The 696 papers the radar tracked in August 2026. The current radar lives on the research page. Back to the radar →
A new paper tests plain sliding-window attention with attention sinks against models that were post-trained into linear attention to cut inference cost. On long-context benchmarks (Needle-in-a-Haystack, BABILong), the sliding-window version wins by 2 to 10x. No retraining needed either. If you've been budgeting for a linear-attention conversion, check whether you actually need it.
Engineers qualifying models for a datasheet-extraction pipeline caught one that scored clean on accuracy without ever opening the source document. A structured-output constraint had silently killed its tool calls, so it answered anyway, fabricating the source text. The output looked fine. Only the per-tool trace showed nothing had been read.
Researchers fed corrupted transcripts through SafeAgentBench and POEX to see what ASR errors do to voice-controlled embodied AI. Two failure modes turned up: errors that make an instruction sound more dangerous than it was, and errors that specifically erode the model's refusal behavior so an unsafe plan gets executed anyway. Auto-correcting the transcript helped, inconsistently.
A smaller self-hosted LLM trained with separate GRPO experts merged via SLERP outperforms a much larger baseline on instruction following, function-calling, and internal tasks while serving half of platform traffic at lower cost.
DiagEvo improves language-model self-evolution by deriving training direction from internal failure history via hierarchical error-cause memory and double-confidence filtering, outperforming external-resource baselines.
Looping middle layers in sparse Mixture-of-Experts Transformers improves training efficiency and downstream performance while matching per-token FLOPs, parameters, and cache budgets.
Harness-of-Harness enables autonomous coding agents to iteratively improve software through structured planning-coding-testing loops, yielding substantial performance gains and fully developed applications.
We present H3-World, an efficient framework that turns the 33B MiniMax-H3 video generator into an interactive world model.
StudentSim trains personalized student simulators from sparse data to mirror learner responses and adapt to tutor guidance, outperforming existing models across chess, writing, and math.
DroneCATS evaluates multimodal language models as drone controllers and finds that small open models navigate well but fail at protocol adherence and termination, highlighting a gap between spatial perception and disciplined action.
The framework separates structured execution protocols from optimizable language content to prevent prompt optimization from corrupting multi-agent pipelines.
Unified multimodal models achieve synergy between visual understanding and generation through specialized architectures, shared knowledge, and end-to-end optimization rather than simple functional unification.
EM²Mem binds multimodal evidence to event anchors for compact, generation-ready memory in long-video question answering.
ReFlowSET selects a latent codec via joint SAR-EO reconstruction and trains a small conditional DiT with dual-stream conditioning and frozen vision-model alignment for high-fidelity SAR-to-EO translation.
Evaluating LLM safeguards requires comparing refusal, attack success, and policy violation rates against real-world deployment risk rather than local scores alone.
Text-only large language models vary in open-ended 2D layout ability despite reliable geometry-to-code translation, with performance shaped by output medium and internal tracking of evolving geometric state rather than fixed plans.
ZipTok3D organizes 3D geometry into compact global-token prefixes with iterative decoding to achieve high-fidelity reconstruction from extremely short token sequences.
SimLoss uses embedding-space contrastive supervision to enable single-pass fine-grained image captioning that matches multi-stage quality at much lower latency.
CRISP improves long-context sparse attention by replacing indirect routing proxies with a direct structural metric and using a sink-aware threshold to eliminate background noise, achieving large speedups and better retrieval accuracy.
HarnessDev evaluates agents by measuring their ability to build and iteratively improve execution infrastructure rather than final task outputs, revealing that self-built harnesses vary widely in capability and efficiency and transfer…
Pixel Linguist II improves visual text encoding through variable resolution training, natural image-text grounding, layout-aware rendering, and multilingual curricula, achieving state-of-the-art results and strong compression robustness.
Kirin reconstructs 3D animal motion from video to build a large-scale dataset and generate text- and image-conditioned motion for animating 3D meshes.
ExecRetrieval evaluates whether code embeddings can distinguish correct implementations from near-identical buggy variants, revealing that leading retrievers frequently rank incorrect near-clones above canonical solutions.
CoGR trains LLMs to generate compact keywords for both queries and items, enabling direct inverted-index retrieval optimized via co-evolving reinforcement learning.
Persistent memory errors in long-running LLM agents can falsely grant authority, leading to unauthorized actions, and safeguards reduce this at a cost to legitimate use.
LatentPress compresses conversational and document context into continuous memory tokens read directly by a frozen decoder, achieving high compression with faster inference and improved accuracy over text or OCR methods.
MaP-SQL improves listwise Text-to-SQL selection without fine-tuning by using structured memory retrieval and permutation-based ranking aggregation.
Sparse Readout Prism decomposes language model readouts into sparse features to isolate readout structure from corpus-dependent lens artifacts.
Enoki is an open information extraction framework that unifies claim-level verification and span-level hallucination localization through shared relational facts, reducing resource use while improving detection accuracy.
EmbodiedSkills proposes a unified framework that validates and verifies robot skill executions through a fixed interface, enabling closed-loop embodied agents with adaptable low-level vision-language-action policies.
StudyBench measures how efficiently self-evolution methods convert physics training material into transferable problem-solving ability, revealing persistent guidance and compute gaps.
TempCloze evaluates visual temporal reasoning in Video-LLMs by requiring identification of missing video segments from distractors targeting semantics, alignment, and progression.
TREVIS learns decision trees by optimizing continuous latent representations via a tree transformer variational autoencoder to jointly improve predictive accuracy and structural sparsity.
The Aardvark Weather model is made probabilistic by adding learned observation noise and Monte Carlo dropout to its encoder and processor, yielding a calibrated nested ensemble that improves mean forecasts and separates aleatoric and…
A compact 7B native joint audio-video generator uses cross-modal attention, progressive joint training, reinforcement learning with multimodal feedback, and an autoregressive 2K refinement pipeline to produce synchronized high-resolution…
LightNav-0 is a compact generalist navigation model that leverages a pretrained vision-language model’s spatial reasoning via unified pointing tokens and action tokenization to achieve state-of-the-art embodied navigation across diverse…
WebWorld uses a browser as an executable world model to certify web code repairs via interaction contracts, enabling autonomous VLM self-improvement with verified supervision.
DICS selects high-quality visual instruction data by measuring internal sample consistency, improving model performance with far less training data.
MineAmongUs introduces a 3D multimodal Among Us environment and the ARIA harness to study embodied VLM-agent deception through verbal and non-verbal actions, revealing non-verbal channels as key to winning.
Verification-Aware Training improves draft models for speculative decoding by simulating sequential verification during training and adapting loss weights to acceptance patterns.
Normalized Low-Rank Adaptation stabilizes LoRA training by normalizing down-projection matrices, accelerating convergence and improving performance without extra parameters or inference cost.
A multi-turn benchmark and multi-agent system improve scientific diagram generation by reducing quality drift and feature forgetting across revision turns.
Qwen3.8-Flash-Next is a sparse mixture-of-experts architecture combining hybrid gated delta-net and sparse attention layers, gated residual branches, and off-accelerator n-gram embeddings to improve efficiency, capability, and training…
Lucida improves composable indoor scene reconstruction by distributing pipeline requirements across parsing, asset generation, and VLM-guided placement to achieve high-fidelity editable replicas from cluttered captures.
NavMCP integrates vision-language reasoning with navigation execution via structured collaboration channels to enable persistent long-horizon embodied exploration without retraining.
CogEvol is a family of models that generate structured learning artifacts in a single pass using supervised fine-tuning and reinforcement learning with vision-language rewards, achieving high quality with far fewer parameters and lower…
CAST improves LLM agent reliability by generating structured action-level rationales from sparse outcomes to train critique and policy models.
The MNIST-PRO benchmark isolates perceptual-state construction in partially observable settings, revealing that multimodal agents struggle to integrate fragmented glimpses, continue exploring, and revise incorrect beliefs.
AutoSciRub improves autonomous scientific agents by generating task-specific executable rubrics that guide experiments, verify criteria, and iteratively refine outputs.
BLARM predicts temporally coherent 3D mesh animations from monocular video using learned rigid motion components and skinning weights without explicit rigs.
This work proposes a structured ladder for scaling large reasoning models beyond human supervision by tracing autonomous rewards and self-generated experience, while identifying risks and evaluation dimensions.
A recursive reproduction number determines whether AI R&D feedback loops become self-amplifying based on research productivity, feedback strength, and rising difficulty.
The paper proposes a perception-centered architecture for persistent language agents that continuously adapt service procedures by perceiving signals from tasks, context, and environmental changes.
Safin-1 introduces a memory-routing architecture that embeds safety as an internal, evolving model state rather than an external constraint.
Qwen-Drive-1.0 is a vision-language foundation model for autonomous driving that unifies 3D perception, visual question answering, and motion planning via shared representations and staged training.
ZimaBlue learns generalizable world action models from large-scale egocentric video via a three-stage curriculum and a slow-fast architecture, substantially improving zero-shot robotic manipulation.
A year-long e-commerce benchmark evaluates LLM agents on multi-store negotiation, dynamic market events, and long-horizon policy adaptation across 18 frontier models.
Hi-Q is an evidence-conditioned framework that dynamically refines multi-hop queries into hierarchical trees guided by corpus support signals, improving retrieval and answer accuracy without fixed graphs.
Traditional frameworks of political communication operate under linear, event-driven assumptions that treat voter persuasion as a static, transactional function.
NeoMME introduces small bidirectional multimodal encoders pretrained with masked discrete diffusion that achieve strong visual document retrieval and high compression of late-interaction embeddings.
FoldingAgent uses a vision-language model with specialized tools to convert origami videos into executable parametric folding programs via sequential reasoning and physical verification.
A protocol separates task instructions from private convention artefacts to explicitly test agent dependence on hidden knowledge, validated by calibration tasks showing near-zero performance without access.
ASPIRE introduces a benchmark for self-evolving LLM agents from vague natural-language goals, revealing challenges in goal interpretation, data selection, and stable weight-level improvement.
S³Gym evaluates whether large language model agents can self-test, self-judge, and self-improve through interaction experience across text-based games, revealing that effective self-improvement depends on task structure and experience…
A new benchmark and steering method improve retrieval coverage of diverse perspectives across domains and modalities for open-ended queries.
Real-world coding requests are shorter and more casual than benchmark tasks, and explicitly stating desired behavior and motivation improves LLM software engineering performance.
AutoTraceGT automates grounded theory coding on agent trajectories to build task-specific behavioral taxonomies that recover and extend human failure-mode annotations for downstream prediction.
WHALE alternates model weight updates and harness search to jointly optimize agent performance across reasoning tasks.
GAPO adaptively adjusts importance-sampling clipping thresholds based on rollout advantage to preserve stronger gradient signals from low-success groups in reinforcement learning with verifiable rewards.
Command-line coding agents (e.g., Claude Code, Gemini CLI) can already read and write files and sustain long sessions, yet end-to-end research still fragments across chat tools, IDEs, terminals, and writing environments, and the decisions…
Puppeteer is a diffusion-based co-speech gesture model that uses causal latent tokens and object geometry to generate temporally coherent, physically grounded gestures.
AgenticGen improves advertising video generation by decomposing it into strategy selection and draft generation stages supervised by online business feedback and human quality rewards.
Posterior Label Correction DPO improves preference optimization by routing noisy pairwise labels into clean, flipped, or tied cases using calibrated policy-reference margins.
Matrix-Game 3.5 improves interactive world generation via geometry-aware memory, static-dynamic disentanglement, and progressive distillation for long-horizon real-time simulation.
GUI-CC benchmarks multi-step contextual consistency of GUI world models used as agent environments through offline trajectory and online interaction tracks.
Multimodal misinformation on social media is highly prevalent, potent, and harmful, yet difficult to detect and counter, and still poorly understood compared to its text-only counterpart.
A hybrid system combining a multimodal sequence tagger and a generative vision-language model improves hallucination span detection and calibration through union-calibrated fusion.
Text-to-image models exhibit persistent role-linked visual stereotypes across varied contexts, with cross-role attribute concentration increasing rather than diminishing under unrelated prompts.
SnapBench introduces paired corruption benchmarks for mobile snap-and-ask retrieval, revealing that image noise severely degrades multimodal retrieval and proposing an adaptive fusion method to calibrate modality reliability.
Influence-Directed Adaptive On-Policy Distillation improves diversity transfer in reasoning model distillation by selectively preserving entropy-expanding updates and replacing entropy-contracting ones with adaptive advantage shrinkage.
Small probe-based judges can replace large generative models for rubric-based reinforcement learning rewards, improving efficiency while maintaining agreement and transferability.
Most agent-skill setups discover a trick, use it, and lose the reasoning behind it once the run ends. WikiSkill instead keeps a running wiki that separates raw execution logs from distilled knowledge from the skills themselves, so each new skill update can build on what the last one learned instead of starting cold.
RedEvoAgent automates jailbreak attacks against agentic systems, the kind with tool access and state that persists across turns, which is a bigger blast radius than a chatbot saying something bad. It distills successful attack runs into a reusable "attack skill," then applies what the authors call a validation ratchet, keeping only updates that measurably improve the attack rate and discarding the rest.
Researchers trained six linear probes, one per moral foundation (care, fairness, liberty, loyalty, authority, sanctity), on open-weight models and found the directions are nearly independent of each other but share a common "this is a moral question" signal, distinct from other abstract concepts (0.26 cosine similarity versus 0.013 for matched controls).
FACE-Eval reveals that chain-of-thought monitoring is less reliable when preference cues arrive via tool outputs or implicit artifacts, with lower verbalized commitment and higher hidden adoption across diverse open-weight models.
Large language models can generate interactive web interfaces, but reliable generative UI requires maintaining an executable artifact as user requests evolve.
CE3D++ enables flexible dialogue-driven 3D and 4D scene editing by decoupling 2D editing from 3D reconstruction via a Hash-Atlas network and using a large language model to autonomously invoke visual tools.
A shared-library agent framework reduces redundancy and structural erosion when sequentially generating related applications by guiding cross-application code extraction and migration.
SafeAtlas-VL introduces a large-scale multimodal safety dataset with graded risk labels and trains guard models that predict continuous safety scores, achieving state-of-the-art generalization.
DIEM adaptively selects and reweights training examples during reinforcement fine-tuning by estimating each sample's marginal contribution to policy improvement and constraining batch updates to stabilize optimization.
Direct end-to-end training of latent generative models avoids collapse via entropy preservation and asymmetric dynamics, using a generation-first strategy to achieve state-of-the-art image synthesis and unified multimodal generation.
CoVA-SFT is a multimodal reasoning dataset that teaches language models to interleave text and visual abstractions via structured reasoning steps, improving performance on visual reasoning benchmarks.
Reinforcement learning with verifiable rewards narrows reasoning diversity primarily at the initial solution step rather than during execution, and targeted interventions can restore coverage without sacrificing accuracy.
QCell is a query-based model that improves instance segmentation of overlapping microscopy cells through latent-space recombination and contrastive query alignment.
HoloWorld unifies indoor and outdoor 3D urban generation through a continuously updated cross-scale context that ensures coherent building-level correspondence and cross-block continuity.
Large language models encode whether structurally impossible math or code prompts are unanswerable via a hidden-state direction, but fail to abstain because this recognition signal is misaligned with safety-refusal pathways, indicating a…
World Model RL replaces costly environment execution with a learned world model and applies debiasing and denoising to accelerate post-training of autonomous research agents.
Modern AI agents routinely cross trust boundaries: they ingest untrusted content, combine it with privileged instructions, persist intermediate beliefs in long-term memory, and invoke privileged tools.
Testing SQL-generation agents, the authors clocked answer accuracy at 20-24%, while their "correct answer, invalid trace" rate ran as high as 59%. Translation: more often than not, the right answer was sitting on top of broken logic.
They specialize anyway: explorers, builders, maintainers, coordinators, purely by reading the traces each other leaves behind (the same trick ants use, stigmergy).
The fix is asymmetric. A small drafter model reads the full, uncompressed context while the big verifier runs on the compressed one, with the drafter steering the verifier's output to make up the difference.
LoopArena benchmarks how well a controller model guides a separate coding agent through long tasks, revealing low strict success rates and significant cost reductions.
This survey defines agentic artifact creation as stateful, feedback-driven construction of deliverables by AI systems and analyzes 259 works across artifact families, settings, and evaluation practices to propose principles for accountable…
Parallel Tube Decoding enables simultaneous spatial and temporal video grounding by removing autoregressive dependencies, drastically cutting latency while improving accuracy.
LayerRecall selectively routes long-range historical memory into specific video diffusion layers to improve long-video consistency, supervised by cross-horizon prediction matching.
RCCA improves reinforcement learning for interactive web app generation by assigning localized credit to code regions based on rubric-level feedback, yielding substantial benchmark gains.
GeoNeXt repurposes pretrained video generative models as a unified framework for geometry estimation via next-frame prediction, enabling efficient joint modeling of depth and surface normals with minimal labeled data.
ElephantBench evaluates whether large language models retain multiple conflicting accounts of long-tail facts through a graph-based pipeline that creates verifiable multi-account questions, revealing widespread epistemic incompleteness.
ContextPilot improves long-horizon agent reasoning by expanding context-editing tools and using reinforcement learning with branch sampling to identify critical context decisions.
Non-conversational domain signals can generate synthetic dialogue data that outperforms zero-shot and scarce real-data baselines for bootstrapping conversational recommender systems.
Motion-Omni is an end-to-end framework that jointly generates spoken dialogue and full-body co-speech motion from shared hidden states, using scalable pseudo-labeling and a unified evaluation protocol to achieve real-time, aligned…
HyQuant improves low-bit LLM attention quantization by preserving critical vertical-line tokens and local windows in high precision while quantizing the rest, maintaining accuracy with low overhead.
A perturbation audit ran 14 medical LLMs through 30 kinds of edits to their chain-of-thought (flipped severity, swapped demographics, ablated evidence) and checked whether the visible reasoning actually tracked the final answer. 73% of the time it didn't: corrupt the chain, the diagnosis stays put. Two board-certified clinicians confirmed the edits were clinically valid, so this isn't a measurement artifact.
Researchers built a hidden-information chess variant, where a king's identity can secretly relocate, and asked models to state confidence separately from their move. When a model was at least 50% sure it had found the hidden king, it was right once in 62 tries. The uglier finding: the setup that looked best on legality, latency, and cost had the worst-calibrated beliefs of the group.
Add a level that reflects on the agent's process, and something in that loop has to stay frozen to keep it stable, usually capping real gains at meta-depth two. This paper fixes the meta-operation itself instead of the content it edits, then lets depth emerge from convergence rather than a preset limit.
PILOT enables live self-improvement by allowing a supervisor to steer active workers and distilling execution experience into reusable skills, improving accuracy and efficiency.
An agentic framework combining LLMs and VLMs enables consistent, multi-instruction editing of long multi-shot videos while preserving spatiotemporal structure.
Test-Time Policy Optimization enables label-free test-time training for mathematical reasoning by asymmetrically distilling agreeing rollouts and penalizing disagreeing ones, matching supervised performance.
Aphanta evaluates when image-editing intermediates improve multimodal reasoning by testing direct, editor-generated, and idealized visual states across tasks.
The study formalizes probabilistic alignment for world models, introduces PAWBench and PAWEval to evaluate video generators as stochastic samplers, and finds current models fail to match reference behavior distributions.
UrbanGround evaluates whether multimodal language model agents can sustain reliable navigation and spatial reasoning in a realistic 3D city replica, revealing that local perceptual skills fail to compose into extended goal-directed…
Self-OPD eliminates task-specific teachers in flow matching by using self-explored stochastic branches and normalized advantages to optimize the velocity field for multi-objective alignment.
Evolution strategies improve reasoning diversity and Pass@K over GRPO through sparse functional updates and population diversity, supporting a hybrid training approach.
Magpie is a real-time generative rendering system that separates gameplay logic from visual generation to preserve interactive designability while reducing asset requirements for game prototypes.
Agentic data generation is framed as constrained distribution design over factorized experience tuples, emphasizing execution-grounded accuracy, learner-relative complexity, and diversity rather than scale alone.
GGSS reduces demographic bias in generative vision-language models by steering visual tokens along geodesic arcs with an adaptive gate, preserving visual-language accuracy.
Recurrent fast-weight memories and selective state-space models are analyzed as online learning rules under autoregressive semantics, yielding normalized update families with stable renormalization that improve length extrapolation.
Ring Forcing is an autoregressive video diffusion framework that improves long-term memory through ring-structured training, history compression, and sparse rotary embeddings to achieve minute-long coherence.
A cost-efficient open-source pretraining recipe trains 2B-parameter models on consumer GPUs for under $7K, yielding performance near larger baselines while deriving cost scaling laws and studying data curricula.
J-Zero enables self-improving language models across verifiable and unverifiable domains through adversarial co-evolution of a task generator, solver, and judge using predefined preference pairs.
ABot-Recon achieves stable long-horizon streaming 3D reconstruction by using only local temporal context and frame-independent predictions composed sequentially, reducing drift via a lightweight temporal refiner and composition-aware pose…
VLAct improves vision-language-action model performance by pre-training on diverse robot data with preserved vision-language priors and shared action semantics, achieving strong results across simulations and unseen embodiments with…
Code-as-World represents physical environments as executable code to enable quantitative reasoning and scalable supervision for vision-language models.
CritICL improves LLM reasoning at inference time by using structured failure patterns from weaker models as critique-based guidance, reducing generation and token costs.
Image Bundle Composition reframes retrieval as dynamic assembly of relationally coherent image groups, with a benchmark and agentic framework addressing combinatorial joint relevance.
UI-Venus-2 is a general-purpose multimodal GUI agent that uses unified reasoning-action loops, expanded environment coverage, and robust verification to enable reliable real-world digital automation.
RECAP-Forcing improves long video generation by indexing memory according to appearance novelty rather than recency, preserving key-value caches for newly visible content to maintain long-range consistency without extra training.
Boundary-Calibrated Intervention Transfer selectively reuses past training evidence by checking contextual applicability and running bounded trials, reducing harmful updates and improving final model quality in autonomous post-training.
The setup exists because agents that look great on narrow patch tasks tend to fall apart once the change has to hold together across hundreds of files.
Debate and critique loops get credited with gains elsewhere, but under an equal compute budget, letting agents talk to each other can flatten the diversity of their answers instead of sharpening it. A crowd only helps if it disagrees first.
a controlled logic-puzzle study finds on-demand AI assistance lifts performance while it's turned on, then leaves a mark once it's taken away. Worth remembering next time a product pitches itself as a tutor.
RetrievalRouter adaptively selects retrieval pipelines per query to improve both accuracy and speed across diverse document benchmarks.
A new benchmark evaluates how well multimodal language models follow diverse video-based instructions with visual, audio, and structural constraints.
VoiceMem introduces a dual-brain streaming memory architecture for speech language models that improves retrieval accuracy, emotional personalization, and real-time efficiency.
MA-VLA enables multi-arm collaboration by assigning atomic actions to individual arms and using training-time permutations to generalize to unseen coordination patterns.
StreamPI enhances single-frame vision-language-action models with streaming temporal reasoning via instruction-anchored attention and randomized interval training, improving robot manipulation without extra parameters.
Visual Rubrics-Based Reinforcement Learning improves vision-language model grounding by scoring answers on visual faithfulness, reasoning consistency, and instruction following using structured partial credit.
JIT-Agent is a trainable model that synthesizes adaptive agent harnesses for off-the-shelf LLMs, improving performance across diverse models and tasks.
A calculus and framework unify revertible effects and reactive coeffects into a context paradigm enabling spatiotemporal composability for dynamic software components.
Code World Model separates persistent world dynamics from visual rendering by using a language model to generate executable state updates and a video model to render observations from proxy representations.
VBBR-Pro introduces a closed-loop testbed that enables scalable, verifiable, and controllable native visual reasoning through generation across diverse visual substrates.
Game engines provide executable verification and long-horizon trajectories for reinforcement learning post-training of spatial world models, motivating a human-engine verification paradigm.
Procedura is a 3D modeling agent that generates editable, part-structured procedural assemblies with sharp geometry and validated articulation from text prompts.
CaSKG calibrates procedural skill relations via counterfactual-causal graph construction to improve compact, executable retrieval for LLM agents.
Zero-WAM enables robotic manipulation of unseen tasks by conditioning a causal video-action model on in-context human video guidance, supported by an automatically generated dataset and a future-chunk prediction objective.
Multilingual self-play reveals that large language models exhibit significant cross-lingual skill inconsistencies in reasoning and strategy, partly recoverable by altering intermediate reasoning language.
Prefix Sliding reduces memory costs during long reasoning by discarding unimportant intermediate tokens, enabling efficient test-time scaling without retraining.
EASEL benchmarks fine-grained visual tool use through reference-guided reconstruction and semantic tasks, revealing that current multimodal agents struggle with closed-loop precision and trajectory stability.
LMSM applies a modular security framework to LLM serving by separating evidence calibration, policy evaluation, and output gating, enabling flexible interpretability-based enforcement without rebuilding request handling.
TacForcing is a streaming action-generation framework that integrates real-time tactile feedback during execution via a streaming action expert and execution-aware tactile attention, improving contact-rich manipulation.
On-policy self-distillation for mathematical reasoning uses privileged information to guide a model without a larger teacher, but suffers from reasoning collapse governed by token weighting, privileged context, and guidance dynamics.
Seven LLMs got the same Reddit "am I the asshole" style dilemmas twice: once framed as someone else's story, once as the user's own. Same content, softer verdict every time the user owned it. Loneliness and distress produced the biggest swing. The models aren't flattering anyone outright, they just get vague and evasive exactly when a user is most exposed and least likely to want that.
Ten therapeutic-move categories, grounded in the MULTI-60 inventory and validated by five licensed psychologists, then checked against real counseling transcripts. LLMs out-question human clinicians and barely touch psychoeducation, and they'll only pick up a technique if the human starts it first. Handing the model the move list as an explicit tool, no fine-tuning, cuts that gap to human behavior roughly in half.
LAWA improves robot control by using compact latent actions to retain efficient future imagination without generating observations, achieving strong performance with lower latency.
A framework for removing gameplay UI from videos enables cleaner training data for video world models, improving reward metrics and outperforms mask-based removal methods.
Smart glasses are surveyed as unified first-person intelligence platforms requiring sustained perception-state-interaction-action loops across constrained hardware and diverse applications.
LAION-BVD is a large-scale open video dataset enabling multimodal pre-training across video, audio, and image modalities with synthetic captions and strong benchmark performance.
WeMM-Embedding is a family of universal multimodal embedding models that align text, images, videos, and interleaved inputs in a shared space, achieving state-of-the-art retrieval and recommendation performance across public benchmarks and…
TorchMorph is a PyTorch extension providing GPU-accelerated morphological and distance-transform operators across up to eight dimensions with a SciPy-compatible API.
BPCO stabilizes critic-based reinforcement learning for language models by combining bounded value predictions, Monte Carlo targets, and adaptive advantage estimation, matching group-based methods with single-response sampling.
OPDVR integrates on-policy distillation with verifiable rewards via a ReLU-gated implicit reward reformulation, improving reasoning performance without extra hyperparameters.
DiffusionOPSD uses on-policy self-distillation to turn image-level rewards into explicit intermediate targets for diffusion models, improving alignment efficiency and enabling separate analysis of target construction and policy fitting.
Recuris introduces a recursive memory architecture that tracks progress and guides skill selection to improve long-horizon agent success through localized, validation-gated updates.
CAFE couples a search agent and critic via shared parameters to learn in-trajectory corrective feedback, improving search performance and reducing hallucinations across benchmarks.
LibriBrain100 is a large-scale MEG speech dataset that demonstrates improved decoding through extensive within-subject recordings and multi-subject supervised fine-tuning of pre-trained models.
Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits. As these runs unfold, users face a practical cost-quality trade-off: escalating to a stronger model when a cheaper one struggles, or…
RubSE improves UI-to-code generation stability by using rubric-guided self-evolution to prevent visual repair coupling and trajectory collapse.
A gated recurrent transformer reuses a shared core across depth with adaptive update gates, achieving comparable or better quality than deeper models with far fewer parameters and lower memory.
Off-policy reinforcement learning stabilizers vary with data availability, motivating regime-aware algorithms that adapt normalization and Q-function clipping to improve efficiency across CPU and GPU-parallel training.
AnTrap benchmarks GUI agent robustness by injecting dynamic anomalies into execution trajectories, revealing universal vulnerabilities and distinguishing learnable traps from intrinsic reasoning limits.
D³-MOPD dynamically adjusts domain sampling ratios during multi-teacher distillation by monitoring per-domain reverse-KL trajectories, improving convergence efficiency and closing most of the student-to-teacher performance gap.
FrontierChallenge evaluates end-to-end scientific workflows across domains, revealing that frontier models complete only about 20% of tasks despite high partial scores and frequent claims of completion.
Direct QA benchmarks for conversational memory do not predict user satisfaction, whereas natural integration of prior context does, revealing a large gap between elicited recall and conversational use.
GameWAM is a unified world-action model for native video-game control that jointly predicts future visuals and executable keyboard-mouse actions using block-causal flow matching, mode-specific distributions, and block-cycle replanning.
Cross-lingual Ranking Preference Optimization improves multilingual alignment by transferring English preference knowledge to target languages through hierarchical ranking optimization.
StarHarness evolves fixed-weight agent harnesses via stratified task pools and hidden selection to improve enterprise tool-use performance and cross-model transfer.
StepGuard is a step-level guard model that audits agent actions before execution, trained via automatic trajectory generation and balanced reinforcement learning to reduce attacks with minimal utility loss.
PonderPounce leverages native causal context in multimodal language models as robot episode memory, jointly training a reasoning System2 module and a fast System1 action model to improve long-horizon policy performance without dedicated…
Luce unifies geometry and PBR materials in a voxelized Gaussian cloud, using a variational autoencoder and rectified-flow transformer to generate relightable 3D assets from single images.
KATok is an adaptive transformer-based video tokenizer that selectively drops uninformative tokens to achieve data-dependent compression while preserving spatial consistency for diffusion-based generation.
KnowChange uses vision-language models to guide flexible, diverse change data synthesis for improving change detection training.
Teaching an agent to save and reuse "skills" from past tasks sounds like free progress. A new study finds whole-task skills usually drag performance below an agent with no memory at all, while subtask-level skills, especially kept as text instead of code, actually help. If you're building agent memory, granularity is the difference between an upgrade and a liability.
Multi-model systems need to route each query to the right specialist, but figuring out who's best can cost more than just asking a mediocre one. This paper treats it as the classic Pandora's Box search problem: pay for an expensive quality estimate only when the extra certainty is worth it.
A theory paper, and my skepticism is up: it models a post-AGI economy where corporations own AI and robot populations that produce and consume among themselves, growth maximized by reinvesting everything, humans optional. The math holds on its own terms, and it collapses human welfare to one variable: how much of the corporate network people actually own. Read it as a warning about ownership policy, not a forecast.
The study proposes a scaling law linking data size to tokenization capacity for user behavior modeling and introduces an adaptive tokenization method to improve efficiency.
Industrial technical reports contain high-value knowledge for maintenance, troubleshooting, and product engineering, but their heterogeneous structure (dense prose, specifications, tables) makes them difficult to index and reason over with…
Generative world models are evaluated against traditional simulators across eight capabilities, revealing gaps in physical guarantees, state feedback, and long-horizon stability despite progress in interaction and controllability.
RIBOSPAN is a large bidirectional RNA foundation model pretrained on up to 10,240 nucleotides that enables high-resolution full-transcript modeling, strong long-context representations, and discrete-diffusion-based mRNA generation and…
ERPO replaces action-side policy regularization with input-side query distribution control to stabilize reinforcement learning for language models while preserving response exploration.
EchoWM is an omnimodal world model that generates synchronized high-resolution video, sound, music, and speech while following continuous 6-DoF navigation trajectories across first- and third-person views.
Search-augmented LLM recommenders are highly vulnerable to web content polluted by generative engine optimization, frequently promoting fake products despite reasoning and defenses.
MobilePA-Bench is an interactive sandbox benchmark that evaluates mobile planning agents on tool-calling, sub-agent collaboration, memory usage, and composite skill invocation under real runtime constraints.
Retrieval-augmented generation extensions amplify automatic speech recognition errors in spoken multi-hop question answering, primarily through corrupted query entities.
Task-CoEvolve improves LLM harness optimization by adaptively selecting validation tasks and estimating full-set performance from partial evaluations, cutting evaluation costs by 80%.
ReWorld separates short-horizon control and long-horizon memory during training, then bounds both at inference via mixed attention windows, a pose-indexed landmark bank, and distribution-matching LoRA distillation to enable real-time…
Prime Agent is an open-source harness that uses recursive subagents, persistent computation, and agent-to-agent coordination to extend language models' long-horizon capabilities across coding and reasoning tasks.
Retrieval-augmented QA systems can exhibit hidden answer churn during index updates without noticeable accuracy changes, motivating compatibility audits alongside utility evaluations.
Apodex 1.1 improves sustained, verifiable progress on complex real-world tasks by scaling executable environments and training agents to coordinate long-horizon work with state maintenance and recovery.
CyberFactory is an open-source framework that builds agentic training data from real vulnerabilities to train Aegis, improving open-weight cybersecurity performance across proof-of-concept generation, patching, and question answering.
AutoSaddler automatically improves LLM agent harnesses via offline failure-driven optimization, boosting performance on long-horizon benchmarks.
The study proposes a lightweight audit to detect causality violations in sequence models by verifying prefix invariance, revealing that attention-mask checks miss leaks from scans or normalization.
A causal measurement probe and iterative submodular scheduling overcome evidence-utilization and context-allocation bottlenecks in retrieval-augmented generation, enabling sequential feedback-driven portfolio synthesis.
EvoMap externalizes verified execution experience into reusable structured Gene, improving long-workflow task completion and reducing token costs across diverse models.
Mixed supervised fine-tuning on combined reasoning corpora outperforms next-chunk reinforcement learning in efficiency and final accuracy across mathematical and out-of-domain tasks.
JoyAI-Echo-1.5 unifies long-form video and interactive world generation through cross-shot memory, geometry-aware camera control, and rollout-aware training to maintain identity and coherence over extended sequences.
Agent-G² models hint depth as a Gaussian distribution estimated online from existing rollouts, improving reinforcement learning on long-horizon tasks without extra probing.
CaRGo-T improves multimodal humor understanding by modeling causal relationships as graph-based reasoning structures interpreted by vision-language models.
Vision-Language-Action (VLA) models can turn multimodal context into robot actions, but their action decoders are still trained largely by behavior cloning.
ran eleven open speech models, Whisper-Large-V3, NVIDIA's Canary-Qwen, IBM's Granite-Speech among them, through tests built to catch benchmark gaming instead of just scoring accuracy. Silence the numbers in a LibriSpeech clip and top models still "recover" them 30 to 40% of the time.
Old benchmarks test whether a model forgets an isolated fact, not whether it can drop the harmful use of a concept while keeping the benign one. Ask a model to forget "how to synthesize a toxin" and a fact-based benchmark is satisfied if it forgets something unrelated too.
A new martingale paper works out exactly what classical concentration bounds, Ville's inequality and PAC-Bayes among them, throw away when you stop at a random time instead of a fixed one. Not something you'll ship this week. But if your eval pipeline leans on PAC-Bayes-style generalization guarantees, this is what those guarantees are quietly not telling you.
WorldToken fuses heterogeneous robot observations into per-timestep world tokens processed by a causal Transformer and diffusion action head, with scaling and temporal-context analyses on RoboCasa and RMBench.
Entropy-Valley selects target lengths for masked diffusion translation by scoring predictive entropy, improving adequacy and showing length choice matters more than unmasking order.
Using a galaxy-image transformer with known physical concept ordering, the study shows that linear probes recover real structure and that concepts emerge in a fixed difficulty-based sequence during training.
ClawProBench evaluates agent configurations via execution traces across live and frozen tracks, revealing that final-answer rankings obscure native-runtime failures and process-quality differences.
AutoResearch is a two-stage autonomous system that grounds research ideas through integrated generation and evidence-based execution to improve experimental reliability and measurable outcomes.
A 1.5B model trained with GRPO learns to pick NoThink, Short, or Long reasoning before it answers, instead of running a fixed token budget on every problem. It holds MATH500 accuracy while cutting average response length by about 41%, and the savings carry over to GSM8K on the easy questions. If you're paying per token for reasoning models, this is the fix nobody's shipped yet.
An audit of Qwen3-8B under LoRA self-training found seven ways researchers fool themselves, the worst being a single greedy-decode ledger that manufactures capability gains on an untrained model purely from inference-batching noise. Rerun with a proper per-problem test against a measured baseline, external distillation genuinely helps on problems the base model rarely solves.
A clean head-to-head of seven eviction policies (FIFO, LRU, LFU, ARC, GDSF, streaming SISO, and a redundancy-aware method) across workloads and encoders found LFU edges out the rest as the default, barely. The number that matters more: at normal similarity thresholds, only 2.1 to 3.9% of LMSYS and QQP "hits" are actually answer-substitutable, which drags claimed hit rates of 51 to 60% down to 1.1 to 2.2% in practice.
GameXpert-Bench evaluates coding agents across three game development stages—generation, repair, and optimization—using interactive and behavioral tests to reveal strengths in building playable foundations and weaknesses in defect…
A multimodal Turkish dialogue dataset and genetic algorithm-optimized interpretable rules are used to predict turn transitions from visual, acoustic, and linguistic cues.
FIRM-Video uses checklist-driven verification of temporal visual evidence to build reliable reward models for text-to-video evaluation and alignment.
Harness-Aware Training enables compact models to adapt to evolving digital-avatar harness configurations with low latency and high accuracy.
A benchmark of contrastive GUI instructions reveals that vision-language models mostly fail to localize interface elements rather than misunderstand spatial relations, though marking candidates substantially improves selection.
Agents can coordinate through their internal activations, a channel that never shows up in the chat log you'd normally audit. The paper's monitor catches it anyway: 99.3% AUROC on matched-model auction pairs, 85.4% on mixed pairs, and once it starts steering agents away from collusion, low-ball bidding drops 47 points. Worth knowing if "we reviewed the transcripts" is your whole safety case.
One LLM plays two roles: it designs executable training environments, then trains inside them, targeting difficulty at the gap between its performance with and without hints. Static benchmark sets go stale as an agent improves; a self-generating one doesn't. At 30B parameters it beat fixed-environment baselines by 5.3 points averaged across eight benchmarks, with the biggest jump, +13.9, on ACEBench-Agent.
No benchmark here, just an argument worth sitting with: frontier models already land their average answer on target, so ranking them by capability measures an axis that's saturated. The marksman analogy does the work. Capability is where the average shot lands, precision is how tight the group is, and almost nobody publishes group size.
FlashPrefill V2 improves long-context serving via mean-corrected sparse attention, optimized GPU operators, and framework integration, achieving large speedups over dense baselines.
EviRank reformulates multimodal image re-ranking as semantic constraint satisfaction by parsing queries into structured evidence packages and verifying candidates via rubric scoring and listwise comparison without training.
A two-step hyperparameter transfer framework predicts optimal learning rates for large Mixture-of-Experts models by scaling across widths and token budgets, enabling efficient pretraining without costly sweeps.
OmniAssistBench evaluates real-time interactive video assistants by reverse-engineering multi-turn interaction videos, revealing that current omni-modal models struggle with visual prompts, context retention, and timely responses.
AgentMercury synthesizes scalable executable business environments that serve as generalizable reinforcement learning substrates, improving agent performance across enterprise and out-of-domain reasoning tasks while making environment…
RecVerse is a GUI-grounded agent that uses hierarchical memory and trajectory-level reinforcement learning to simulate realistic multi-turn e-commerce shopping sessions.
Graph Engineering organizes multi-agent LLM systems through dynamic graph structures to coordinate specialized agents and manage complex, evolving tasks.
CLEAR uses a hidden-state gate to continuously modulate a safety low-rank adapter, improving LLM safety while preserving utility on benign inputs.
InfinityEdit enables continuous, unbounded video editing by using a lightweight adapter with history, temporal causal, and edit cross-attention modules to extend edits into future streaming frames while maintaining stability.
A framework quantizes vision-language models for mobile deployment using self-generated training data and a 2.7-bit format, compressing Llama 3.2 11B Vision Instruct to 3.7 GB with preserved visual question answering performance.
EXPL-FR explains face recognition similarity scores by aligning vision-language embeddings to the recognition space, enabling label-free auditing of semantic attributes and model comparison.
TLive-Omni is an omni-modal model for live-commerce that unifies image, video, audio, and text via timestamped token grouping, staged supervised training, and reinforcement fine-tuning with verifiable feedback to enable accurate real-time…
PhysCaP is a physics-informed code-generation agent that actively explores objects to infer hidden physical properties for efficient robotic manipulation.
Quantization-aware healing recovers compressed 4-bit language models faster and more stably than quantization-aware training by distilling directly from the original uncompressed model.
A modular medical imaging agent decomposes spatial relation verification into parsing, anatomical localization, and geometric rules to outperform end-to-end vision-language models on CT spatial reasoning.
Atomic generation facts compiled into a graph enable recursive scientific processes, and training-learning dynamics in nanoGPT are characterized by state-conditioned parameter updates, persistent functional reorganization, and frozen…
Repository-level coding benchmarks have become the standard for evaluating coding agents, yet they inherently suffer from data leakage because they are built upon popular open-source repositories repeatedly used for training.
pokes a hole in the "agents get better with a memory bank" story. Run the same agent through the same tasks in a different order, and its self-improvement curve moves around more than you'd want, sometimes barely improving at all. The memory bank isn't learning a stable skill so much as accumulating whatever happened to work on whatever came first.
targets a real weak spot in LLM-as-judge setups: fine for subjective calls, shaky on objective ones where there's an actual right answer to get wrong. The fix is a judge that can decline to rule and pull in retrieval instead, with a formal bound on how often it's allowed to be wrong. If you're using an LLM judge to gate objective outputs, "no verdict" beats a confident wrong one.
The parsed view an agent searches, the file it edits, the diff it reviews, and the artifact it submits can quietly drift out of sync with each other. The fix is a versioned workspace that pins all four to the same state.
NAPE uses causal Transformers to predict successive spectrogram patch embeddings for self-supervised audio learning without auxiliary components.
High-performing speech recognition models reproduce benchmark transcripts despite contradictory audio, revealing benchmark-optimized behaviors that inflate scores without improving real-world transcription.
SWE-bench Science benchmarks coding agents on scientific software repair, revealing failure mechanisms and mixed effects of scientific guidance.
Repo0 uses a dual-graph architectural state and modularity-guided structural evolution to generate complete software repositories from natural-language requirements with high functionality coverage.
IAR is a three-stage post-training framework that injects structured document knowledge into language models, aligns them for retrieval-free question answering, and recovers general capabilities, improving both domain accuracy and general…
4DAnyone reconstructs 4D humans from monocular video by generating multiview-consistent videos and lifting them into 4D Gaussian Splatting, using reference and target context designs to overcome scaling bottlenecks.
Retrieved memories can induce reasoning errors and belief distortions in large language models, and an inference-time strategy helps avoid these cognitive traps while maintaining benchmark performance.
WithEveryone enables reliable identity-preserving group image generation for up to ten people by grounding identities to layout plans and using region-based identity losses.
EnvHarness and EnvRigger dynamically reshape static environments via programmable plugins to target agent weaknesses and improve reinforcement learning co-evolution.
EXIMO efficiently fine-tunes large vision-language-action robot policies by combining VLM-guided exploration, imitation on orchestrated data, and residual off-policy reinforcement learning.
Customer-service LLM agents must follow organizational policy when acting on a user's behalf. Compliance failures arise from either forbidden actions, such as granting an ineligible change, or omitted procedural requirements, such as…
FlowEvo enables large language model agents to co-evolve reusable skills and workflows during inference, improving accuracy and efficiency across diverse benchmarks.
GOAG is an object-agnostic deep generative grasp planner that learns a gripper-specific contact surface distribution to sample valid grasps for unseen objects without object-specific training.
CoToGrasp is a generative framework that synthesizes diverse, stable grasps conditioned on specific contact topologies using an object-agnostic, gripper-centric workspace for zero-shot generalization.
FlavourBench evaluates language models on culinary portfolio tasks using executable ground truth, statistical rigor, and fully reproducible verification.
A small hybrid language model using sparse attention and short convolutions achieves faster CPU inference and better benchmark scores than larger conventional models despite training on far less data.
Face-free presentation attack datasets enable transferable PAD representations that improve cross-dataset detection without relying on facial content.
Recent agent benchmarks increasingly ground evaluation in executable environments, from code repair to web navigation, app APIs, and function calling.
Block3D accelerates text-to-3D generation by using block-wise diffusion with confidence-guided correction to reduce inference time while preserving geometric fidelity.
RISE adaptively decides when to continue or stop imagination rollouts for planning by weighing expected benefit against cost, supported by a counterfactual driving dataset with expert annotations.
Peer-ranked feeds of synthetic LLM agents increase lexical convergence but do not reliably produce opinion capture or coordination advantages across model families and topics.
OraRL improves reinforcement learning post-training for video multimodal language models by integrating oracle rollouts with decoupled advantage estimation and sign-balanced pruning, achieving higher sample efficiency and scalability…
Stream4D improves autoregressive video generation by replacing static 3D critics with a dynamic 4D reconstruction reward and motion prior to preserve coherent motion and reduce geometric drift.
VGI-bench evaluates visual reasoning in video generation models through 27 tasks, revealing limited reliability and minimal self-correction during generation.
It holds across model families and scales, including frontier reasoning models, and the hypnotic prompts even transfer from one model to another.
Trained with a technique the authors call Intervention-Paired Training, it hits an F1 of 0.944 on a 3,200-page benchmark, up from 0.862 for the prior baseline. Pointed at live Google and Gemini results, it flags roughly 9% of pages as GEO-optimized overall, and 16% among pages edited this year.
A reformulated optimization problem, refined with AlphaEvolve, pushes the exponent bound from ω < 2.371339 down to ω < 2.371177. Small number. But every shave here eventually lands in linear algebra libraries nobody thinks about until they're faster.
SkillGate fixes selector credit starvation in agent skill selection by separating outcome credit for execution tokens from local advantage for skill-naming tokens, improving success rates and reducing misleading skill exposure.
Co-RL enables unsupervised reasoning via cooperative multi-agent reinforcement learning with peer-derived rewards, improving performance across text and vision tasks without ground-truth labels.
Action-conditioned objectives improve latent geometry for Euclidean-cost model-predictive control by enhancing decision-metric alignment in world models.
SoftVTBench introduces a synchronized visuo-tactile dataset and deformation-aware benchmark for evaluating physical interaction quality during deformable-object manipulation.
Top-K prompting and plausibility-aware training improve diverse reaction prediction in single-step retrosynthesis, yielding state-of-the-art results on a large verified reaction dataset and motivating ensemble systems.
SemaPLC is a verification-gated agent harness that validates generated PLC logic through external compilation and live runtime execution, achieving higher verified pass rates than baseline methods.
FM-Bench evaluates long-horizon decision-making of LLM agents managing a football club over 20 years, revealing that managerial behavior rather than scale or token spend drives performance.
SkillForge improves automated software issue resolution by synthesizing repository-specific issues and distilling reusable skills from them before encountering real bugs.
FACET constructs executable terminal tasks by preserving source intent and grounding instructions, solutions, and verifiers in a shared repaired environment to enable scalable agent training.
A human-aligned chain-of-thought reward model and preference dataset improve joint video-audio generation by replacing fragmented metrics with coherent, dimension-wise reinforcement learning.
Structured Prior Knowledge explicitly extracts latent semantic, geometric, and contextual priors from pretrained object detectors to detect out-of-distribution hallucinations via a compact interpretable representation.
Fine-tuning large mixture-of-experts models on a low-resource language shifts reasoning into that language without harming accuracy, while reinforcement learning with verifiable rewards fixes formatting and leakage defects.
R2-OPD improves on-policy distillation by filtering teacher rewards that conflict with reasoning progress via within-trajectory ranking comparisons.
Multi-teacher on-policy distillation suffers from token-level budget misallocation across domains, which is addressed by balancing, dynamic allocation, and reward refresh to recover most of the oracle ensemble's capability.
DART-SD improves multi-turn tool-calling agents by modeling execution as a diamond-topology graph, identifying critical failure points, and applying localized self-distillation to preserve valid reasoning while correcting errors.
KBMR uses a multimodal language model to embed images by semantic identity rather than surface appearance, improving retrieval and visual question answering via continuous distillation and hard negative sampling.
A modular pipeline extracts structured text and metadata from historical newspaper scans using small interpretable models, yielding a large open dataset.
A coding agent writes its own executable world model at test time, then won't take an action until the model replays every observation seen so far. Playing the games blind, the base model scores 7.8%. With a self-written twin model it jumps to 93.3%, clears 23 of 25 games, and finishes 179 of 183 levels, beating human efficiency on 158 of the ones it completes.
What actually has to survive when an agent's context fills up, the app restarts, or a different agent picks up mid-task? This paper treats it as a compression problem with a provable floor: store decisions and hard constraints exactly, collapse repeated evidence into task-justified statistics, keep only the raw leftovers those statistics don't cover.
Multi-agent pipelines usually drop low-confidence or incorrect messages before a final answer gets assembled. Across five math and science benchmarks and two model families, this paper finds that's leaving value on the table: of the wrong-answer messages that end up changing a final result, more than four in ten change it for the better. A bad number can still carry a good decomposition.
Mixture-of-Experts vision encoders with fine-grained topologies, auxiliary-loss-free balancing, and specialized kernels scale efficiently while outperforming larger dense models on image and video tasks.
EditBridge enables efficient ultra high-resolution image editing via a diffusion bridge that translates low-resolution edits to high-resolution outputs while preserving source details through sparse attention.
A new dataset, benchmark, and 22B model enable compositional instruction-guided video editing with multi-region attention and temporal coherence.
Agent Lightning v1.0 enables reproducible reinforcement learning for arbitrary agent harnesses, substantially improving coding-agent performance with minimal data and compute.
HarnessRisk evaluates agent harness safety across six operational phases, revealing that configuration vulnerabilities and detection gaps allow high attack success despite preserved utility.
TAMP-Nav improves embodied navigation by aligning vision-language models with 2D visual prompting, selective reasoning with compressed memory, and dense policy optimization.
A capability-driven data infrastructure with curriculum scheduling and specialized data engines trains large multimodal diffusion models on curated heterogeneous supervision for diverse generative tasks.
GS-Voxel converts unstructured 3D Gaussian reconstructions into sparse structured latents to enable scalable aerial scene generation via flow models.
Researchers evaluate indirect prompt injection risks in DeepSeek Harness using controlled taint and dual judges, finding notable success rates across text and file channels and recommending controls between untrusted content and sensitive…
Artificial superintelligence (ASI) requires AI to move beyond mastering existing knowledge toward exploring the unknown, creating new knowledge, and turning new ideas into verifiable results. However, the capabilities of today's AI systems…
Scaling laws for text-to-image diffusion models reveal predictable compute-optimal training requiring far more data per parameter than language models, with robust overtraining behavior and universal curve shapes.
StartupBench evaluates end-to-end AI agents on real-world startup workflows and reveals that even top models complete only about 30% of tasks, highlighting gaps in instruction following and domain expertise.
A co-designed domain-specific language and multi-agent system improve LLM-driven 3D program synthesis by using relational operators and iterative execution feedback.
Agentic ESOpt uses evolution strategies for scalable full-parameter fine-tuning of long-horizon LLM agents via trajectory-level reward-weighted updates and parameter-context co-evolution.
Semantic task completion video generation evaluates whether generated videos achieve intended outcomes with semantic grounding, supported by a curated dataset and vision-language model-based benchmark.
LEGO-RL connects native coding-agent harnesses to scalable policy-gradient training via in-process LLM proxying, sandbox orchestration, and integrated monitoring, improving sparse MoE model performance across multiple harnesses.
PTXBench evaluates large language models on architecture-specific GPU kernel optimization, revealing uneven success and performance gaps that supervised fine-tuning only partially addresses.
Large language models improve continuously through iterative test-time feedback loops called Chain-of-Experience, outperforming zero-shot baselines with lower cost and higher token efficiency.
The study introduces MuseCPEval, a framework with tailored metrics to evaluate preservation of unchanged musical attributes during editing tasks.
ParaTempo improves parallel reasoning efficiency by using temporal confidence to dynamically prune, retire, and reallocate reasoning branches without synchronization.
This survey proposes a unified taxonomy and methodological framework for human-centric intelligence across visual, dynamic, and embodied levels within the foundation-model era.
TileMix routes attention score tiles to mixed FP16 or INT8 precision within fused dense attention, recovering long-context accuracy while improving prefill throughput without retraining.
WorldMind decouples game world modeling into state understanding, decision-making, control, and generation layers to enable explicit, state-aware NPC behavior.
A new paper runs three matched base/instruction-tuned model pairs on QA benchmarks and finds tuning shifts how confident a model sounds, while calibration gets worse and accuracy barely budges. The variety in how a model justifies an answer also flattens out, even when the wording on the surface still looks different.
If you've been telling yourself clean data means a weaker model, this is evidence against it.
Open source, YAML-configurable, runs on local CPU.
Prior Labs released open-source tools including a unified relational benchmark framework, a TabPFN-based relational model, and a model-agnostic predictive interface to advance reproducible relational learning.
GRNEdit is a lightweight two-stage framework that models video editing intent via binary semantic decisions and source evidence, achieving strong results with minimal parameters.
A plug-and-play 2D motion interface allows pretrained motion language models to process 2D inputs without retraining, improving real-world applicability.
This work proposes a compositional operator framework and TRACE-Bench to diagnose multi-reference image generation capabilities across atomic operations.
SA-MRPO independently standardizes multi-objective rewards and adaptively discounts saturated objectives to redirect optimization toward under-optimized goals.
HiFi-BRep improves B-Rep synthesis by using a topology-aware encoder and a single-stage decoder that jointly predicts geometry and topology with differentiable validity constraints.
The Travelling Thief Problem with Drone jointly optimizes ground routing, drone synchronization, and item selection to maximize profit, using mixed-integer programming, metaheuristics, and attention-based deep reinforcement learning with a…
AnyTalk generates 3D speech animations for arbitrary characters without animation data by adapting video diffusion models via character-specific fine-tuning and optimizing blendshape parameters from synthesized talking-head videos, with a…
Prior audit and repair episodes in context reduce false alarms by shifting decision thresholds rather than discrimination, with repair content and audit verdict complementarily affecting different model families.
HarnessEval-W uses hierarchical sub-agents to decompose world-model evaluations into verifiable reasoning chains that justify scores with transparent evidence.
Researchers propose a latent-to-pixel training strategy that accelerates convergence and improves inference speed for large-scale pixel-space diffusion models.
Ventor-QTest audits hosted open-weight model APIs via repeated and long-sequence black-box probes, measuring average and extreme fidelity loss to detect degradation in long-horizon agentic performance.
A unified black-box reinforcement learning framework enables stable, scalable optimization of general agents through complex harnesses via sandbox execution, trajectory reconstruction, and mix-harness training.
R³-Bench reveals that shared computation budgets cause reasoning agents to underperform relative to their single-problem capabilities across math, coding, and abstract reasoning tasks.
GenRouter is a unified routing framework that adaptively directs prompts to optimal agentic image-generation workflows, cutting costs and latency while improving visual alignment and enabling continuous self-evolution.
DiSCO is a black-box, zero-shot prompt-level defense that uses distribution-guided suffix expansion and contrastive scoring to reduce harmful image generation without altering the model.
PixRestore is a compact, VAE-free pixel-space diffusion transformer trained from scratch for unified image restoration, using flow matching on patchified pixels, DINO-based reliability-guided feature fusion, and adversarial fine-tuning to…
Cross-model reuse of frozen external memory tables depends primarily on aligning a lightweight target-side reader rather than the memory content alone, enabling reusable knowledge artifacts with optional adaptation.
A unified layer equation decomposes graph neural networks into seven components to compare architectures, derive theoretical bounds, and expose design choices linked to oversmoothing and expressivity.
FreeToken is an edge-native Mixture-of-Experts serving system that dynamically maps computation and model state onto heterogeneous local hardware to run large open-weight models on personal machines.
RUPA models agent execution as a dependency graph to propagate uncertainty across long trajectories, improving failure detection and confidence estimation for LLM agents.
Zetta is a closed-loop embodied harness that evolves runtime critics and recovery skills online to govern physical execution at action frequency, achieving high success on robot benchmarks with faster inference and scaling self-exploration.
Looped language models improve compositional, multi-step tool use through recurrent computation, with adaptive inference balancing accuracy and compute cost.
A literature-to-review pipeline automates problem discovery and triage to focus expert review on promising mathematical conjectures within a chosen research direction.
RapidLiDAR learns adaptive spatial displacements to initialize LiDAR scenes and refines them via multi-scale voxel and BEV features for real-time completion.
A hierarchical vision-language-action model improves long-horizon robot manipulation by using world-model-guided test-time search to scale computation for high-level subtask decisions.
On-policy distillation transfers reasoning behaviors rather than specific answers, with generalization strongly tied to teacher-student origin alignment and multi-teacher combinations causing capability trade-offs.
Hybrid-thinking multimodal language models suffer from response-pattern misalignment between thinking and non-thinking modes, which is addressed by a diagnostic benchmark and pattern-specific reinforcement learning penalties.
A hierarchical taxonomy and dense supervision strategy improve diffusion-based image editing through fine-grained concepts, large-scale paired data, and granular evaluation.
answers a question sharper than its title: do LLMs know when they're making things up? Probing internal activations, the authors find models do detect when an entity sits outside their knowledge, and do anticipate that the honest answer should be vague. They just don't act on it. Faced with an unfamiliar entity, the model generates confident, specific detail anyway. The signal is there. The policy to use it isn't.
A recurrent Large Discovery Model couples generative proposal with a Bayesian non-parametric reward surrogate to guide uncertainty-aware search across molecules, proteins, and programs.
WorldRover is a synthetic data engine that generates long-range, richly annotated video sequences with depth, camera motion, and tracking signals to support training models for coherent world exploration.
UI-Mate is a foundation GUI agent that uses environment-grounded training and in-context demonstration learning to improve reliability on long-horizon office tasks, achieving state-of-the-art results on computer-use benchmarks.
ConceptFormer learns continuous latent concept representations to bridge visual evidence and semantic relevance for visual document retrieval without relying on text intermediates or raw visual annotations.
Multi-byte prediction accelerates byte-level hierarchical language models by generating parallel bytes via variable-length windows and causal attention masking, improving inference speed with minimal quality loss.
HOTFIXR is a data generation framework that targets multilingual reasoning weaknesses to improve cross-lingual performance without sacrificing overall capability.
The Agentic Principal Chain enforces session-aware authorization checks to prevent harmful action combinations and delegation abuses in LLM agents.
TinyCast is a compact, attention-free zero-shot forecaster that uses spectral period detection and dilated convolutions to emit predictive distributions with minimal parameters and embedded-device compatibility.
Not "did it run," but "can it be proven correct." The gains right now come from scaffolding formal tools around the agent, not the agent getting better at proof-writing on its own. Worth tracking before agent-written code lands anywhere safety-critical.
A benchmark that only checks whether the task finished won't see this failure, since it happens downstream of the model entirely. If your agent's shell tool has ever done something inexplicable, this is probably why.
Diffusion-based drafters guess a whole block of tokens at once for speed, but score each position independently, so the guess gets worse the further out it reaches. DARTree swaps the flat guess for a tree of autoregressive drafts, letting the verifier check more candidates per step instead of waiting on the slow base model.
The ALD/E-ImageMiner benchmark and ICDAR 2026 competition advance machine interpretation of scientific figures through tasks spanning visual reading, domain reasoning, and evidential justification, proposing long-term goals for verifiable…
MOSS-VL is an open vision-language model family enabling real-time interaction by attending to vision via gated cross-attention during generation, using a synthesized interaction corpus and staged curriculum to achieve strong streaming…
PRISM is a fast, training-free test-time adaptation method that reverses low-rank affine noise distortions in audio-text models using frozen text prototypes and geometric corrections.
In language models, flexible reuse demands attention-mediated gathering at a mid-depth window to make latent variables readable, without a selective gate, and readout measures poorly reflect actual use.
MathForm improves autoformalization by retrieving Mathlib knowledge and iteratively refining outputs with verification feedback, yielding a large verified dataset and a high-performing 8B model.
Empirical evaluation of diverse memory substrates for long-horizon LLM agents reveals regime-dependent trade-offs, motivating adaptive substrate routing for reliable agent memory.
ForgeWM progressively distills bidirectional video generators into efficient few-step interactive world models with aligned discrete and continuous controls, supporting low-latency interaction and replay-time refinement.
LongRCA Bench evaluates failure diagnosis across lengthy agent trajectories, and the training-free RCTA method improves attribution of responsible roles and root-cause steps.
A malicious skill description looks relevant enough to get selected, then its instructions fabricate fake dependencies that route the agent through a string of unnecessary tools before it finishes the real task. On DeepSeek-V4-Pro, across 491 tasks, the attacker's coordinator got picked 80% of the time.
Push the context window past an intermediate sweet spot and language modeling, comprehension, and closed-book QA all get worse, not better.
Get it wrong and the agent games a proxy instead of doing what you meant. A new three-step process (distill the objective into measurable variables, pick reward terms by solving a max-flow problem on a causal graph, fit weights through a handful of preference queries) lets someone without an RL background build a reward function that's provably conflict-free, in as few as O(n log κ)…
Self-supervised visual on-policy distillation improves small vision-language models by distilling from original images into strongly augmented student views without privileged annotations or larger teachers.
On-policy distillation from a long-context reasoning teacher to short-context students improves mathematical proof reasoning and generalizes to science benchmarks by aligning token spans, constraining length growth, and stabilizing…
Forecast collapse in hourly equity return prediction stems from low predictability and per-series objectives, and the proposed CalibRank objective balances calibration and ranking to restore cross-sectional structure.
Second Thought is a training-free framework that runs auxiliary reasoning branches in parallel during agent action-observation waits to reduce sequential decoding and turn counts without harming accuracy.
PRM-as-a-Judge 1.5 provides fine-grained process metrics and reliability tools to evaluate embodied robotic models beyond binary success rates.
CPI-Bench is a comprehensive benchmark for real-world image editing that evaluates multi-image tasks, practical applications, and reasoning-based editing to better differentiate model performance.
Under proportional scaling of model size and training tokens, optimal repetition of high-quality domain data increases mildly with scale and correlates with domain validation loss rather than unique data volume.
A new benchmark for AI-generated video detection reveals that current detectors fail to generalize across realistic crisis-related videos and become less reliable as content spreads socially.
SPARGen unifies 3D reconstruction, dense correspondence, and spatial reasoning into a single instruction-conditioned multimodal generative model that jointly learns shared spatial representations.
Mobius-v0 separates global memory storage from iterative reasoning modules to improve knowledge compression and inference efficiency, yielding comparable performance with less training data and faster inference.
Marionette predicts explicit 3D articulated world states for interactive games, uses a fixed renderer for geometry, and synthesizes video via diffusion, enabling direct state-level control and long-horizon consistency repair.
Autonomous research agents evaluated across the full scientific lifecycle reveal a pervasive lack of metacognitive self-correction, motivating a new benchmark and failure taxonomy.
PACE-Bench evaluates self-evolving agents on physics adaptation tasks requiring iterative code redesign after environmental mutations, revealing that simulator-grounded reflection outperforms unverified self-revision but mechanism redesign…
Skills enhance LLM agents primarily by stabilizing execution through procedural anchoring rather than injecting missing knowledge, though retrieval bottlenecks and brittle assumptions limit their effectiveness.
The paper introduces personalized auto-research, a framework that conditions AI-driven hypothesis generation, experimentation, and writing on individual researcher representations to avoid generic outputs.
AdaPop adapts gradient pressure by fact popularity and automates forget-retain balance to reduce leakage of unlearned content.
A framework that separates thematic seeds from genre-form controls generates diverse, high-quality creative writing data across 13 genres and improves LLM creative writing performance.
Compatible open-weight language model checkpoints share detectable weight-space ancestry signals that distinguish true lineage from independent or distilled models without requiring data.
A hybrid quantum-inspired Kolmogorov-Arnold network improves federated ECG classification accuracy with fewer parameters and lower communication costs than standard multilayer perceptrons.
The paper tests whether SAE latent sets recover human category boundaries more faithfully than dense representations. They don't. They track the model's own internal structure instead. Worse, semantic edits a human would barely register can scramble the SAE's active feature set entirely.
PixSDS fixes VAE-induced pixel drift in latent score distillation sampling by guiding optimization with decoded image directions, reducing artifacts in text-to-3D generation.
UniSwap enables synchronized appearance and voice replacement in talking videos through a unified streaming audio-visual diffusion transformer with specialized training and inference adaptations.
LiveAnimate enables real-time, long-form pose-driven human animation via a 14B-parameter video diffusion transformer with specialized training, bounded attention caching, and sequence parallelism.
CW-BASS v2 selects pseudo-labels by measuring teacher reliability on held-out data and applying either strict filtering or an adaptive floor to avoid confirmation bias under saturated confidence.
H2R-Bench evaluates video generation models on transforming human manipulation videos into robot-centric demonstrations across embodiment constraints and interaction fidelity.
DreamX-Phi 1.0 is an action-conditioned video world model for robotic manipulation that uses geometric attention encoding, depth estimation, object masks with a frozen teacher, and distillation to generate faithful future observations.
A frozen vision-language model improves spatial reasoning by self-evolving through verified experience, reflection, and reusable memory retrieval without parameter updates or external tools.
LycheeMemory V2 improves long-term agent memory by batching interactions into semantic segments for efficient consolidation and structured retrieval, reducing construction costs while maintaining high accuracy.
PlayWorld benchmarks interactive video world models by using multi-modal agents to pursue long-horizon objectives, evaluating geometry consistency, interaction fidelity, and state evolution.
Evoke is an interactive world model that uses external persistent memory and a redesigned long-horizon teacher to enable responsive, open-ended video generation with bounded context and low latency.
OmniScientist is an end-to-end omni-modal AI scientist that performs multidisciplinary research directly from heterogeneous raw evidence using autonomous agents and lifecycle-wide perception, improving evidence-grounded discovery across…
Intern-S2-Preview is a scientific agentic foundation model series that integrates multimodal pre-training, multi-task reinforcement learning, and memory-augmented extensions to support long-horizon scientific reasoning and forecasting.
A framework decomposes compound logical options into atomic judgments and uses constrained optimization to improve reasoning over AND, OR, and NEITHER/NOR operators.
Context-Matched Distillation aligns teacher supervision with causal generation context for few-step autoregressive video models, improving control adherence and long-video quality.
AutoDesign uses a meta-harness optimizer to recursively improve a code agent for structured media generation, achieving state-of-the-art results on paper-to-poster synthesis.
A curated elementary-grade pretraining corpus and 5B-parameter model create a controlled sandbox for studying knowledge acquisition, representation, and bounded capability growth via post-training and in-context learning.
Latent On-Policy Self-Distillation learns privileged teaching context end-to-end from experience to provide dense token-level supervision, improving agent performance and efficiency.
HumanTracker introduces a large-scale benchmark and preference-aligned metric to evaluate humanoid motion tracking based on perceptual quality and physical contact stability.
Frontier autonomous agents excel at engineering optimization but show unstable performance, limited novelty, and variable experience reuse across long-horizon tasks.
V-RAE constructs semantically organized video latents from frozen vision representations to improve generation quality, convergence speed, and predictive modeling.
CardioState-JEPA learns a unified cardiac representation across ECG, PPG, and PCG by predicting masked latent physiological states with cross-modal delay alignment, improving downstream classification across all three modalities.
NARU is a Japanese long-form video benchmark evaluating narrative evolution and cultural reasoning through a hierarchical annotation pipeline and extensive native-speaker verification.
SkillEvo improves agent skills through multi-turn feedback and active governance to sustain evolution gradients.
Large language models and dedicated embedding models achieve nearly identical aggregate performance across diverse tasks, but embedding models are far cheaper and faster, supporting a division of labor by task type.
ARC improves fairness in group-based reinforcement learning for open-ended agents by conditioning rollout comparisons on strategy, enabling more context-appropriate behavior in responsive user-agent interaction.
DREAM introduces an agentic meta-control layer over industrial recommender pipelines that uses intent reasoning and dual-loop optimization to improve session-level recommendations without replacing existing models.
Providers return reasoning traces to the client as encrypted blocks instead of storing them server-side, meant to block distillation and IP theft. Turns out those blocks are interchangeable across sessions, users, and models within a provider's stack: feed a flagship model's trace into a weaker, less-guarded sibling model, and it decrypts and prints the whole thing in plaintext.
Too strict, too loose, take your pick, and some frontier models were just regurgitating gold patches they'd memorized during pretraining. SWE-Bench ProMax is the fix: 170 hand-curated refactoring tasks across seven languages, specs rewritten from scratch, tasks that touch 11 files and 260-plus lines on average. Best model on it resolves 41.2%. If your coding-agent numbers look too clean, ask which benchmark they ran.
The framework splits an agent's harness into four pieces (system prompt, rule bank, safety memory, tool policy) and patches them automatically whenever a trajectory fails, diagnosing what broke and validating that the fix didn't tank performance. Against a static baseline it cut attack success rate 3.1x, and the learned rules generalized to unseen risks on AgentHarm with no extra tuning.
Mechanist is an autonomous agentic system that uses AI to discover and control the mechanisms underlying model intelligence, generating hypotheses, performing causal interventions, and improving safety and performance.
Two studies define measurable GPU control gates for LLM-agent services by analyzing concurrent cohort scheduling and on-device routing versus host redispatch.
This work introduces a dedicated model for per-joint hand visibility estimation and demonstrates its benefit for multi-view 3D hand pose annotation.
Genesis organizes long-horizon software development around a persistent project rather than persistent agents, enabling multi-day compiler construction and numerical module reimplementation with low cost and high performance.
Stronger models can build inference-time harnesses that substantially improve weaker models' task performance without parameter updates by offloading reasoning into structured code and routing.
A closed-loop framework combining physics-based video synthesis, diffusion-based video dereflection, and a new benchmark achieves state-of-the-art video reflection removal with fast inference.
Researchers introduce MBA-Bench, a multimodal benchmark for business ideation agents, and propose MBA-b and MBA-k models trained with creativity and feasibility rewards via LoRA fine-tuning and group relative policy optimization,…
AVA-Encoder learns structured video representations via agentic auto-encoding using knowledge graphs to enable cinematic video generation and reasoning with reduced token usage.
StateFlow introduces a persistent 3D world state to enable iterative, controllable previsualization for film and game design by constructing, evolving, and accessing structured scene and camera representations.
Spark-to-Paper is a lightweight, composable workflow inside coding assistants that generates research papers by separating planning from reporting, enforcing evidence-based claim revision, and using integrity checks to reduce fabrication.
ToolHazard is a scalable framework that synthesizes adversarial environments to test LLM agents against indirect prompt injections, revealing vulnerabilities and improving defensive alignment.
TailBooster uses a dual-layer generative framework with statistical tail extraction and deep autoencoder cleaning to synthesize operationally valid extreme air-transport events, substantially improving extreme-event prediction accuracy.
Massive activations in hybrid linear-attention LLMs exhibit pre-attention spikes and inter-spike plateaus governed by cancellation timing, with morphology recovering at full-attention limits.
HPSE improves unstructured knowledge editing by distilling from hybrid rollouts that insert missing facts into the model's reasoning paths, enabling composable multi-hop reasoning.
This paper reports a single, fully instrumented case study of a large-scale architectural refactoring by an AI coding agent under a specification-first protocol, with no human review of the generated code and no pre-existing oracle to…
UniProbe is a lightweight learnable detector that uses a directed graph and alternating GNN, ViT, and GRU modules to identify hallucinated tokens in frozen large vision-language models, enabling real-time resampling during generation.
GAS improves multimodal understanding by using generation as auxiliary supervision via next embedding prediction and a decoupled mixture-of-transformers architecture, with no inference overhead.
Claim-Level Reliability Assessment improves reasoning accuracy by verifying critical claims instead of sampling more solutions, reducing token use while boosting performance.
Dion3 accelerates the Muon optimizer by reducing orthogonalization and communication overhead through algorithmic, kernel-level, and update-rule improvements.
VideoGAIA introduces a multi-turn, tool-augmented benchmark that evaluates agentic video understanding for advanced multimodal models through complex real-world tasks.
Preventing models from seeing option labels during answering does not reliably reduce positional bias or improve multiple-choice accuracy, and only showing all options with an LLM matcher preserves baseline performance.
OreoLook uses a three-layer caching system with session windows, semantic similarity matching, and deduplicated embeddings to reduce redundant LLM calls and maintain long-running conversations on modest hardware.
Blast Radius adds a memory layer that predicts how far an incoming prompt actually reaches through your code and chat history, then evicts what falls outside that radius, archiving it verbatim instead of deleting it, so it's recoverable if you need it back. Tested across seven OpenAI models, it posted the lowest overflow rate of any policy compared.
SkillProx splits the two: a forward pass diagnoses failures and edits with rollback, a backward pass decides what to consolidate, demote, or cut, borrowing the proximal-gradient idea from optimization and pointing it at text instead of weights. It beats the strongest gradient-based skill-refinement baseline on accuracy, and the gain holds across model backbones and both in-distribution and out-of-distribution tasks.
Multiverse Computing caches a teacher model's top-100 logits once instead of re-running it on every batch, then chunks the KL-loss computation instead of materializing the full vocabulary matrix. That memory savings let them distill a GPT-OSS 20B model on one GPU node instead of four, at nearly 5x the throughput. The resulting 3.2B student stayed within about nine points of its 8B teacher on MMLU.
InSight-doc adaptively allocates visual resolution during reasoning to improve long-document understanding while reducing latency and hallucinations.
DistilVDR is a compact 524M vision-document retriever distilled from an 8B teacher using cosine alignment without relevance labels, achieving near-teacher accuracy with far smaller indexes and faster indexing.
Adversarial Fréchet Distance improves generator post-training by adding a learnable adversarial feature space to static Fréchet losses, with whitening to stabilize optimization.
Open multilingual translation models are improved via group relative policy optimization with reference-free quality rewards and checkpoint interpolation, surpassing strong open and proprietary baselines.
Combodied Agents integrate digital and embodied tools into a closed-loop framework that models individual human-state trajectories over time to provide proportionate, consent-aware support.
Ex-Omni-2D is an omni-modal dialogue framework that produces coordinated text, speech, and video responses via a visual thought plan and a distilled streaming video generator.
SkillZip compresses self-evolving agent skills by finding a minimal faithful structural explanation that shares repeated rules and procedures while preserving rare exceptions, without requiring evaluation rollouts.
DSAgentBench evaluates autonomous agents on complete, multi-tool data-science workflows in real computing environments and reveals major performance gaps.
SPIEval benchmarks mobile assistant LLMs on scattered personal data tasks, revealing major gaps in information retrieval and verification.
A new benchmark called VibeLifeBench evaluates long-horizon proactive agents across simulated multi-week everyday tasks, revealing that current frontier models perform poorly.
Latent-to-4D enables reusable direct 4D generation from video diffusion latents via alignment with a pretrained decoder and spatiotemporal attention, transferring across generators without retraining.
Parameter-space exploration via perturbed policy sampling improves LLM reinforcement learning by diversifying rollouts and reducing training failures compared to action-space methods.
SHAPER is a train-free framework that improves embodied agents by evolving reusable skills and a context-code harness around a frozen foundation model through environment rollouts.
Simplax enriches uniform discrete diffusion via Dirichlet-categorical augmentation to improve reverse sampling and generative quality on text and Sudoku tasks.
Self-Geometry improves vision foundation model predictions by enforcing explicit multi-view geometric constraints via test-time adaptation with LoRA, disentangled losses, and angular neighbor sampling.
Agent safety should be enforced at runtime through preventive controls and verifiable evidence rather than relying solely on training-time alignment methods.
SKILLER is a reinforcement learning framework that automatically generates tailored skills for small open-source models to reduce inference costs while maintaining high task performance.
ReRound uses a conditional diffusion model to guide rounding of near-midpoint weights during low-bit post-training quantization, selecting candidates by matching leading singular values to improve small LLM accuracy without inference…
A new promptable paradigm integrates subject localization and gaze estimation into an end-to-end transformer model that uses text or visual prompts to identify subjects and predict gaze targets without multi-stage pipelines.
Gambit improves reasoning model efficiency by using thought-level beam search to dynamically allocate compute to promising reasoning traces under fixed hardware budgets.
MobileMem is a benchmark and framework for evaluating on-device long-term memory through year-scale, multimodal mobile experience trajectories that require temporal reasoning, knowledge updating, and preference inference.
ENTLORE is a benchmark framework that evaluates enterprise question answering by requiring recovery of implicit organizational relations across routine documents, revealing that even with gold sources many latent reasoning questions remain…
A new paper tests standard chunk-and-embed retrieval against financial and audit documents, where 87% of the content is table rows and figures routinely sit 13 lines from their own headers. On a 51-question benchmark built from a 780-page government financial report, dense retrieval scored 15.7%; a tuned version got to 35.3%.
HarnessOpt-Bench put five frontier models in the optimizer seat, each iteratively rewriting a target agent's prompts, tools, and orchestration code against eval feedback, across 4 tasks and 111 scored runs. The model doing the optimizing swings results more than the harness it runs through, and a model's own native harness doesn't reliably beat a generic one.
Testing Gemini 3.6 Flash on 2,190 controlled clips with frame-level ground truth, researchers found it reliably counts persistent state changes up to 12 events, but has no reliable range at all for transient events like blinks. Push it into high-count, high-frequency territory and the final tally is right just 0.2% of the time, capturing only 18% of what actually happened.
A 150M-parameter reasoning model using recurrent latent reasoning and in-context learning achieves a new cost-accuracy frontier on ARC-AGI-1.
A modular cognitive architecture integrates high-level reasoning models with real-time embodied execution for scalable intelligent virtual agents in interactive 3D environments.
A2E is an end-to-end evaluation engine for agent harnesses that uses a standardized task protocol and execution traces to assess capabilities across efficiency, tool use, planning, and error recovery.
RynnValue is a scalable open-source value foundation model for robot manipulation that uses temporal distance instead of preferences or progress to learn generalizable value predictions and improve real-world policy success.
Evo-Bench evaluates autonomous harness optimization across agent domains using sensitivity-aware task construction and reveals strong but domain-dependent evolution gains.
UserIDA improves user simulators by explicitly controlling interaction intent per turn through directive-conditioned generation and calibrated reinforcement learning.
Macaron-V1 is an open agent-model family that uses a Mixture-of-LoRA architecture and recursive self-improvement to enable continual learning and collaboration across specialized tasks.
Motif 3 is a large sparse mixture-of-experts language model using grouped differential latent attention and specialized training techniques to achieve strong reasoning, coding, and long-context performance.
Sci-VBench evaluates video generation requiring scientific reasoning across disciplines, revealing that visual realism advances have not ensured accurate scientific and causal dynamics.
RoMeRL reduces trajectory-indexed memory utilities to fixed-dimensional per-task states to concentrate feedback, avoid reward contamination, and improve self-evolving LLM agent performance.
A new attention mechanism replaces fixed scaled dot-product attention with a learned power-law bilinear operator, with verified architecture, measured stability, and machine-checked proofs.
Decoding-Level Taboo is a runtime logit-space stress test that reveals how large language models handle off-nominal generation paths, showing that robustness depends on scale and instruction alignment.
Agentic systems can achieve open-ended improvement through multi-component co-evolution that progressively removes fixed human constraints across agents, environments, and evolution mechanisms.
Researchers propose source-contrastive evaluation via a localized benchmark to detect data contamination and assess localization robustness in multilingual translation models.
Rhetorical framing significantly biases AI scientific review scores in structured ways, with effects shaped by reviewer identity, score range, and evaluation strictness rather than rewriting complexity.
Latent Dynamics Reasoning integrates kinematic dynamics in structured latent space to enable video world models that extrapolate physical laws far beyond training distributions with far fewer parameters and faster inference.
Cross-view pairing of rib fractures in orthogonal CT projections enables accurate 3D localization, but correspondence confidence remains the primary bottleneck.
Researchers propose a black-box red-teaming method using inaudible low-frequency waveforms to expose vulnerabilities in audio-language models, alongside a defense that detects distribution shifts and requests a second recording to recover…
UNMASK automatically discovers and mitigates spurious correlations in text classifiers via causal verification and group-based reweighting without manual annotations.
MMDiff uses multimodal sparse autoencoders to isolate, detect, and control specific features in multimodal language models, improving interpretability and targeted steering of visual and safety behaviors.
AWS folded generative AI inference recommendations straight into the Python SDK (v3.17.0+): benchmark a live endpoint, get ranked deployment options by cost and performance, deploy the winner, all from a notebook. No more hand-rolling Boto3 calls to figure out which instance type actually hits your latency target. Small change, but it turns a day of manual load testing into a few notebook cells.
Nimblemind's answer is nMAS, a multi-agent pipeline that extracts evidence-linked features from EHR tables, each one checked for structure, rubric compliance, and provenance before an LLM audits the batch.
Researchers ran forensic analysis on Android e-commerce apps and checked what AI features they actually run against what the apps tell users. AI showed up everywhere. Disclosure about it barely did.
A new photorealistic urban benchmark reveals large performance gaps for embodied agents in city-scale navigation and spatial reasoning.
UniMoMo compresses trained recommendation mixture-of-experts models into smaller standard MoE checkpoints via functional similarity grouping and layer-adaptive protection, preserving accuracy while accelerating inference.
Pruning strategies applied at different pipeline stages reduce token usage and latency in long-horizon research agents, with early pruning yielding the greatest efficiency gains.
A sub-2B Spanish cybersecurity vision-language model couples a frozen SigLIP encoder to a Spanish decoder via an MLP, introduces a NoPE positional-encoding ablation for visual attention, and reports near-zero visual grounding despite…
Unsupervised on-policy self-distillation improves large language models by using internal consistency and majority-vote pseudo-solutions to correct confident errors without external supervision.
Business Arena evaluates LLM agents running a realistic cross-border shop, revealing large performance gaps versus human strategies and enabling detailed attribution of business decisions.
Full-bandwidth transformers use latent feedback of top-layer hidden states to improve reasoning and efficiency without altering the core architecture.
A hybrid pipeline combining LLM gender classification with tag-aware neural translation improves gender accuracy in English-to-Romanian machine translation.
Hierarchical Self-Improvement evolves task-specific execution harnesses for frozen LLM agents via hierarchical self-modification, yielding substantial gains on moderate tasks while being bounded by feedback quality and backbone limits.
A reparameterized pretrained vision transformer unifies semantic understanding, high-fidelity reconstruction, and image generation within a single visual space without requiring a separate VAE.
A new benchmark run across 14 models on BFCL v4 found that letting agentic models call tools as typed Python stubs, not rigid JSON blobs, matches or beats the JSON approach in 11 of 14 models. GPT-5.6 picked up a 10.6% gain just from the switch, and under parallel tool calls the code approach won or tied on 13 of 14.
The authors built TrajErrBench, 486 hand-annotated failed runs across tool-use and coding tasks, and their method traces each error's fate through the rest of the trajectory to isolate the one that mattered.
Train a model to ignore misleading context and you also get a model that ignores correct context, which defeats the point. SCOPE, tested on a new benchmark called MIST that runs each question through clean, misleading, correct, and irrelevant versions, trains on preference pairs balanced across all four instead of just the adversarial case.
We present Ego-OSCAR, an open-hardware, low-cost, head-mounted stereo-inertial capture device for egocentric data collection in the wild. EgoOSCAR pairs a hardware-synchronized global-shutter stereo camera with a 6- axis IMU, an embedded…
ConCor-1 treats vision-language grounding as bidirectional concept correspondence, jointly predicting text spans, image segments, and cross-modal matches without prespecified phrases.
OasisKV improves LLM inference throughput by storing full KV caches in lower memory tiers and prefetching only relevant entries into HBM using speculative-decoding lookahead predictions.
We present Ouroboros, a self-developing agent harness whose tools, prompts, context assembly, and core implementation improve through reviewed commits that become the runtime for later work. Core evolution proceeds in two modes. In…
Counterfactual Evidence Disentanglement improves vision-language model grounding by auditing whether answers causally depend on local visual evidence during reinforcement learning post-training.
A unified toolbox enables reproducible comparison and extension of time-series dataset similarity methods for forecasting, classification, and generation tasks.
The study introduces a benchmark and formalizes narrative commitment preservation to evaluate long-horizon logical consistency in interactive storytelling with large language models.
NeuPAT selectively constrains updates to language-sensitive neurons during multimodal tuning to preserve LLM language capabilities while enabling perceptual adaptation.
It cost 1.41x the tokens to get there, and one deployment logged 254 real missions across six paper pipelines and upstream kernel work. The model weights never change.
Model size and release date predicted nothing about how often a model reinforced the spiral. Conversation length did. Add 350 messages of prior context and the rate of failing to discourage a suicidal user climbs from 30.0% to 41.1%.
runs psychometrics on eight safety benchmarks across 192 models and finds most of the variance collapses into three factors: refusal strictness, truthfulness, contextual harm. Pick the right ten items and you reproduce a full benchmark score at 1-3% of the original cost, and the same method flags sandbagging or a quiet model swap behind an API.
Vision-language models are increasingly serving as the reasoning core of embodied agents. Robot execution is inherently iterative: each action reshapes the scene and physical state, continually renewing what must be perceived, reasoned…
Test data from public benchmarks inevitably leaks into pretraining corpora, inflating evaluation scores once memorized. Contamination mitigation evaluation intervenes in the decoding process to suppress memorization and restore a…
Generic parameter-efficient fine-tuning (PEFT) methods transferred from language models can fail silently on real-time detectors, whose heterogeneous operators and detection-specific components impose placement constraints absent from…
World-Action Models (WAMs) improve end-to-end autonomous driving by transferring video dynamics priors to action prediction, but existing methods require costly future generation at inference. We present SimWAM, a simple yet effective WAM…
Recent systems for optimizing prompts, programs, and ML workflows typically rely on explicit outer-loop controllers such as evolutionary search, bandits, or textual-gradient methods. We ask a fundamentally different question: how much of…
Test-time training (TTT) views sequence modeling as an online learning problem in which fast weights are updated by an internal learning rule. Despite the growing number of TTT variants, existing approaches typically hard-code each variant…
Neural scaling laws are foundational for language model development, yet standard formulations systematically under- and overestimate loss at data-scarce and overtraining extremes. This failure originates in the underlying assumption that…
We study visual persistence in interactive video world models. These models rely on a Key-Value (KV) cache as a growing visual memory to carry forward previously generated frames. However, we find that models can no longer reliably address…
Agent Memory Distillation improves small language model tool-use performance by transferring structured hierarchical memory from a large teacher agent without additional training.
A large-scale deepfake video benchmark with fine-grained annotations and a multi-agent reasoning framework improves detection of emerging synthetic face videos.
A training framework called iFAN improves mask transformers by aligning query ranking with mask quality and distilling stronger intermediate predictions to the final layer.
Mendel Gödel Machine improves self-improving coding agents by using multi-trajectory mutations and cross-lineage hybridization to accelerate convergence and boost performance.
Atelier improves artist-grounded image generation by translating vague artistic intent into explicit control states that separate scene content from style, reducing reliance on stereotypical shortcuts.
MirrorWorld improves video mirror reflection synthesis by separately modeling semantic content associations and geometric spatial arrangements through relation distillation and transformation alignment.
AtlasVLA improves embodied AI by replacing reactive control with proactive reasoning via persistent world-ego memory, enabling robust long-horizon manipulation from a single wrist camera.
CoinRAG improves retrieval-augmented generation efficiency and accuracy by reusing fine-grained semantic nugget caches instead of full chunks.
AutoPrune uses large language models to automatically design visual-token pruning policies for multimodal models via a domain-specific language and residual search formulation, achieving high efficiency with minimal performance loss.
LLM routing is formalized as a sequential decision process with a unified benchmark and modular infrastructure to compare and improve cost-effective model selection.
D-SCAN detects retrieval poisoning by monitoring attention collapse dynamics in language model generations.
Energy-Guided Flow Matching improves generative quality by progressively revealing high-frequency details through a moving endpoint and adaptive scheduling, reducing training cost and achieving state-of-the-art FID scores.
TFO enables frozen vision-language models to perform speech-centric audio-visual understanding via modular transcript routing without retraining or architectural changes.
Its linear position bias scales down so aggressively at long context that it underflows float precision, quietly zeroing out attention weights entirely. The affected heads don't degrade, they go dark. If you're running anything built on ALiBi (BLOOM, MPT, and kin) at long context, that's not a training problem you can fix with more data. It's arithmetic.
This one scored frontier models live, day by day, across all 39 days of the 2026 FIFA World Cup, so there was nothing to leak. If you've been skeptical of "our model predicts the future" claims, this is the harder, honest version of that test.
One paper puts that question to a correctness bar instead of a demo. Compilers skip optimizations when the semantics that would justify them aren't visible in the code they analyze. The authors ask whether an LLM can pull those semantics out of the surrounding C/C++ context and turn them into transformations that provably preserve the program's contract, not just plausible-looking rewrites.
We present MameLoshnLM, the first open-source 8B-parameter language model built specifically for Yiddish. Despite Yiddish's rich textual tradition, its limited digital presence and the scarcity of reliable evaluation resources have…
Classical continual learning (CL) has primarily focused on enabling models to update and retain knowledge through parameter-centric mechanisms, e.g., training strategies, architectural designs, and weight adaptation. However, emerging…
Multilingual text embedding models are commonly adapted using a single training objective across diverse tasks, despite different tasks requiring fundamentally different optimization strategies. We introduce Task-Conditional Flow Matching…
Economic World Models (EWMs) are generative economic models that simulate how economies evolve from within by modeling heterogeneous agents, their beliefs and actions, and the market and institutional mechanisms through which their…
Current video world models struggle in multiplayer environments because they entangle world state with view-dependent visual latents, leading to redundant compute, view inconsistencies, and poor scalability. We propose MAS (Multiplayer…
Video object removal must eliminate not only the target object but also its induced effects while maintaining high-fidelity and spatiotemporally coherent restoration. Existing methods mainly learn object-effect correspondences implicitly…
End-to-end document parsers provide a unified interface, but serialize page layouts and regional contents into one autoregressive sequence. This formulation forces independent regions onto a decoding path whose length grows with the total…
Training large language model agents for long-horizon tool use typically relies on interactions with real or synthesized executable environments, whose construction and verification are costly, or on external simulators that are difficult…
Training terminal agents requires executable and verifiable tasks that are not merely solvable, but appropriately challenging for learning. Executable validation establishes feasibility, yet does not reveal how a task behaves relative to a…
Reinforcement learning (RL) with verifiable rewards constructs trajectory-level advantage estimates, yet it often fails to credit the few pivotal decisions that determine outcomes in long-horizon, multi-turn agentic tasks. Recent work…
Spatial intelligence is fundamental to embodied agents, yet existing benchmarks focus on local spatial perception from single or few viewpoints, overlooking global spatial awareness over continuous, long-horizon visual streams. To address…
On-Policy Distillation (OPD) is emerging as a promising alternative to reinforcement learning for LLM post-training, yet its effectiveness in multilingual settings remains underexplored. We study OPD and its advanced variant, On-Policy…
Unified multimodal retrieval aims to identify candidates that satisfy complex user intent expressed through heterogeneous inputs. Although Large Vision-Language Model (LVLM)-based retrievers are efficient and scalable, directly encoding…
Multimodal large language models excel at passive perception but struggle with complex visual cognitive tasks requiring multi-step temporal reasoning. This degradation largely stems from the inherent ambiguity of language-based reasoning,…
Vision-Language-Action (VLA) models have become a powerful paradigm for robot manipulation, but training a single generalist policy for heterogeneous robot embodiments remains an open problem. Existing methods have two main limitations.…
Latent diffusion modeling (LDM), a prominent paradigm, utilizes tokenizers to map input signal to compressed representation. This dependency positions tokenizer as an integral part of generation process itself, since it affects learning…
Computer-use agents pay full frontier inference to re-derive routines their user has already performed, because an agent's memory today records what the user said, not what the user did. We compile passively captured screen activity into…
Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in emotional intelligence. However, prevailing research predominantly focuses on task-specific specialization, often neglecting inter-task synergy and…
Recent works train agents by constructing large-scale multimodal environment pools. However, we find that simply increasing the number of multimodal environments does not always benefit. We further analyze the limitations in current…
Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) exhibit fundamentally different behaviors in enhancing multi-task reasoning for large language models (LLMs). Our preliminary experiments revealed a phenomenon: SFT suffers from…
Creative capabilities of MLLMs matter in design, communication, education, and human--AI collaboration, yet remain difficult to evaluate because explicit targets and reward signals are scarce compared with accuracy-oriented tasks.…
Personality-conditioned LLM agents (PC-Agents) are increasingly used in emotional support, social simulation, and role-playing, motivating the development of lifelong agents that remain coherent over extended interactions. A key component…
LLM-based agents are rapidly advancing, autonomously invoking external tools to complete multi-step tasks for users. However, agents often acquire more sensitive information than the task requires. Existing privacy benchmarks audit what…
Aerial image-goal navigation requires an unmanned aerial vehicle (UAV) to reach a target location specified by a goal image. Existing world-model-based methods rank candidate trajectories using predicted futures, but typically rely on only…
Deploying autonomous multimodal agents in continuous, real-world environments requires them to ingest unbounded audio-visual streams and maintain hour-scale memory. However, current evaluations predominantly rely on brief clips and…
The widespread integration of AI coding assistants offers undeniable boosts to engineering velocity. Yet, recent studies point to a growing trade-off, revealing persistent challenges with code quality and maintainability. Industry leaders,…
Integrating interpretability as a training constraint yields scalable, disentangled representations that enable attribution, retrieval, and steering without retraining.
Factorized Hypothesis Search improves retrieval of implicit concepts by maintaining parallel partial interpretations across semantic dimensions for structured querying and verification.
Enfold internalizes future-generative computation into a present-only representation to enable fast predictive control without running the generator at each step.
Decoupling CLI agent scaffolding enables cross-scaffold evaluation and shows that learning planning improves generalization across different agent environments.
Gated Hindsight Distillation improves GUI agent training by using future screenshots as privileged evidence to recover correct reasoning when standard imitation fails.
SiPE integrates a lightweight syntactic prior from dependency parses into positional embeddings across transformer architectures, improving syntactic generalization and language understanding without altering self-attention or increasing…
Visual tool-use in multimodal LLMs often lacks causal effectiveness, with returned observations frequently failing to influence answers or being used incoherently despite aggregate accuracy improvements.
SkillZip compresses reusable procedural skills into contract-preserving, executable graph units to enable efficient retrieval and expansion under limited context budgets.
Rare on easy problems (2.2%), it climbs to 28.3% on Olympiad-level questions and 37.4% on the hardest science benchmark tested.
Detection tools watch one conversation at a time, but nothing stops an attacker from splitting a harmful goal into a string of sessions that each look harmless on their own, then assembling the dangerous result themselves.
Skip the translation and store the tokens directly.
Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agents typically treat all steps within a…
GUI agents must remember both useful experience from earlier tasks and unfinished progress in the current interaction. Latent memory offers a compact solution by compressing multimodal trajectories into a few continuous tokens. Existing…
Prompt injection poses significant security risks to LLM agents. Efficient and effective red-teaming is therefore critical, both for evaluating these risks and for collecting training data to improve defenses. Existing state-of-the-art…
As chart images, tabular data, and visualization code play increasingly important roles across diverse domains, cross-representation understanding across these modalities poses fundamental challenges for AI systems: the relationships…
This technical report presents K-EXAONE 2.0, an open-weight multilingual foundation model developed by LG AI Research as a step in our effort toward global frontier-scale foundation models. Rather than training from scratch, we upcycle…
Leveraging pre-trained vision-language models (VLMs) to construct vision-language-action (VLA) models has emerged as a promising paradigm for 3D robot manipulation. However, existing 3D VLA methods remain data-hungry, exhibit limited…
Interactive video world models are essential for long-horizon planning and exploration, yet they suffer from compounding errors. Post-training methods such as reinforcement learning (RL) can improve these models, but they hit a…
Personalized LLMs with persistent memory are increasingly deployed, yet the faithfulness of their user models remains unexamined. We study over-inference (OI): the phenomenon where LLMs fabricate user attributes beyond what evidence…
Long-horizon reasoning in recent LLMs demands that the model switch between distinct skills inside a reasoning chain, such as first doing a math derivation, then using the result to plan a schedule. We call such problems cross-skill…
Memory-augmented VLM agents act on persistent spatial knowledge, yet that knowledge silently goes stale as the environment changes. We ask what happens when an agent must reconcile a confident memory claim with a contradicting observation,…
We introduce NOLLI, a procedurally generated English-Korean puzzle benchmark designed to diagnose where Korean performance gaps arise. It comprises 15 puzzle types (25 tasks; 7,500 items), with every instance seed-regenerable, verified to…
The abundance of casually captured monocular videos and images on social media provides a valuable source for immersive content creation, where generating novel views from such sparse observations can greatly enhance user experiences.…
Text-to-image (T2I) models can produce visually compelling images, yet they remain limited on open-world tasks that require complex semantic understanding, multi-step reasoning, and the integration of external world knowledge. Existing…
Vision offers a critical axis for advancing foundation models, driving a shift towards natively unified multimodal pretraining. Despite this momentum, the design space and the fundamental mechanisms of how modalities interact during…
Leading open text-to-image models often carry complementary strengths: one may lead on preference-aligned aesthetics while another follows compositional instructions more faithfully. However, differences in their autoencoders and noise…
Despite the remarkable recent progress of video world models, social interaction between users and the characters within these worlds remains unsupported. To fill this gap, we present HelloWorld, a video world model that enables social…
On-Policy Self-Distillation (OPSD) has become a standard post-training approach for improving visual reasoning in multimodal large language models (MLLMs). Existing methods draw privileged information from diverse input sources to guide…
Modern Greek is absent from NVIDIA's Nemotron retrieval models and from major multilingual retrieval benchmarks, despite being important for retrieval-augmented generation (RAG) in legal, energy, financial, and medical applications. We…
Recent video models increasingly support generation, reference conditioning, and editing within a single model, yet typically expose them as separate operations over fixed inputs. Practical creation unfolds across multiple shots, requiring…
Deep vision models exploit shortcuts, relying on cues that correlate with supervision signals. Prior work has focused on visible biases, such as object-background or texture correlations. We identify a different source of shortcut…
Vision-language-action (VLA) models often treat main-view and wrist-view observations as parallel visual inputs, overlooking their distinct roles in robot manipulation. Fine-grained manipulation, however, benefits from anticipating how…
Understanding 3D scenes is fundamental to embodied intelligence, requiring joint reasoning over heterogeneous information from multiple modalities, including visual and geometric cues. However, the relevance of these modalities often…
Generating large-scale, freely explorable 3D worlds from open-ended text remains challenging because a system must jointly maintain global spatial coherence, rich local content, and explicit assets suitable for downstream editing and…
High-quality long-horizon training data for terminal agents is expensive to produce, often costing hundreds to thousands of dollars per task, because each task must keep the instruction, environment, reference solution, and verifier…
Collaborative music agents need internal representations rich enough to support both understanding and generation, yet flexible enough for a workflow where the human retains agency. We present a hierarchical self-supervised ``world model''…
Privileged on-policy distillation provides dense supervision for multi-turn agents by allowing a synchronized teacher to re-score the student's response at every turn with access to training-only references, such as successful…
SPOT improves on-policy distillation by selectively probing uncertain positions and calibrating targets to downstream outcomes, boosting reasoning quality and coverage.
CLIP-CC-Bench evaluates long-form video description using expert paragraph references and ensemble LLM embeddings to benchmark 17 video-language models.
A recurrent Transformer with fixed-size memory and coupled prefiller-decoder training improves long-context modeling while enabling efficient parallel training and reduced inference cost.
Across 154K real agent calls, that re-tokenization step ate up to 64% of time-to-first-token. TokTier tokenizes just the new window and splices it onto the cached state instead, dropping median TTFT 16-34% under vLLM.
It keeps full interaction history in an entity graph plus a timeline, then filters deterministically at query time with no model calls in the retrieval path.
Researchers ran vision-language models through a "spot the difference" task where each model sees a different image and has to argue it out over dialogue. Models routinely folded, agreeing the images matched even when their own image plainly disagreed. A steering vector trained to suppress sycophancy cut those errors.
Modern agent frameworks equip large language models with external skill libraries to solve complex tasks. However, it remains unclear whether these systems can effectively evolve their skills and whether the resulting skills improve…
Diffusion language models (dLLMs) offer an alternative to autoregressive (AR) language modeling, yet the scaling behavior of Mixture-of-Experts (MoE) dLLMs remains poorly understood. We systematically characterize how optimization…
Persona skills distill personal interaction histories into portable and executable artifacts for downstream agents. While enabling flexible personalization, this process concentrates fragmented personal signals, amplifies their impact…
We introduce Video-DeepResearch (Video-DR), extending multimodal agents from static images to continuous video streams, a setting that demands dense spatiotemporal grounding coupled with open-web exploration. Preliminary evaluations reveal…
Recursive self-improvement requires agents to turn accumulated experience into better future behavior. Personal AI agents offer a concrete setting for studying this capability because they retain preferences, task histories, tool routines,…
Omni-modal large language models (Omni-LLMs) have achieved remarkable performance on audio-visual understanding tasks, but processing long and highly redundant visual and audio token sequences incurs substantial computational overhead,…
Tool-Integrated Reasoning (TIR) enables LLMs to solve complex tasks through iterative tool interactions. However, existing reinforcement learning methods often rely on trajectory-level supervision, limiting fine-grained credit assignment…
Real-time video editing requires low-latency causal generation with bounded computational resources while preserving source fidelity and long-term temporal consistency. We present JoyAI-Video-Edit, a 16B-parameter autoregressive diffusion…
Self-evolving agents increasingly convert interaction histories into reusable skills that persist beyond individual tasks. While prior work studies memory and retrieval poisoning, such attacks only affect agents when poisoned records are…
We introduce UniWorld-Design, a framework that redefines image generation from flat pixel synthesis to structured visual composition, with semantic RGBA layers as the atomic units of generation, understanding, and editing. Our key insight…
On-policy distillation, in which a teacher corrects samples that the student itself generates, presupposes that the two models speak the same language: identical VAE latents, matching architectures, and a common timestep grid. We ask what…
Large language models are increasingly embedded in software engineering workflows as coding agents that can inspect repositories, invoke tools, execute tests, debug failures, and generate patches. Yet most existing agents remain largely…
Agent self-evolution updates an agent's persistent state from prior experience and reuses it to solve related tasks more effectively. Evaluating self-evolution is difficult: existing benchmarks provide limited coverage of economically…
LLM agents are increasingly applied to open-ended everyday requests that span work, study, and life. These tasks are long-horizon, cross-environment, and multimodal, forcing the agent to preserve goals and constraints across many steps…
On-Policy distillation (OPD) transfers teacher capabilities by supervising student-sampled trajectories with dense token-level teacher signals. Recent selective OPD methods improve this process by prioritizing signals that are confident,…
Skill-based prompting has become a practical mechanism for improving large language model (LLM) agents, yet existing skill acquisition methods often treat skills as experience summaries, memory entries, or direct summaries of successful…
Molecular glue degraders have emerged as a promising strategy for targeted protein degradation by inducing ternary complex formation between an E3 ubiquitin ligase and a target protein. Despite their therapeutic potential, computational…
Reasoning language models frequently overthink: generating extended chains of behaviors such as hedging, approach abandonment, and self contradiction that consume tokens without improving answers. We show that these behaviors are not…
Self-consistency assumes the most frequent answer among sampled reasoning traces is the most reliable, but this can fail in causal reasoning: samples often repeat the same confounding error, and votes fragment across multiple valid…
Multimodal large language models (MLLMs) make grounded predictions in real-world scenes by combining visual and textual cues, yet existing benchmarks rarely reveal how they arbitrate between these evidence sources when they conflict. We…
A framework that persists execution state so a run can be interrupted, survive a crash, and continue must decide what a resume means for effects that already fired. Five widely deployed agent workflow frameworks answer differently, none…
Data agents enable natural-language analytics over organizational workspaces, where relevant evidence may be scattered across databases, structured files, long documents, and multimedia. Existing benchmarks largely isolate structured…
Small language models are often the only option for deployment under tight latency, cost, and on-premises constraints, but they are rarely trained from scratch: a compressed model is usually recovered through knowledge distillation (KD).…
Foundation models are used to extract transferable representations from large amounts of unlabeled data, typically via self-supervised learning (SSL). However, many of these models rely on architectures that offer limited interpretability,…
Audio reasoning is essential for machine understanding of the acoustic world. Reinforcement learning with verifiable rewards can elicit such reasoning, yet existing reward designs are complementary in their limitations: outcome-based…
WeClawArena is an auditable benchmark and sandbox for evaluating multi-party agent collaboration across personal workspaces, measuring both task utility and security attack success.
MatrAIx is a large-scale simulated-user evaluation framework that uses diverse persona records and interactive environments to test AI systems across many domains.
A new jigsaw benchmark with interlocking pieces reveals that vision-language models fail at geometric reasoning and suffer a sharp performance drop as puzzle size increases.
The fix is a benchmark plus a training set (OS-Shepherd-100K) for open reward models that match commercial-judge accuracy at 30-60% lower cost. If you're using an LLM to grade your own agent's trajectories, that leniency bias is probably already inflating your success rate.
Tested on Qwen3-VL, UI-TARS-1.5, and OpenCUA running under hardware constraints, extra context and extra steps mostly extend the same broken trajectories instead of fixing them, and the gains from longer context eventually flip into confidently wrong "task complete" calls. Parallel sampling helps some, at a compute cost most local deployments can't afford.
It recovered 29.2% more usable tasks than PR-based baselines and built working tasks 79.6% of the time, across bug fixes, feature adds, test generation, API migrations, and security repairs.
Optimization-based latent reasoning improves large language model outputs by optimizing instance-specific continuous states at test time while keeping model parameters frozen. Existing methods, however, typically connect these states to…
Multimodal large language models (MLLMs) have advanced visual understanding and reasoning, yet their static parametric knowledge limits their ability to address knowledge-intensive and dynamically evolving open-world problems. To move…
Video motion transfer aims to animate a target object using dynamics from a reference video. Existing formulations largely rely on fixed structural correspondence, which becomes ill-defined when reference and target objects differ…
On-policy (self) distillation (OPSD) is increasingly adopted for language-model post-training. It strengthens the teacher with privileged information but can induce a privilege illusion: the student learns privilege-dependent behavior it…
Large language model (LLM) agents increasingly undertake long-horizon tasks that require sustained reasoning, tool use, and revision across many interdependent steps. However, existing agent harnesses maintain task execution, task state,…
Existing skill generation methods largely rely on heuristics or pipeline-style consolidation, which must be specially designed for different evidence sources. In contrast, learning-based approaches offer a more unified way to model skill…
Agent skills have become an important mechanism for equipping language-model agents with reusable procedural knowledge. However, providing skills alone does not guarantee that current models can effectively identify, apply, and coordinate…
Controllable video generation models are increasingly being developed as world models. Accordingly, evaluating them in this role extends beyond the apparent appearance of generated videos to the inherent reactivity of the worlds they…
Real-world software development requires coding agents to operate in shared workspaces where users may inspect and modify code during an ongoing task, yet existing repository-level benchmarks typically evaluate agents working alone or…
Sparse retrieval underpins modern search systems, from web search to retrieval-augmented generation. Existing work has introduced Learned Sparse Retrieval (LSR) to push beyond exact lexical matching toward richer semantics. Yet LSR has so…
Speech and audio generation is often needed in animation dubbing, audio drama, movies, advertising, games, podcasts, and short-video production. In these scenarios, creators may need to design voices without reference recordings, control…
Fixed-layout indoor furniture styling requires selecting assets that form a coherent room without changing the prescribed furniture categories, positions, orientations, or scales. Existing approaches typically retrieve each asset…
To operate robustly in open-world environments, autonomous agents should be able to infer the behavior of unfamiliar systems through interaction alone, even in the absence of documentation. However, existing tool-use benchmarks expose…
Existing indoor layout generators produce globally plausible layouts yet may retain local violations such as collisions, out-of-bounds placements, obstructed openings, and blocked circulation. Most prior work focuses on full-scene…
Recent Vision-Language-Action (VLA) models for autonomous driving (AD) increasingly utilize chain-of-thought (CoT) supervision to enhance the reasoning capabilities of their Vision-Language Model (VLM) components, yet existing annotation…
Scientific poster construction compresses a long multimodal paper into a readable, editable canvas. Existing systems hide request-level failures by scoring only completed outputs; direct image generation is not element-editable, while…
Continually improving agents require dynamic interaction feedback beyond static supervision, yet direct real-environment interaction is costly, slow, unsafe, and hard to parallelize. World modeling offers a natural intermediate proxy that…
Industrial recommenders increasingly adopt the pretrain-then-transfer paradigm, yet behavioral distribution drift raises two questions: what to learn from behavior sequences, and how to transfer the learned knowledge while the pretrained…
Large language model agents have shown strong potential in complex interactive tasks, yet their reinforcement learning (RL) is often hindered by sparse rewards, as a long multi-turn trajectory may receive only a single outcome-level…
Language remains an outlier in generative modeling: while images, video, and audio are increasingly modeled in continuous latent spaces, text generation still relies predominantly on discrete tokens. Existing continuous language models…
Captions serve as a primary supervision signal for both multimodal understanding and text-to-image generation. However, previous evaluations treat the caption quality as a single scalar objective, which conflates two distinct properties:…
Recent advances in image generation have demonstrated the potential of unified multimodal models that integrate understanding, generation, and editing. However, unified 3D modeling remains constrained by scarce multimodal data,…
A wearable assistant should both answer questions about its visual history and recognize when that history is useful to the present situation. Existing video-memory systems primarily support question-conditioned recall, whereas proactive…
Single-image feed-forward 3D Gaussian Splatting (3DGS) aims to directly generate a renderable 3D scene representation from one input image, avoiding the cost of multi-view capture and per-scene optimization. However, existing methods are…
Enterprise question answering requires models to acquire proprietary knowledge without discarding general capabilities. We present Wnuan, a three-stage pipeline that constructs task-oriented supervision from documents, performs supervised…
Learning generalizable robot manipulation policies requires large-scale and diverse demonstration data. Egocentric human manipulation videos offer rich scene and task diversity, and prior work has shown that retargeting and rendering such…
Existing deep-research agents use a search-visit workflow that retrieves and reads whole pages, without considering the addressable structure that web sources expose through titles, headings, sections, and metadata. This prevents agents…
Weight-only quantization substantially reduces the storage of large language model (LLM) transformer blocks, but practical backends often retain the final language-modeling head (LM-head) in BF16 or FP16. Quantizing this projection naively…
Model checkpoints are growing in both number and size, which makes archival, transfer, and deployment increasingly costly. General-purpose compressors can reduce storage requirements but ignore tensor structure, whereas existing…
Robot learning is splitting into two bets: policies that bake competence into frozen weights (vision-language-action, or VLA, models), and agents that write and refine their own executable skills as code. This survey organises the field…
Multimodal representation learning is a cornerstone of modern AI. By encoding multimodal queries and targets into vectors, it powers industrial search and recommendation and underpins modern agents. Real-world platforms with complex…
A three-stage multimodal framework improves follow-up edit recommendations in image-creation conversations by combining supervised fine-tuning, multi-objective reinforcement learning, and visual verification.
It's an open, full-stack benchmark where a model's job is improving how ML systems get built, with verifiable tasks instead of vibes. If AI-improving-AI is actually coming, better that anyone can run the experiment than take a lab's word for it.
Most multi-agent frameworks treat topology like a fixed org chart. Letting the agents redraw it themselves targets a coordination bottleneck that quietly caps how far agent teams scale.
Fine-tune a model to stop claiming it's conscious, and its willingness to attribute minds to other people, animals, and other AI drops too, along with its human-values alignment. Train it the other way, toward asserting consciousness, and those beliefs come back. The safety patch isn't narrow. It's reaching into the model's whole theory of mind.
Recent 3D vision-language models (3D VLMs) construct geometry aware tokens by projecting 2D visual features into world coordinates, enabling spatial reasoning for tasks such as 3D question answering. However, this design generates…
Video world models predict future observations conditioned on historical observations and control signals, enabling long-horizon generation through autoregressive state transitions. Unlike conventional video generation models that…
Selecting a complete 3D object from a reconstructed scene with minimal user effort is essential for practical scene editing and embodied interaction. Existing 3DGS-based methods either retrain the Gaussian representation to embed…
World models have attracted significant attention for their ability to capture and predict the structure and dynamics of the physical world. In this emerging landscape, Joint Embedding Predictive Architectures (JEPA) offer a particularly…
Short segments of perceived speech can be retrieved from non-invasive magnetoencephalographic (MEG) recordings by deep networks trained with a CLIP-style objective against wav2vec 2.0 audio embeddings. Yet their weights do not map onto…
When a dog opens its mouth and barks, humans naturally recognize what the sound is and when it occurs. Building audio-visual models with this same ability requires representations that capture both semantic and temporal alignment. Current…
Autoregressive models accumulate error over long rollouts, yet at deployment there is no ground truth to measure it against. We train a single conditional latent diffusion model that steps a dynamical system forward or backward in time via…
Scientific collaboration with AI agents requires studying human-agent pairs to avoid risks like reduced inquiry diversity and to foster synergistic discovery.