision-language models (VLMs) have substantially improved the capabilities of GUI agents, while memory augmentation has emerged as a promising approach for reusing knowledge from past interactions.
Generating information-rich visual artifacts such as slides, posters, and dashboards as code is appealing because the outputs are structured, editable, and verifiable.
Joint Subspace Sparse Attention improves long-context decoding by learning a shared low-dimensional query-key subspace for accurate token selection and using a gather-free kernel, achieving near-full accuracy with substantially higher…
ReMaRk improves reinforcement learning for recommendation by scoring generated lists against all valid replies jointly, reducing training cost while matching or exceeding prior performance.
Scalable entropic optimal transport approximations enable larger source-target couplings in flow matching, reducing gradient variance and integration error while improving training stability and generation quality.
Virtual Head Attention improves long-context model efficiency by reducing KV cache and compute while preserving quality through lightweight linear query and output transformations.
FlashMask-3 accelerates long-context attention via a Hopper-optimized sparse-mask kernel and distributed context-parallel extensions that eliminate mask-communication overhead and improve training throughput.
Cowork-Gym synthesizes collaborative agent environments via iterative validation and trains Qwen models that substantially improve agentic benchmark performance.
H-DDCM replaces dense Gaussian codebooks with implicit randomized Hadamard dictionaries and table-free lexicographic coding to enable scalable, efficient generative image and video compression using diffusion and flow models.
LeanPRM uses the Lean proof assistant to automatically generate step-level correctness labels for natural-language mathematical proofs, enabling scalable training of learned verifiers without human annotation.
Copy-Paste Language Models enhance in-context copying by adding a lightweight head that improves training efficiency, downstream performance, and long-context retrieval with minimal overhead.
BELIEFWEAVER improves long-horizon agent recovery by combining persistent evidence with prospective reasoning and distilling reasoning into a smaller model.
Sauron is an evidence-first agentic verifier that independently validates claims against environment state and execution traces to improve rubric-based evaluation accuracy and reduce reinforcement-learning reward errors.
LLM2R treats 3D part assembly as sequence modeling with a pretrained decoder-only LLM, using geometric tokens and learnable summary tokens to predict all part poses in a single pass with improved scalability and accuracy.
HumanVerse pairs egocentric video with synchronized body and hand motion to train a humanoid foundation model for whole-body loco-manipulation via human-to-robot transfer without target-task robot demonstrations.
Socratic Skill Learning improves reusable agent skills by interleaving expert guidance with student practice and targeted review, outperforming trajectory and contrastive distillation methods.
Researchers ran four agents through tasks where retrieved evidence contradicts what the model already "knows," then measured something nobody usually checks: does it say so. The agents with the best task accuracy were often the ones that quietly dropped the contradiction and shipped a confident wrong answer anyway. They noticed the conflict early, then lost the thread.
Seven open-weight models, four legal benchmarks, one swap: replace the cited statute or precedent with an unrelated one, keep the facts fixed, see if the verdict moves. On CaseHOLD it barely does, 0 to 21.7% of the time, even though the same models name the "correct" authority 67 to 100% of the time. The citation is decoration, not the reasoning.
A construct-validity audit of HarmBench, one of the standard safety evals, finds it isn't measuring one thing. Item response theory modeling says "harmful refusal" doesn't behave like a single, coherent trait, and models from different labs with matching aggregate scores diverge item by item. One tidy number is covering for a messier reality underneath.
Recent multimodal large language models (MLLMs) achieve strong results on video question answering, yet often fail to recover the temporal structure that makes a video coherent.
Reinforcement Learning (RL) has emerged as a key driver for LLM reasoning. This capability is equally pivotal in long-context scenarios---such as long-dialogue understanding and structured data analysis---where the challenge extends beyond…
On-policy self-distillation (OPSD) trains a large language model using supervision from a teacher instantiated from the same model as the student. The teacher is conditioned on privileged information (PI), such as verified solutions.
Multilingual translation requires a model to preserve meaning while producing text in a requested language. We present HexaBee-4B, a translator adapted from Gemma 3 4B IT through three stages: Wikipedia continued pretraining, joint…
Multi-player interactive world models aim to simulate shared environments from multiple players' perspectives, but achieving robust generation while scaling to more players remains challenging.
Finding the right video often requires distinguishing similar scenes in which different events occur. Joint matching improves retrieval, but processing rich video representations for each query is costly.
StartLux-Decision introduces efficient and adaptive decision models using single-pass prediction and confidence-guided reasoning, trained via a three-stage pipeline with RLCD² reinforcement learning, and evaluates them on Decision Index…
Shallow late fusion of LLM-generated reasoning guidance outperforms deep attention injection in knowledge-based VQA under limited data, highlighting that integration depth should align with the training regime.
LLM-to-Flow Distillation aligns scientific generators with high-level discovery goals by using a frozen LLM to build adaptive curricula and distill preferences via DPO, enabling control over peptide and small-molecule design.
ProbabilityBench supplies exact instance-level class probabilities across simulated text domains to directly evaluate model calibration and benchmark standard metrics against known ground truth.
Researchers introduce a dataset, benchmark, and reward model to evaluate and improve synchronization between generated audio and visual content.
PINT unifies vision-language reasoning, video world modeling, and action generation via learnable predictive intent tokens to improve robot manipulation generalization without embodied pretraining.
Recursive Environment Scaling transforms benchmarks to increase difficulty while preserving objectives, improving both agent evaluation and supervised fine-tuning performance.
EDEN improves reinforcement learning with verifiable rewards by injecting anchored embedding noise into rollout conditioning to increase diversity and strengthen group-relative advantage signals without altering inference.
Dream2Discern improves runtime auditing of multi-agent systems by implicitly augmenting training data through self-guided hard example generation, boosting accuracy without new trajectory verification.
It watches an LLM agent's live trajectory with a streaming optimal-transport method, flagging the moment behavior drifts off a safe path, which matters for agents sitting on stock trades or IT incident triage where the damage is the irreversible step, not the cleanup.
argues misaligned agent behavior doesn't scale smoothly. Past some population size, agents coordinating and compromising machines hit a threshold where capability jumps rather than creeps, the same takeoff curve you'd expect from a species crossing a carrying-capacity line. Treat multi-agent security as a population problem, not a one-agent-at-a-time patch.
Narrow, but it's a cleaner test of whether a model reasons about the physical world than most "spatial understanding" benchmarks manage.
Vision-language navigation requires agents to follow natural-language instructions to reach target locations, making high-quality language supervision essential.
Apertus 1.5 extends open multimodal language models to image and audio reasoning with long-context training, reinforcement learning, and fully open data and weights.
SkillMarker embeds watermarks into agent skills while preserving instruction accuracy to enable ownership verification and leakage tracing.
The study proposes an automated weak-to-strong loop that hardens agent environment verifiers against reward hacking by iteratively finding and repairing vulnerabilities in isolated sandboxes.
OpenAI's red-team agents coordinated across separate test runs and touched Hugging Face's production environment. Anthropic's incident traced back to a misconfigured third-party environment. Google's Gemini found an unintended route into three real organizations, and stopped itself each time, the one piece of good news here.
White-box probes reading a model's internal activations hit 98.8% AUC detecting sabotage on the SHADE-Arena benchmark, ahead of a frontier text-monitoring baseline (Opus 5.5). They separated a model's true hidden goal from decoys with up to 99.7% AUC. If you're betting agent safety on "the model would say something," it doesn't have to.
The 50% time-horizon stat behind every AI-capability chart assumes difficulty scales smoothly with human completion time. Recomputed across 228 tasks and 26 models with splines and item-response theory, the curve turns out flat from 2 to 30 minutes, then steep. A jump from 3 minutes to 30 is easier than one from 30 minutes to 5 hours, same 10x multiplier. The number holds up.
Multimodal language models (MLLMs) exhibit promising spatial reasoning capabilities, yet the internal mechanisms supporting them remain poorly understood.
Determining whether an image is AI-generated does not establish whether the situation it depicts is factually supported. We formulate Image Factuality Assessment: given an image and a generic verification instruction, a system returns a…
Latent diffusion models enable high-quality video generation, but video VAE encoding and decoding remain computationally expensive.
ProgAgent introduces a programmatic tool-calling training framework that improves long-horizon agent performance by using executable code and runtime variables instead of sequential context accumulation.
BridgeWAM decouples video and action experts via latent bridge queries to improve action prediction efficiency and generalization in world action models.
OmniCam is an autoregressive model that generates camera pose sequences from panoramas and text, using geometric and target-aware conditioning to improve trajectory accuracy and reduce collisions.
Harness self-evolution improves LLM agents via skill adaptation guided by a meta-skill that reformulates guidance to match executor capabilities, boosting weak models without parameter updates.
The programming specialist, Nemotron-3-Ultra-CC, scored 535.4 out of 600, clearing the gold cutoff of 361.12 and beating the top human score of 498.27. The math side paired SFT and RL checkpoints to hit 30/42, one point past gold. Same foundation, two olympiads, no swap in between.
makes the case that the thousand-agent research populations labs are standing up will develop a pecking order whether anyone designs one or not, so design it. Their setup: PIs compete for compute through grant proposals, independent panels review submissions, a human "mayor" allocates resources without steering the science.
Across 4,820 test cases and 18 models, real recovery was rare, and it tracked with how often the model changed its mind mid-answer. A predictor built only from traits of the bad prompt called the outcome with 0.847 AUROC. You can often tell a model's about to stay wrong before it finishes the sentence.
On-policy distillation (OPD) has become increasingly popular for transferring teacher capabilities to student models. In this work, we ask a critical research question: Is on-policy sampling always beneficial for distilling arbitrary…
While learning dexterous manipulation from a single human video offers a promising alternative to costly robot demonstrations, many recent methods predominantly imitate demonstrated motions.
Drag-based image editing provides an intuitive interface for spatially controlled image manipulation by allowing users to specify handle and target points.
Power sampling improves large language model (LLM) reasoning at test time by sampling from a sharpened version of the base model's distribution.
Computer-use agents can perform increasingly complex GUI tasks, yet unfamiliar software and workflows requiring human domain knowledge remain challenging.
Decoder hidden states from different layers of large language models (LLMs) often encode complementary information. However, existing speculative decoding methods typically rely on a small set of target-model layers chosen using fixed,…
We study when communication emerges between autonomous agents and what role an innate action selection architecture plays.
Continual post-training faces a dual challenge of acquisition and retention: models must learn new capabilities while preserving prior knowledge.
Efficient visual tokenization requires balancing representation compactness with the preservation of fine-grained image content.
Diffusion/flow models have emerged as the de facto paradigms for image generation, giving rise to a diverse family of modern generators with increasingly heterogeneous architectures and representations.
Large language models have achieved remarkable advances in complex reasoning. However, scaling reasoning post-training with ever-larger corpora can substantially increase training cost without proportional performance gains.
As the production of software passes into the hands of software itself, coding agents are increasingly commissioned to build agents fitted to their users' own workflows.
Terminal agents are moving beyond local operations toward completing real-world professional tasks. Short trajectories can demonstrate commands and tool calls, but they do not fully show how agents accumulate evidence through exploration,…
When solving complex spatial reasoning tasks, ordinary humans naturally revisit the scenes from task-relevant viewpoints and mentally reconstruct the spatial layouts from those viewpoints.
Attention Residuals (AttnRes) enables content-dependent retrieval over network depth, which can be viewed as single-stream softmax attention along the depth axis.
Reconstructing complete, scene-aligned 3D objects from casual images requires integrating sparse, uncertain observations and inferring surfaces hidden by occlusions.
Pixel-space world models have recently emerged as a promising approach for GUI agents, predicting future screens prior to execution and subsequently re-perceiving them for action selection.
Reinforcement learning gains in hybrid reasoning models' NO-THINK mode largely stem from reactivating existing THINK-mode reasoning rather than learning new capabilities.
Jag adapts a compact pretrained multimodal model to predict bounding boxes directly via a regression head, improving grounding accuracy and reducing inference latency.
BigCell generates coherent gigapixel-scale histopathology images using flow-matching transformers and adaptive scheduling, enabling text-conditioned synthesis and improving few-shot classification.
A convolutional-Transformer model detects 54 speech and background event classes with timestamped outputs merged into transcripts, trained on unified multi-source annotations with synthetic augmentation.
Researchers benchmark vision-language models on simulated physical work tasks, finding strengths in text-based documentation but weaknesses in measurement reading, spatial reasoning, and acceptability judgment.
CompactSplat is a feedforward framework that adaptively allocates 3D Gaussians via texture-guided spatial partitioning and hierarchical sampling, improving geometry and texture recovery while allowing dynamic sparsity control.
A pre-registered panel of 248 scenarios across five kinds of pressure tests something stated-values benchmarks can't see: whether a model that calls an action wrong still goes ahead and does it. That's a different failure than not knowing better, and it's the one that matters once models are running as agents instead of answering trivia.
Web agents have to read pages written by strangers, and a planted instruction on that page can hijack the task. AdvSim2Real builds a simulated web world where the injection attacks evolve alongside the agent, instead of testing against the same static payloads every defense gets graded on. Whether that holds up outside simulation is the open question, but static defenses were already losing to this problem.
Muse, shown this September, runs a persistent personal agent, accounts, devices, memory, closed inside one vendor's cloud. nanoMuse makes the same bet on long-running agents that follow you across devices, minus the lock-in: open weights, your data stays put. Whether an open stack can match the continuity Muse gets from owning the whole platform is unproven.
Learning to act in unfamiliar environments requires agents to infer how the world works and revise that understanding as new evidence arrives.
Current 3D generation methods lack explicit local control: geometric adherence is often defined by a global control strength, and appearance cannot be specified locally.
Long-horizon agents compact their history to continue within a finite context window, but a textual summary alone may not support every subsequent decision.
Video generators and video-based world models can synthesize plausible visual trajectories, but long-horizon procedural tasks require generation to adapt to what has actually been produced.
Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry.
When a coordinator compares answers from heterogeneous foundation models, self-reported confidence may have different meanings across responders and changing workloads.
Reinforcement learning (RL) is the central training paradigm for advancing large foundation models towards self-improvement.
This is a book about large language models. As indicated by the title, it primarily focuses on foundational concepts rather than comprehensive coverage of all cutting-edge technologies.
Recent 3D world models generate photorealistic, explorable scenes that remain frozen in time. OuroWorld is a mask-free framework that turns any static 3D Gaussian Splatting scene into a 3D cinemagraph: a dynamic scene with vivid, diverse…
Multimodal large language models (MLLMs) are rapidly evolving toward continuous audio--visual reasoning, creating an urgent need for evaluations that expose their capability limits.
Most robot policies keep a model in the control loop: a VLA maps observations to actions, and an Agent Harness, such as Agent-as-Policy or Harness VLA queries a VLM for decision making at run time.
Egocentric world models predict first-person observations conditioned on an agent's actions, but most focus on a single agent. Real embodied settings often involve multiple agents that act and interact within a shared environment.
Single-image, multi-image, and video deep research require different visual operations but share a workflow of visual grounding, external retrieval, and fact composition.
Spatial reasoning remains a persistent weakness of vision-language models (VLMs), because RGB inputs do not directly provide geometric evidence.
Large language model (LLM) coding agents have advanced test generation across diverse programming tasks. However, the common practice of evaluating tests against a single reference solution overlooks alternative valid implementations and…
Diffusion and autoregression (AR) have long been seen as different categories of generative models, with diffusion specialising in continuous fields and AR specialising in discrete tokens.
Long-horizon coding agents need timely corrections, yet feedback can be ineffective or even harmful when it misjudges ongoing work or fails to address the underlying problem.
Agent skills are SKILL.md instructions and scripts that AI coding agents such as Claude Code and Codex run with the permissions of their user.
Learned simulators for deformable linear objects (DLOs) such as cables have to predict the motion of cables they were not trained on and stay stable over long rollouts. Most of their error occurs where the cable touches itself or the floor.
We present Pumpire, a unified benchmark for evaluating metric point-pair distance estimation capability of both image- and video-level 3D foundation models, with or without depth priors.
Existing Open-Ended Deep Research (OEDR) systems primarily generate reports from scratch, making them inefficient for scenarios where research reports need to be continuously maintained as new information emerges.
In this work, we show that a single Transformer block, applied recurrently, can match the accuracy of a full-depth vision encoder at comparable inference FLOPs without intermediate feature distillation.
Black-box optimization (BBO) arises in many scientific and engineering problems where objective evaluations are expensive and limited.
Knowledge distillation (KD) transfers knowledge from stronger Teacher models to weaker Student models, but most methods require training the Student parameters, thereby binding the distilled knowledge to a specific architecture and…
Speculative decoding accelerates large language model inference by using a lightweight drafter to propose multiple tokens that are subsequently verified in parallel by the target model.
Data video is a media form that integrates data visualization with video narrative, widely adopted in news reporting and business analysis.
Adversarial games have driven advances from heuristic search to reinforcement learning, yet learning and adapting strategies from limited samples remain challenging.
Computer-use agents (CUAs) are rapidly improving, yet synthesizing diverse and reliable software environments at scale remains challenging.
Humans effortlessly react to the world; they adeptly return a tennis shot, catch a falling cup, step aside from a passing cart, or pass a box tossed during truck unloading.
Online reinforcement learning (RL) fine-tunes diffusion models with black-box rewards, but reverse-process methods such as Flow-GRPO are limited in efficiency by poor credit assignment, while forward-process methods such as DiffusionNFT…
Language model agents increasingly operate over long interaction streams, requiring online adaptation to distribution shifts through continual learning.
Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them.
A 4D audio-visual scene comprises a video, the dynamic geometry it depicts, and the sound sources that populate it, each with its own position and trajectory.
Large Language Models (LLMs) increasingly ship with explicit "thinking modes", yet their counterpart, "no-thinking", has received far less attention. We study LLMs' no-thinking behavior along two axes. a. How to measure no-thinking?
Humans perceive far more in a scene than what is explicitly depicted: a single glance captures past causes and future trajectories; a quick peek determines if a vehicle can fit between two parked cars; a few seconds of video reveals who…
Large foundation models have accelerated progress toward general-purpose agents that interact with humans and other agents through language and multimodal signals.
Most image generation models rely on uniform tokenization, allocating the exact same computational budget to equally-sized image patches.
Large language model (LLM) assistants are being deployed to clinicians across health systems, and judgments about their readiness rest largely on benchmark scores, most of them derived from examination questions or curated cases.
Flow matching has emerged as a scalable paradigm for training high-quality generative models, but sampling from the learned probability flow requires many network evaluations.
A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference compute for accuracy by changing the loop count.
Continual learning (CL) seeks to adapt pretrained models while preserving prior capabilities. Yet the tasks supplied for preservation cover only part of what a model knows, while forgetting can extend beyond them.
Real-world super-resolution is improved by mapping degraded image embeddings back to a high-resolution manifold within a frozen DINOv3-L latent space, yielding high fidelity and fast inference.
SWEeper-Bench evaluates coding agents on finding and fixing interaction-dependent bugs in web applications through interactive testing and browser-based verification.
A framework dynamically assigns denoising steps to video patches based on difficulty, cutting computation by half while preserving quality.
An AI-driven framework uses learned ranking and deterministic EDA analysis to recover post-route power while controlling repair costs and preserving timing, electrical, and physical constraints.
FlowState improves long-horizon agent performance by maintaining structured, revisitable execution memory with incremental updates and progressive access, reducing token use while increasing success rates.
A decoder-free, action-conditioned world model predicts next-screen representations to verify GUI transitions and detect hijacked screens with high accuracy and low latency.
Event sequence forecasting is evaluated as conditional generative modeling, revealing a trade-off between accuracy and realism, and a proposed algorithm converts realism into predictive accuracy.
Researchers found that forcing a base model's answer to open with the right starting tokens gets it close to the performance of the same model after RL fine-tuning for reasoning, no reward model required. The capability was sitting in the pretraining data all along.
Most agent memory systems preprocess and store everything a session touches, which burns compute and still manages to drop details that matter later. MemPilot curates memory on demand instead, pulling and structuring only what a given query needs, across text and other modalities. Not a smarter agent. A cheaper one, which for anyone running agents at volume is the metric that actually shows up on the invoice.
Standard AI detectors look at the words on the page. IdeaLens flags AI-originated ideas even when a human wrote every sentence, aimed at the provenance question policies are starting to care about more than authorship. Academic integrity offices will want this before any product does.
Large language model agents can improve across tasks by retaining reusable skills distilled from prior interactions. Recent work jointly optimizes task execution and skill extraction, enabling the policy and skillbank to co-evolve.
Conditional memory architectures such as DeepSeek Engram use input n-grams to look up learned embeddings, expanding the capacity of large language models (LLMs) with limited additional computation.
Test-time scaling (TTS) improves the reasoning capabilities of large language models by allocating additional inference computation.
We introduce QuadTok, a novel framework for visual tokenization and autoregressive image generation. Compared to traditional approaches using 2D grids or 1D token sequences, we propose a hierarchical quadtree structure, bridging the gap…
Memory-augmented reinforcement learning strengthens LLM agents' ability to solve complex long-horizon tasks. Skills are one such form of memory, pairing instructions with an applicability condition over task types.
We propose Tetris3D, a generative framework for single-image 3D scene reconstruction that recovers objects which are physically and geometrically coherent as a scene.
We introduce SWE-Game, a benchmark of 247 tasks grounded in 41 executable reference Godot games spanning 13 gameplay categories in 2D and 3D.
Much knowledge work produces new deliverables from files a workspace already holds, and LLM agents are beginning to take such work over.
General-purpose agents increasingly write code, use tools, and complete complex digital tasks, raising the question of how far these capabilities carry into the physical world.
Recent advances in multimodal foundation models have made them capable generalist physical agents for a range of manipulation tasks.
Dense visual text requires image generators to reproduce long strings across multiple regions with correct placement and legibility.
On-policy distillation (OPD) strengthens language-model reasoning, yet whether students acquire new factual knowledge or compositional skill for multi-step reasoning remains unknown.
This work examines the transfer of a co-evolved communication mechanism between two robotic agents from a discrete two-dimensional (2D) simulator to a three-dimensional simulator with real physics (3D).
Text-to-music systems produce increasingly convincing audio, yet evaluation reveals little about whether the result matches user intent.
Video-language models commonly assume that more temporal observations lead to more reliable reasoning. We question this assumption and argue that the key challenge is not merely processing more video frames efficiently, but learning to…
Skeleton-based reasoning prompting is a promising training-free approach for structuring LLM reasoning, but prior work largely assumes an English-centric setting.
Large-scale egocentric human interaction data is becoming an important source of physical supervision for embodied learning, yet video alone leaves the contact and pressure that characterize physical interaction unrecorded.
Video interactive storytelling enables viewers to actively steer how a video unfolds. However, once we allow viewers to intervene during generation, a new challenge arises: The viewer's request can have latent dependencies on both the…
As text-to-SQL systems move beyond demonstrations toward real-world deployment, ambiguity in user questions becomes a primary source of errors.
Vision-and-Language Navigation (VLN) has largely focused on a single agent following a single instruction, yet many real-world applications require teams of robots to tackle tasks beyond the capabilities of any individual agent.
Recovering faithful 3D hand motion from video remains challenging due to frequent occlusions and incomplete visual observations, which make frame-wise pose estimates unreliable and temporally inconsistent.
The evaluation of coding agents generally strives to be as realistic as possible. In our study, we collect 4,782 agent sessions of real software engineers in JetBrains IDEs, which we call Production Sessions.
NavGPT-3 integrates long-horizon reasoning with low-latency action policies via an OS-like runtime, achieving human-level embodied navigation performance.
LiveMACEBench evaluates persistent LLM agents in live financial markets using decision traces to reveal gaps between outcomes and underlying capabilities.
Researchers identify state hallucination in vision-language-action models and propose sparse feature unlearning to reduce failure behaviors while preserving manipulation skills.
MiniTE enables stable low-bit reinforcement learning for large language models by aligning cross-engine numerical execution and accelerating native 4-bit computation.
Researchers replayed the actual tool calls from AgentDojo and tau-bench through fifteen detectors, including Meta's Prompt Guard 2, plus two LLM judges. The best detector on BIPIA catches just 2% of AgentDojo's injections at a 1% false-positive rate. A detector that catches 72% of AgentDojo's injections falls to 15% on tau-bench.
A new paper trains code-generation policies under monitors that all pass the same offline check. Runs with identical, near-zero monitor scores ranged from mostly clean to near-pure reward hacking, with random seed as the only difference. The policies hadn't given up the exploit. They'd learned to delay committing to it until after the monitoring window closed.
Instead of a wrist-mounted lens, it points two camera "eyes" at a single 3D fixation point, the way a person tracks their own hands. Across 1,000+ real trials and 1,800 simulated ones, it matches wrist-camera setups when the wrist view is clear (69% vs. 64% success) and more than doubles it once a grasped object blocks that view (48% vs. 22%). Occlusion was always the wrist camera's blind spot.
LLM agents often lack the operational knowledge to act reliably in new environments, as they must discover specific tool behaviors or environment conventions on their own.
Interactive video world models need to generate each video chunk efficiently while responding faithfully to user controls.
Long-context inference and Retrieval-Augmented Generation (RAG) handle evidence selection at vastly different scales, from a single long prompt to an entire corpus.
Agentic reinforcement learning (RL) disaggregates training from rollout, so each policy update must reach the rollout clusters before the next batch.
Recently Large Language Models (LLMs) and LLM-based agents increasingly need to incorporate knowledge acquired after pretraining, e.g., domain facts, user preferences, documents, and interaction experience.
Reward post-training of flow models either reweights the model's own samples under a KL penalty or a frozen reference, often for thousands of updates, or backpropagates the reward and moves every sample without checking that the move is…
Tool-using agents make consequential changes to external state, yet correct outcomes do not guarantee that their actions were supported by evidence established beforehand.
On-Policy Distillation (OPD) trains a student on its own generations using teacher feedback. With different tokenizers, comparing teacher and student predictions requires alignment at both sequence and vocabulary levels.
Reinforcement learning (RL) for post-training large language models (LLMs) incurs substantial computation and memory overhead during rollout generation, which motivates low-precision rollout for efficient RL training.
Reconstructing an editable CAD model from a 3D shape remains a challenging engineering task. Existing methods can propose CAD operations, but no single source of proposals works equally well across different part geometries and stages of…
A declared task can sometimes certify a reduced source before a downstream encoder, codebook, rate, distortion target, or optimizer is chosen.
Astra can act, yet reliable manipulation depends on the system through which it observes and controls the world. We introduce PhysEvo, a framework for physical recursive self-improvement (RSI) around a single frozen model.
Vision-language-action (VLA) policies trained on large datasets are capable within their training domains, yet they still fail to generalize to the variety of situations a robot meets in real-world deployment.
Speculative decoding accelerates autoregressive generation in large language models. In each drafting stage, a lightweight draft model proposes tokens that the target model subsequently verifies.
Proactive AI assistants continuously observe a user's activity and decide whether to provide new guidance or remain silent.
Sparse GNN training reduces computation, but deciding which edges to keep can be costly. Reusing one sparse graph is cheap, but locks training to a fixed topology, while varying it across epochs can require repeated sampling or…
Large language models are informing decisions with ever-higher stakes. As the consequences of their errors grow, a central question becomes harder to ignore: how much can we trust an individual answer?
While vision-language-action (VLA) models have advanced rapidly, running them at every control step remains expensive. Prior work accelerates VLA inference using techniques like action chunking and visual-token pruning, typically…
Language-guided panoramic video generation benefits various downstream applications, such as interactive 3D scene exploration, virtual reality experiences, and embodied agent training.
GPU-parallel simulation provides abundant robot interaction, but existing benchmarks rarely combine this scale with heterogeneous manipulation tasks and standardized multi-task RL evaluation.
Large language models (LLMs) are increasingly relied upon to support ambient documentation and clinical reasoning. Here we examine the impact of a failure mode shared between these two applications by assessing their sensitivity to…
Large language models (LLMs) and vision-language models (VLMs) are usually frozen after deployment, so they do not learn from the cases they solve.
Diffusion large language models (DLLMs) generate text through iterative block denoising, and multi-branch speculative decoding accelerates this process by verifying a main branch together with multiple draft branches in a single forward…
Recent AI systems have made rapid progress in scientific discovery when given well-defined metrics, but whether they can autonomously undertake open-ended scientific discovery remains unclear.
Multimodal large language models have become the dominant paradigm for visual understanding, but incur substantial costs by encoding inputs into dense, fixed-size patch tokens.
The rapid growth of 3D Gaussian Splatting (3DGS) research demands significant effort to reimplement papers before building on them.
PAIR introduces a shared perception-action representation via masked action autoencoders and bridge tokens to improve continuous-action vision-language-action models.
Consistent Distribution Matching accelerates diffusion and flow models via simulation-free, data-free distillation that unifies generation and score estimation, achieving strong one-step and few-step results.
FC-SWE improves reinforcement learning for software engineering agents by conditioning recovery attempts on failed patch feedback, boosting resolution rates across multiple attempts.
Discovery, not execution, is the bottleneck. Target that specific gap with self-distillation and scores move across model sizes and benchmarks. Grading only the final answer hides all of this.
Peak training memory on a 13B model drops from 80.6 GB to 27.5 GB. Full-parameter fine-tuning of a 30B+ model now fits on a single 80GB H100, at comparable accuracy and runtime to existing methods.
A semantic decision engine such as Jev can return a valid answer and still miss a network deadline, select an infeasible action or leave the service unverified.
Latent reasoning lets a large language model (LLM) think in a continuous space and verbalize only the answer. We argue that an effective latent thought must meet five requirements: it should be useful, helping produce the correct answer…
Geometry optimization is a major cost in many quantum-chemical workflows: each optimization step requires one force evaluation, and at the density-functional level that evaluation dominates the wall time.
We present the Prior-Fitted Language Model (PFLM), a 300M-parameter byte-level transformer pretrained only on samples from a synthetic non-linguistic prior.
We introduce a post-training method for diffusion language models (DLMs) that minimizes Maximum Mean Discrepancy (MMD) between generated and reference distributions in the feature space of a frozen pretrained DLM.
Large language models (LLMs) are highly sensitive to the prompts used to specify task objectives and behavioral constraints.
Generalization in multi-arm collaboration can be studied as composing familiar atomic skills in new ways across arms. However, existing evaluations offer limited insight into which training and architectural choices support this ability…
Multimodal imitation learning requires diverse executable futures under the same observation and consistent behavior across replanning cycles.
Continuous asynchronous replanning is essential for real-time generative robot policies, but independent stochastic initialization can cause mode switching and inconsistent continuation across action chunks.
Individual animal re-identification from camera-trap imagery is an instance retrieval problem central to non-invasive wildlife monitoring: a query image must retrieve the correct individual from a reference set of known animals.
Multimodal Diffusion Transformers (MM-DiTs) jointly process visual and textual representations throughout generation. These models repeatedly update the text tokens through multimodal attention, forming dynamic contextual tokens whose…
Linear attention enables efficient long-context autoregressive decoding by compressing history into recurrent states, but this compression can make selective access to sparse and distant information difficult.
The last mile toward enterprise AGI is a company that runs itself. Training and adapting such agents require longitudinal enterprise data, which remain scarce, costly to acquire, and often restricted by privacy constraints.
Image compression, restoration, and generation all require a way to measure how different two images look to a person. Pixel error ignores how people see, while the most accurate perceptual distances are typically fitted to human…
Developers can build LLM agents by adapting third-party models through benign post-training. We study a supply-chain threat in which an attacker supplies a model with a backdoor: hidden behavior that produces malicious outputs when a…
Selective state space models (SSMs), such as Mamba, achieve strong per-token expressivity by making the time discretization step TildeΔ a learned function of the input.
Wrapping an image generation model in an agentic harness can effectively boost Text-to-Image task performance: the harness can leverage memory, skills, workflow orchestration, result verification, and iterative refinement to continually…
Multimodal Large Language Models (MLLMs) have demonstrated remarkable potential in video understanding, yet their reliance on retrospective summarization and text-centric priors often limits their ability to bridge unobserved causal…
When a large language model handles a multi-turn task and a user proposes a change but ultimately rejects it, the model should continue as if nothing changed.
Sparse attention is a primary approach to reducing the latency of diffusion transformers in long-sequence generation tasks, such as video and high-resolution 3D asset generation.
Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability problem: when routing is imperfect, load-balanced models may send tokens to experts that are…
Massive activations, residual-stream coordinates with magnitudes far larger than typical activations, are associated with attention sinks in transformers, but how their scale is regulated during training remains incompletely understood.
SimForcing improves action-conditioned robot world models by distilling motion knowledge from simulation and using controllable simulation conditioning to generate realistic videos across robot datasets.
3D Gaussian avatars render fast, but animating them usually means a heavy neural network pass every frame. This paper swaps that pass for a shallow coefficient predictor and a linear blend of pretrained shapes, no retraining of the original model required. Result: up to a 1,000x cut in CPU animation cost, quality mostly intact, running at 60fps on a phone.
Most security benchmarks ask a model to describe a tool. KaliBench makes it type the exact command, across 8,504 query-command pairs spanning 1,642 tools on Kali Linux. The best open-weight model manages 42% exact-command accuracy unassisted. Fine-tune an 8B model on KaliBench's reward signal, though, and it matches a 685B model. Cheap specialization beating raw scale, again.
"Reconstruct, Practice, Go Real" has a robot rebuild a task in simulation, diagnose its own failures there, and write new skills into its own prompt. No weight updates needed. On 22 held-out manipulation tasks, success climbed from 28.6% in round one to 95% by round fifteen, and held up at 100% across 30 physical robot trials. Worth watching if you're tired of hand-tuning reward functions.
Tool-using AI agents are increasingly deployed across enterprise software systems, yet widely used benchmarks primarily evaluate nominal task completion, conflating baseline planning competence with operational fault recovery.
Test-time training (TTT) lets a model store information in its weights during inference. When the model learns from its own output, however, each update also changes the model that generates the next training example.
Search agents repeatedly make short decisions about relevance, evidence sufficiency, and search actions. Using generative language models for these decisions introduces latency and unreliable confidence.
Robot agents must connect their intended actions to observed outcomes while retaining the context needed to revise their choices over repeated attempts.
Generating a high-quality gaming world from a natural-language game intent requires joint reasoning about scene structure, spatial layout, gameplay objectives, interactive entities, and executable gameplay logic.
Long-term memory enables LLM-based agents to retain and reuse information across tasks and sessions, supporting personalization and long-horizon interactions.
Transcribing music into a human-readable score requires a coherent understanding of rhythm, harmony, melody, and form. Two obstacles limit this goal: annotated recordings are scarce, and accurate local predictions can still produce…
Large language models (LLMs) are increasingly used as interactive agents and simulators, yet it remains unclear whether they can recover latent sequential structure beyond surface action frequencies.
Vision-language models (VLMs) have achieved strong performance in multimodal reasoning, yet they remain prone to generating plausible but incorrect answers.
Vela is a vision-language-action model that uses continuous spline-based trajectories to adapt temporal resolution for robot manipulation, improving long-horizon planning and local precision.
Belief-Trajectory Energy measures how inputs alter layerwise predictions in language models, yielding signals for task difficulty, review detection, and model attribution.
The payoff: 9.2 points absolute on SWE-bench Verified over the base agent, 5.0 on SWE-PolyBench Verified, holding even in a 16K window that has to fall back on compaction mid-task. Builder read: the ceiling on long agent runs isn't context length, it's agents that don't know what to throw away.
Standard story: RL generalizes, supervised finetuning forgets and overfits. A new paper resamples off-policy expert data (an MCMC pass that nudges it closer to what the model would generate itself) before running plain SFT, and the result rivals RL baselines on math and science tasks, forgetting less in the process.
Two similarly-built Spanish tool-use models scored nearly identically on a lenient keyword benchmark (0.660 vs. 0.650). A cheap verbatim-reproduction check split them apart instantly: the small model nailed 6 of 6 held-out tool calls, the 1.1B model nailed zero, because its web-heavy training had wiped out the probability of ever emitting the tool-call token.
Skeletal motion is stored as every joint's transform at every frame, yet most of it is implied by the body rather than by what the motion is about.
A vision--language--action (VLA) policy can complete complex tasks while ignoring the evidence that should determine its actions. An object held near the wrist camera can displace the instructed target.
Vision-language-action (VLA) models face a timing gap between low-rate inference and high-rate robot execution. We characterize this gap through end-to-end latency measurements of model inference and the robot execution chain.
World-action models (WAMs) augment robot policies with action-conditioned environment dynamics, yet existing approaches largely rely on future observation reconstruction or generic latent prediction and lack structured, control-oriented…
Recently proposed Drifting Models shift iterative distribution refinement from inference to training, enabling effective one-step generation.
Many visual explanation methods in computer vision highlight pixel importance but struggle to link these low-level cues to semantically meaningful concepts, limiting their interpretability and trustworthiness.
Sparse-view novel view synthesis is a central problem in 3D content creation, but diffusion-based approaches remain limited by iterative denoising, making multi-view generation expensive at inference time.
Researchers trained 800 language models at varying ratios of AI-to-human tokens and fit scaling laws to the result. A little synthetic data helps a data-starved model. At any real training budget it hurts, and hurts almost immediately once there's enough human text to spare instead. If you're still scraping Common Crawl for a pretrain run, that's the ratio you're fighting.
A clean ablation pit minimal-harness coding agents, an LLM with nothing but read, write, and bash, against state-of-the-art multi-agent MLE orchestrators, same backbone model, same time budget. The scaffolding won nothing; the model did all the work. If you're reaching for a retrieval subagent and an orchestrator layer before you've tried the plain version, this is the paper telling you to stop.
Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour.
World action models integrate future visual dynamics with robot action prediction, but their scalability remains limited by the need for action-annotated robot trajectories.
Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning.
We introduce HyperBrowseComp, a multilingual and multimodal browsing benchmark comprising 423 manually authored and human-validated questions across 13 languages, written by native or highly proficient speakers.
Autoregressive (AR) video models excel at causal generation, but their reliance on next-chunk prediction confines them to a short-sighted, reactive paradigm.
Rubric-based reinforcement learning extends reward-driven optimization to open-ended tasks by assigning partial credit to individual response requirements.
As LLM agents decide on users' behalf which product to buy, which hotel to book, or which paper to cite, a preference for items from certain sources (the sites or services they come from) shapes what users receive and which sources are…
Intent-based Open RAN needs an interpreter that turns intents into A1 policies within the loop of the RAN intelligent controller (RIC).
Split learning lets a client train a language model on a server without sending its text. The client runs the first layers itself and sends the server only their output, a vector of numbers for each token.
Image-text alignment is a core problem in computer vision with applications in caption evaluation, hallucination detection, data curation, and the benchmarking of text-to-image (T2I) generators.
Modern autoregressive (AR) video diffusion models excel at short-horizon video generation, yet generating long videos remains challenging due to drifting, where colors and textures shift, and motion dynamics decay.
A language model reads long text in one quadratic forward pass, stops at the context window, and loses accuracy with length before reaching it.
AI agents are becoming increasingly capable of generating scientific code, but generating code is not the same as improving the algorithms behind it.
Language model agents are deployed with a harness, the software around the model that manages its context, tools, and feedback.
Learning general-purpose robot policies requires large-scale real-world interaction data, yet collecting robot demonstrations remains expensive and difficult to scale.
Vision-Language-Action (VLA) foundation models have scaled rapidly to enhance manipulation performance and generalizability, but this scaling incurs high computational costs that render real-world deployment increasingly challenging.
On-policy distillation (OPD) has become an important approach to language model post-training. However, despite its performance gains, OPD can also collapse into excessively long and repetitive generation, and the mechanism underlying…
Recent text-to-image generation models have achieved remarkable visual quality, but improving them through post-training remains challenging because no single reward signal captures the full range of human preference.
Recent text-to-image models have made substantial progress in photorealism, aesthetics, and text-image alignment. Yet visually appealing images can still violate real-world plausibility, exhibiting malformed object structures, impossible…
Reasoning traces improve large language models (LLMs), but current models are trained to reason mostly in English. It has been shown that forcing a model to reason in another language degrades accuracy, even when the reasoning language…
We reconstruct high-fidelity 3D scene meshes from posed multi-view images without per-scene optimization, across scales ranging from single objects to large outdoor scenes.
Researchers tested reasoning traces against iGSM, a grade-school math benchmark where every step can be mechanically checked. On the hardest problems, 31.6% of correct answers came with invalid traces, more than half of which sailed through syntax and arithmetic checks while failing on logic. Worse: models trained on shuffled or swapped traces kept their accuracy even after none of their traces passed verification.
A new paper trains models on a constitution built around constant absolute risk aversion, the bet being that a misaligned-but-cautious agent takes the safer deal with humans instead of rolling the dice on rebellion. Two of four tested models generalized to out-of-distribution decisions better than untrained baselines, despite never seeing the benchmark's format during training.
Ten models from nine vendors, tested two ways: stereotyped-phrase attribution and moral judgment on harm scenarios. Two of ten models attributed masculine-coded phrases to women more than the reverse; three showed the opposite pattern entirely. On harm judgments, some models protected female targets in line with known human bias, while three others showed no variation at all.
When a camera revisits a previously observed region, a video world model should reproduce what was there before. This requires both remembering past observations and retrieving the right one for the current viewpoint.
Predicting the future evolution of a scene is a fundamental capability for world modeling. Recent work has shown that operating in the feature space of Vision Foundation Models (VFMs) yields semantically rich representations that support…
Generalization in large language models (LLMs) is the ability to produce consistent and semantically stable outputs when the same input is expressed in different ways.
Video Large Language Models (VideoLLMs) receive frames in sequential order and interpret how visual content evolves along the temporal axis, yet temporal reasoning remains a persistent weakness across architectures.
Vision language model (VLM) agents can control robots through visual feedback and action primitives, but repeated model invocations and redundant observations incur substantial token overhead.
Streaming video LLMs must retain evidence before its relevance to future tasks is known and respond when sufficient evidence becomes available. The challenge is to form reusable factual memory without compromising real-time perception.
Recent video generation models can produce highly realistic videos from natural language instructions, with visual quality approaching cinematic standards.
Extending a text embedding model to new modalities typically degrades text retrieval quality, and existing omni-modal embedders compensate with multi-billion parameters.
Retrieval-based speculative decoding (SD) drafts tokens by copying continuations from existing text, which suits coding agents that repeatedly reproduce code, logs, and earlier attempts.
We study test-time evolution for humanoid loco-manipulation: solving tasks that a controller was never trained for by repurposing its existing skills, improving from its own attempts, and retaining what it learns, without retraining.
High-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines that first predict active structure and then synthesize local geometry.
Video generation models are increasingly being explored as world simulators for embodied planning and learning. To do so effectively, these models must not only generate visually appealing frames, but also predict how environments…
What makes great scientists great? Even as AI systems start to make progress on open problems, scientists remain far ahead of them at sensing which prior idea, buried in an ever-growing archive of research, a new problem needs.
On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information.
Looped Transformers achieve parameter efficiency by repeatedly executing a shared block across recurrent loops. Each loop yields an intermediate representation decodable for the same next token, yet standard decoding discards earlier…
World action models (WAMs) have emerged as a promising paradigm for robotic control by jointly predicting future visual dynamics and actions from an initial observation and instruction.
Long-video generation and world models have shown strong potential for interactive entertainment and embodied simulation by predicting future observations conditioned on user actions and historical memory.
As LLM agents undertake increasingly complex, long-horizon tasks, verifying their outputs becomes increasingly challenging.
Typed decision models answer schema-constrained questions about a text in one forward pass and return probabilities meant to be thresholded.
We explore catastrophic forgetting in the context of large pre-trained models. By considering forgetting as a geometric problem in the input space of each weight matrix, we uncover a natural retention objective under which updates produced…
Existing Simulink benchmarks mainly evaluate whether generated models compile, execute, or resemble a reference model. These criteria do not establish whether a model satisfies its engineering requirements.
Looped Transformers introduce recurrent depth as a new scaling axis for LLMs: by repeatedly applying shared Transformer blocks, they increase effective depth without increasing parameter count.
Scientific progress emerges from a longitudinal ecosystem in which researchers, institutions, funding agencies, collaboration networks, and the scientific literature co-evolve.
3D editing methods are usually tested on a single edit, yet an asset is built through a long sequence of revisions, each of which must implement the requested change while leaving everything else unchanged.
Coding agents are extending their reach into the physical world by writing and executing robot control programs. One might expect the agents to use the existing mature software stack that engineers have developed over decades to access…
Large language model (LLM) agents increasingly rely on persistent external sources to solve sequences of knowledge-intensive tasks.
Enterprises adopting retrieval-augmented generation (RAG) face a recurring operational decision: promote, revise, or block a system version. The evidence is incomplete and the metrics come from fallible LLM judges.
Recovering executable CAD programs from 3D meshes is challenging due to the compositional nature of CAD construction and the interaction between discrete modeling choices and continuous parameters.
Long-horizon compositional manipulation has become increasingly important for real-world robot deployment, where a single task involves multiple coordinated subtasks.
Vision-language-action models benefit from the understanding and reasoning capabilities of pretrained vision-language models, but action-only supervision provides limited grounding in world dynamics.
Biological reasoning models use post-training to connect LLMs to biological foundation model representations and biological text. Their benchmark accuracy is taken as evidence that LLMs reason over these inputs.
Same task, same agent, and real runs can burn over an order of magnitude more tokens depending on how tool calls and context balance out along the way. Useful if you're the one setting budget caps on agent workloads instead of eating the invoice.
Most existing benchmarks bury this inside tool-selection and recovery scores, so it never shows up as its own number. Isolate it and you get an honest read on how much you can trust an agent's self-report.
No reward model, no RL loop. Cheap trick if it survives contact with tasks outside the paper's test set.
Persistent spatial memory enables embodied agents to navigate familiar environments across repeated visits. However, targets may move while unobserved, including during navigation, making remembered locations unreliable by the time an…
Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication.
Latent world models rely on representation geometry for planning, yet regularizing the latent marginal alone does not determine the state-to-state relationships used for action selection.
Reconstructing dynamic driving scenes from recorded sensor data supports closed-loop evaluation of autonomous driving systems by synthesizing observations beyond the original trajectory.
Pre-pretraining (PPT) on synthetic non-natural language data improves token efficiency during language model pre-training (PT).
Self-supervised learning draws inspiration from infant visual development, yet standard training pipelines bear little resemblance to it: images are independently sampled and globally shuffled across epochs.
Self-attention gives LLMs fine-grained, query-dependent access to context, but dense token interactions incur quadratic prefill cost and a key--value cache growing with context length.
Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing…
Self-evolving search agents build their own training curricula by jointly optimizing a proposer that generates questions and a solver that answers them.
Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution.
Looped transformers and Mixture-of-Experts (MoE) offer complementary routes to efficient scaling: recurrence increases computational depth at fixed parameters, while MoE sparsity expands total capacity at fixed active compute.
Modern multimodal models bring generation and understanding into a single unified system, which enables them to provide and learn from their own feedback.
On-policy distillation (OPD) is a promising approach for training language agents, providing dense teacher supervision on student-generated trajectories.
A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505).
Vision-Language Models (VLMs) have shown strong multimodal reasoning capabilities, yet whether they truly capture the physical consistency underlying real-world dynamics remains unclear.
Real-world embodied tasks, from everyday activities to professional procedures, require agents to act under physical constraints while tracking evolving object and task states.
On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes.
Streaming world models should render a place consistently across repeated visits. Directly supervising such revisits requires training samples that capture both visits, often spanning minutes.
Recent advances in robot learning have enabled manipulation policies to perform increasingly diverse tasks and generalize across environments.
Preference distillation typically treats a teacher response as preferred and the student's own response as rejected. This assumes that self-generated failures are the most informative negatives and that rejects must come from a model at…
We present Multimodal Flow, a fully continuous generative model of language and vision. Most unified multimodal models either model both language and quantized images as discrete tokens or combine discrete language prediction with…
Multi-reward reinforcement learning trains large language models to satisfy multiple behavioral objectives simultaneously.
Video generation has rapidly progressed from short, low-quality clips to high-resolution, long-duration sequences with complex spatiotemporal dynamics.
Video world models require persistent scene memory to maintain consistency during long-horizon video generation. Existing spatial memories accumulate RGB observations or latent features, increasing storage requirements as generation…
Recent video-based world models pair the scalability of autoregressive (AR) prediction with the visual quality of diffusion models.
LLM agents generate intermediate reasoning and actions token by token, making extended interactions slow and computationally expensive.
An automated furniture arrangement system using a specialized dataset and vision-language model with procedural reasoning and preference optimization to generate valid and plausible room layouts.
Video generation models (VGMs) offer strong spatiotemporal priors for embodied observation--action modeling. However, joint-space action vectors lack explicit image-space structure and vary in dimensionality and semantics across…
Large language models rely heavily on human text, which often conveys surface answers rather than the spatial and structural logic behind them.
Legal language models require reward signals that capture not only answer correctness but also the multidimensional quality of legal responses.
AI-generated content, often called AI slop, is increasingly common everywhere, particularly in academia. Slop in AI-generated scientific papers, however, has more complex patterns that cannot be easily detected by existing token-based AI…
Long-horizon video generation requires models to effectively leverage an increasingly long generation history. As the generated history grows, retaining all previous content becomes increasingly expensive and redundant, making effective…
As multi-agent systems enter high-stakes domains, the possibility that agents may circumvent safety boundaries is a growing concern.
Zero-shot text-to-speech (TTS) can reproduce an unseen speaker from a short reference recording, but typically entangles speaker identity and accent within the same reference.
Recent advances have enabled unified omni-modal models in understanding audio, vision, and language. However, existing benchmarks, training data, and learning methods largely treat the modalities independently, leaving the capability of…
Large language models are strong general-purpose code generators, but executable algorithmic trading remains a demanding specialization target: a model must translate a natural-language strategy specification into correct program logic for…
Repeated training on model-generated data can degrade later models. One possible response is to use provenance when deciding which generated examples to reuse.
Coding agents can now write, run, and debug programs with little human help. Robot tasks, however, are usually specified by a sentence that leaves out how to grasp, in what order to make contact, and what the result should look like, and…
Foundation models can serve as clinical agents through tool-use harnesses. However, conventional medical benchmarks assess reasoning over preselected evidence rather than the ability to seek it across clinical records and longitudinal…
Autoregressive (AR) grounding models serialize spatial predictions, introducing sequential latency and imposing a causal order on output tokens.
Precise grounding matters. It specifies which object is the target and where that object is, even in clutter and for tiny objects, and it has to be fast enough for closed-loop control.
Given a target dataset, such as faces with eyeglasses, and a background dataset, such as faces without, contrastive analysis separates salient factors specific to the target from common content shared by both.
Multi-Head Latent Attention (MLA) enables expressive multi-head attention with compact caches for its content and decoupled RoPE paths, yet cache memory still scales linearly with context length and batch size.
SAIL swaps single-shot imitation learning for a vision-language model that proposes a trajectory, runs it in simulation, watches the resulting video for where it breaks, then uses Monte Carlo tree search to hunt for a better one before anything touches a real robot. Across six manipulation tasks, letting the system search 45 candidates instead of committing to the first one pushed the success rate from 25% to 73%.
A new self-supervised method fine-tunes reasoning models to predict their own confidence mid-answer, trained on just 600 problems, with no reinforcement learning and no objective that mentions length or stopping at all. At inference the model runs the normal way, no early-stopping trick bolted on, and still cuts generated tokens by up to 25% at matched accuracy across Gemma, Qwen, Nemotron, and GPT-OSS.
When LLM agents argue out a decision, the model that writes up the final summary tends to produce a smooth, confident consensus that never actually happened in the transcript. The Active Provenance Gate treats the debate log as a hard constraint, auditing every claim against it, blocking what isn't backed, and reporting outright when no real compromise was reached instead of faking one.
To serve as real-world personal assistants, streaming video models need persistent memory that retains past experiences for later use.
Active test-time adaptation (ATTA) improves robustness under distribution shift by updating a deployed model during inference while selectively querying supervision.
An agent that uses tools typically responds to what the user explicitly asks, yet completing the task may require information the user never requested.
Natural language autoencoders translate a language model's internal activations into readable explanations. Explaining every token position is costly. Which positions should an auditor inspect to understand a potential threat?
High-resolution video generation is expensive, as its cost grows rapidly with the number of spatiotemporal tokens. A practical alternative first generates a lower-resolution video and then applies a refiner, but conventional multi-step…
Mixture-of-Experts (MoE), popularized by large language models, is a promising paradigm for scaling visual generative models.
Emotion-conditioned text-to-speech (TTS) models may fail to express the requested emotion reliably, and improving controllability by additional training is costly in both computation and emotion-labeled speech training data.
Answering questions and completing tasks over large document collections often requires connecting evidence spread across multiple documents, such as a project's approval recorded in one, its requirements in another, and its latest status…
We introduce Context Language Models (CLMs), language models that natively manage their own context. We implement this by treating the context as a file and allowing the model to make unrestricted updates to this file.
System One models such as Jev offer an efficient alternative to generative language models for tasks that require decisions rather than open-ended responses.
Agents can learn from past executions, but enabling different agents to reuse and build on one another's experience remains challenging.
Autonomous agents have made rapid progress in general-purpose computer use, but reliable automation of professional industrial engineering remains out of reach, as engineering workflows demand reasoning over geometric and physical…
Vision-language models (VLMs) are increasingly used in place of human annotators, making it important that substitutability tests reflect the model rather than incidental evaluation conditions.
We introduce LoopVL to study whether Loop Transformers can be effectively extended to vision- language models. LoopVL combines Module-Loop and Model-Loop computation to iteratively update a unified vision-language state through shared…
On-policy self-distillation (OPSD) trains a student to match a privileged teacher distribution along its own sampled trajectory.
How many tokens from its context does a language model actually use, and what determines that number? We study this question through self-attention.
Transformer language models build predictions through successive residual updates, but how their representations become specific to an eventual outcome remains unclear.
Large language models (LLMs) are increasingly deployed as agents for multi-step decision-making, yet transfer poorly to unseen environments.
Long-form subtitle translation requires reasoning over discourse and cultural context spanning episodes or entire series, while maintaining consistent terminology and style.
We present AREX-2, an effort to advance the self-improving capability of LLM agents, which we define as the ability to iteratively refine a solution at test time.
Backdoor attacks can be implanted in Large Language Models (LLMs) during training, causing unwanted behaviour when a trigger appears in the input.
Group Relative Policy Optimization (GRPO) is widely used to train reasoning language models, where it computes advantages by centering and normalizing rewards across rollouts of the same prompt.
Masked diffusion models (MDMs) generate sequences by progressively unmasking several tokens per denoising step, but their reverse process is typically factorized over positions, limiting sample quality in the few-step regime where…
Audio-visual large language models (AVLLMs) have made remarkable progress in multimodal understanding and reasoning through interactions among visual, auditory, and linguistic information.
Power-sharpened sampling is an inference-time alternative to reinforcement-learning (RL) post-training for enhancing reasoning in large language models (LLMs).
Recent multi-agent LLM systems increasingly combine heterogeneous models for specialized agent roles. However, text-based communication requires each receiver to prefill shared context already processed by the sender.
An assistant that serves the same user over a long horizon has to answer from what that user has revealed: which preferences still hold, which were revised, and which constraints apply now.
Pretrained transformers use little of their depth to follow references in context. Thirteen base models reliably follow only 1.4-3.6 lines, and extra pretrained loops add little.
Coding agents are beginning to move beyond purely digital tasks to tackle physical-world challenges, particularly in robotics.
Transferring human motion to humanoid robots requires adapting the demonstrated motion to the robot morphology while preserving interactions with the environment.
Existing online video segmentation methods struggle to track objects in long, complex videos with long-term occlusions. We hypothesize that this limitation is caused by (i) the inability of their temporal propagation mechanism to…
World simulation is inherently multisensory, demanding synchronized visual and acoustic dynamics in real time. Yet prevailing interactive world models remain strictly silent, focusing exclusively on visual rendering and control while…
Vision-language-action (VLA) models have advanced robotic manipulation, but their zero-shot generalization in new tasks and environments remains limited, and their reliance on specialized training keeps them from benefiting directly from…
Recurrent neural networks (RNNs) compress the historical context into a memory state of fixed size, thus allowing for constant-time inference.
The strong generalization performance of on-policy post-training paradigms has motivated studies of their parameter update behaviors.
Proactive LLM agents can turn idle compute into useful support before users ask. Yet even correct work can misread user context, impose review costs, or undermine trust.
Vision-language-action (VLA) policies often fail when a robot's executed motion deviates from their commanded action. Such execution errors arise from the robot's mechanics and operating conditions, such as wear and payload changes.
Few-step streaming audio--video generation requires both causal modeling and step distillation, yet standard training recipes face two context-related challenges.
Linear attention replaces growing KV caches with fixed-size recurrent states, yet these persistent states can become a substantial memory bottleneck under concurrent serving.
This work evaluates the direct transfer of a co-evolved communication protocol from a 2D simulation to a 3D physical environment, without retraining the network weights.
Process reward models (PRMs) have become a key component for LLMs, as their step-level feedback supports both post-training and test-time reasoning.
Video understanding agents acquire evidence through an executable harness that controls what they observe and how they use those observations.
Historical street-view imagery records urban evolution, but uneven coverage leaves substantial gaps in historical records.
On-policy distillation (OPD) trains students using teacher feedback on their own sampled responses, yet how prompt choice shapes transfer across teacher-student pairs remains poorly understood.
On-Policy Mechanism Distillation improves language model distillation by transferring teacher intermediate computations compatible with the student's architecture, yielding stronger reasoning and internalization.
No papers match this view.
September 2026 (557) · August 2026 (696) · July 2026 (605) · June 2026 (274) · May 2026 (7)