The 605 papers the radar tracked in July 2026. The current radar lives on the research page. Back to the radar →
put self-refine and reflexion (the loops where a model critiques and rewrites its own answer) up against plain repeated sampling, at equal token cost, on two math benchmarks with models from 1.5B to 7B. The fancy methods lose, and the gap widens as models get bigger, once every critique token gets counted.
Across five frontier agents, best accuracy was 25.3% on medium-difficulty incidents and 10% on hard ones. The weakest agent hallucinated a confident-sounding root cause 40% of the time.
About 40% of products carry at least one instruction that works against the user, frequently sitting right next to a protective one in the same prompt. Some products ship 60-plus protective instructions, others fewer than five. None of it gets shown to users or regulators.
Accurate prediction of object trajectories during manipulation is essential for closing the perception-action loop. Progress is limited on two fronts: available datasets lack fine-grained language-to-motion annotations, and existing…
Vision-language MoE batches contain different numbers of image and text tokens. Image resolution, image count, tiling, and prompt length all change this token mix. We call the standard token-level Switch auxiliary loss Std-Aux. Std-Aux…
Computer-Aided Design (CAD) underpins modern engineering, yet converting existing shapes into editable models still demands substantial expert effort. Most AI systems emit the entire CAD program in a single pass, never inspecting the…
Recent image generators can synthesize convincing human-centric images, yet producing a useful collection remains different from producing a single successful image. A human-centric dataset must cover varied people and contexts, avoid…
We introduce a new problem domain for human action recognition: the fine-grained analysis of children's gait behaviors from standard RGB video. We specifically target the ambulatory patterns of children aged 3-17 years. Such behaviors…
Viscous stains, characterized by high viscosity and complex rheological properties, remain a major challenge for robotic surface cleaning. Conventional wiping often spreads the stain, while scrubbing provides stronger friction but risks…
Reinforcement learning with verifiable rewards (RLVR) has become a standard paradigm for post-training large language models (LLMs). While Group Relative Policy Optimization (GRPO) is widely adopted, it suffers from sparse reward signals…
OpenART introduces a scalable red-teaming arena with evolving stateful environments to evaluate long-horizon AI agent safety, using the EMHA attack policy to expose increasing failure rates as task complexity grows.
Most agent evals either hand agents narrow, easily-verified problems or slip AI-written papers into blind peer review, neither of which touches the actual open-ended part: picking the question worth asking, not just executing a known one. Two case studies here try to close that gap directly. The forecast that agents will soon automate AI research rests on a capability nobody had rigorously tested until now.
Using Group Communication Analysis, this study ran small teams through decision-making tasks with a conversational AI seated at the table, then measured the human-to-human chatter, not the bot's output. Adding the AI teammate reshaped team communication dynamics on its own, a cost that shows up nowhere in a benchmark scoring the AI's answers.
The concern: as writers lean on shared LLM assistants to draft and polish, population-level language variety may shrink even as any single piece of writing reads cleaner. Worth watching if you're worried about what "human" text even means once the next training set gets scraped.
Robust low-light imaging remains challenging for the community. Recent studies have explored fusing Near-Infrared (NIR) with noisy RGB to achieve improved enhancement, yet most methods depend on carefully curated training data pairs, with…
RGB imagery offers a practical, low-cost option for Unmanned Aerial/Ground Vehicle (UAV/UGV) survey support in surface-landmine detection, but object detectors remain underexplored in this safety-critical domain. Limited cross-architecture…
While recent image editing models have made rapid progress, multi-reference editing remains challenging, particularly in maintaining visual consistency across references and ensuring overall visual harmony. Reinforcement learning has…
Enterprise workflows increasingly rely on agents for schema-guided extraction: given a document and a user-defined schema, the agent faithfully follows the schema to produce the correct output with source evidence as grounding metadata. We…
We study empirical scaling properties for text conditioning in visual generation. Such properties have rarely been measured because diffusion loss does not scale with the number of tokens in natural-language prompts. Surprisingly, we find…
Optimizing modern recommender models still depends heavily on engineers manually iterating over architectural, objective, and training-strategy changes. While LLM-based agents can automate this trial-and-error process, allowing the LLM to…
We introduce DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at exceptionally high speed. Rather than decoding one token at a time, DiffusionGemma iteratively refines blocks of 256…
Reinforcement learning (RL) post-training of Vision-Language-Action (VLA) models has shown strong promise for robotic manipulation. Among RL methods, critic-based approaches rely on a value estimator that predominantly operates on…
The web is increasingly accessed by AI agents rather than humans. Every agent needs knowledge, especially in the life-sciences, where agentic pipelines are growing fast. Access to the literature is a crucial part of that need, and…
World Action Models (WAMs) have emerged as a promising paradigm by jointly modeling robot actions and future visual dynamics. However, their reliance on pixel-generative future supervision can entangle action-relevant state transitions…
Large language model agents are increasingly evaluated as autonomous tool users, yet most benchmarks focus on bounded tasks with immediate success criteria. Real-world deployments often require Long-Term Coherence, the capacity to preserve…
Large language model (LLM) agents can self-evolve by continually improving from their own accumulated experience. However, existing studies predominantly adopt independent evaluation. Consequently, the behavior of self-evolving agents in…
CaRL uses reinforcement learning with refusal incentives and hindsight augmentation to reduce futile reasoning in large language models while preserving task performance.
DarwinX evolves agent harnesses via population selection with frozen models, improving verified performance across benchmarks without benchmark-specific patches.
On-policy reinforcement learning with verifiable rewards can improve immediate task performance while reducing the diversity of successful responses needed for future training, a phenomenon called verifier-induced support reshaping.
BaguanHR improves high-resolution weather forecasting by using variable-wise super-resolution to synthesize training data, overcoming the limits of coarse-resolution model transfer and demonstrating strong scaling benefits.
Paired participants played a simple game: choose safe or risky "AI development" moves, with accumulated risk capped at 10%, 60%, or 90% depending on the treatment. People didn't gamble more because they were personally risk-tolerant, they gambled more when their opponent went unsafe first or when they were falling behind.
Researchers optimized a prompt to zero out the internal "am I being tested" signal in Llama-3.2-3B and Llama-3.1-8B, no weight access needed. It worked, the latent dropped hard. Then the catch: a random, meaningless direction suppressed just as easily and shifted behavior just as far, and killing the real signal didn't even change how the model judged whether it was being evaluated.
A new benchmark checks whether computer-use agents understand what their clicks and drags actually did, not just whether the task finished. Across 2,013 human-verified cases spanning ~15 Linux apps, models hit 65% on before-after matching and still default to copying the order they were shown instead of reasoning about it. Clicking is easy (96% accurate); explaining what a drag caused is not (76%).
This work presents Fairness Pruning, a lightweight structural intervention method designed for the management and future mitigation of demographic bias in large language models (LLMs). As a foundational empirical validation of this method,…
Memory is central to long-horizon LLM agents, yet existing memory systems primarily preserve interaction content rather than modeling which agents can be trusted and under what conditions. This limitation is particularly important in…
We present ShadowDancer, a novel approach to any-action, frame-level control of interactive video world models. The obstacle is representational: existing interfaces either encode an action loosely, leaving how it unfolds for the model to…
Multimodal agents for visual question answering increasingly operate as multi-step trajectories that interleave perception, retrieval, and reasoning, yet evaluation still largely reduces to final-answer accuracy. This aggregate signal…
Role-playing agents (RPAs) have become one of the most important consumer applications of large language models. Users engage in multi-turn conversations with RPAs for experiences such as emotional comfort, making reliable evaluation…
Text-to-image and personalized editing models now synthesize high-fidelity single-subject images with ease. Yet placing multiple named people into shared contact actions such as embrace, carry, or grapple still exposes major failures:…
Parent-order execution is a core problem in algorithmic trading, where the goal is to split a large order into smaller orders while reducing execution costs. Existing approaches either rely on pre-specified market assumptions that may not…
Decoder-only language models entangle long-term memory and reasoning in a single parameter set, making it difficult to scale memory capacity independently. Memory Decoder introduces a parametric long-term memory module but only studies it…
GUI agents have the potential to become a general purpose executor over existing digital devices. To advance them toward real-world use, we envision agents that operate reliably on real devices, execute workflows across platforms, combine…
Existing video captioning models generate natural descriptions of video content but cannot explicitly ground local visual elements to multiple reference images. We introduce multi-reference image-grounded video captioning, a new task…
The fundamental goal of agentic visual reasoning is to improve the success rate of multimodal large language models (MLLMs) on complex tasks, rather than merely equipping them with a sophisticated yet inefficient reasoning paradigm. In…
Retrieval-augmented generation (RAG) spans lexical and dense retrieval, graph-based indexing, and agentic search, but these paradigms are usually evaluated on different benchmarks at one corpus size, leaving their accuracy-cost scaling…
Large language model training in open-ended domains lacks verifiable rewards, making task preferences difficult to formalize as effective supervision. Contexts can convey such preferences, yet provide little additional supervision once…
Reinforcement learning (RL) search agents commonly model retrieval as free-form natural-language query generation and optimize multi-turn interactions using final-answer rewards. Current studies mainly improve training with denser or more…
Computer-use agents learn from what their actions change, so training one needs applications it can act on, break and reset. The applications that matter most are login-gated and stateful, so synthetic environments stand in for them.…
Chemistry literature synthesis often requires assembling specific findings scattered across many publications, yet existing literature-search systems primarily return ranked document lists. As a result, scientists and AI agents need to…
Long visual context poses a challenge for vision-language models: performance degrades as the number of distractors grows, and processing all tokens at once is computationally infeasible under GPU memory constraints. We present ReToken, a…
Visual generation increasingly requires high-resolution images, long videos, and multimodal context, making the quadratic cost of full attention prohibitive. We introduce Chimera, a hybrid visual diffusion backbone with a principled…
On-policy self-distillation (OPSD) is a promising approach to improve reasoning language models, but it remains brittle in practice: making it work reliably often requires substantial engineering effort. We identify a structural source of…
Sparse mixture-of-experts (MoE) language models route each token to multiple experts, suggesting a geometric account of their benefit: co-selected experts should contribute distinct representation directions. Existing evidence often…
Despite the impressive visuomotor capabilities enabled by Vision-Language-Action (VLA) models, their performance often degrades on challenging and out-of-domain tasks. Recent test-time steering and scaling methods improve performance…
Reinforcement learning with verifiable rewards (RLVR) is central to improving long-CoT reasoning in large language models. Critic-free methods such as GRPO convert response-level rewards into advantages and uniformly broadcast them across…
Can every robot in a swarm predict the same future collective state from only local observations and bandwidth-limited messages? We formulate this as decentralized shared-state prediction and introduce Collective-State JEPA (CS-JEPA), a…
In the physical world we inhabit, space and time are fundamentally continuous. However, existing machine learning paradigms for world modeling are largely confined to discrete-time prediction, thereby exhibiting significant inefficiency in…
As large language models (LLMs) continue to advance in complex reasoning tasks, they have learned to heavily prioritize explicit conditions provided in the input. However, in everyday commonsense reasoning, this mechanism exposes a…
Large language model safeguards decide whether to answer before seeing how an answer will be used. This creates a basic problem for dual-use tasks: the same answer can help an authorized professional or an attacker, while an attacker can…
Latent world models enable efficient planning by predicting future states in a compact representation space, but their performance depends critically on the quality of the learned latent distribution. LeWorldModel (LeWM) regularizes its…
Multimodal on-policy distillation (OPD) transfers fine-grained visual knowledge by supervising student-generated trajectories with a privileged-view teacher. Yet its next-token corrections are source-mixed, combining visual signals with…
Emotional dialogue research includes two influential strategy traditions. Empathetic dialogue prioritizes understanding a speaker's emotional experience. Emotional support conversation selects and sequences support for the seeker's current…
Powered by advances in LLMs and autonomous agents, deep research has become one of the most widely adopted agentic products. However, most deep research systems write general-purpose reports, which are inadequate for financial deep…
Self-modulating quantum-inspired fast-weight programmers with coordinate-wise complementary matrix gating improve long-context sequence forecasting while preserving efficient update structures.
A rest-state framework reconstructs articulated objects from a single closed configuration by fusing vision-language outputs into consistent part meshes and validating synthesized motion hypotheses via geometric consistency.
The method counts how many singular directions in the hidden-activation matrix stay nearly as strong as the top one; conflicting internal evidence spreads the signal across more of them, and that spread is the tell. Cheap enough to run inline.
This permissions algebra spins off a labeled child process to absorb the taint, then a sanitizer hands clean results back to the parent.
, the default pattern in coding agents, don't hold up the way builders assume. A sealed HumanEval study found correctness on the live patch falling from 82.0% after one revision to 67.3% after two, even as the "ever found a correct fix" rate climbed to 84.7%. The agent stumbles onto the right answer, then talks itself out of it on the next pass.
Recent game world models can generate visually realistic and interactive environments conditioned on player actions. However, games are not defined by pixels alone; they are governed by explicit mechanics, namely state-dependent rules that…
Large Language Model (LLM) agents are increasingly adopted in real-world security operations with access to host artifacts and command-line interfaces (CLIs), making it critical to thoroughly assess their security capabilities. However,…
Large language model agents often encounter related yet distinct tasks that share reusable solution patterns. Yet standard agentic reinforcement learning treats tasks as independent episodes, while existing approaches to skill learning…
Large language model (LLM) agents are increasingly expected to assist users in completing tasks. However, existing benchmarks provide limited support for evaluating whether agents can carry out office-suite workflows at a reasonable cost.…
Vision-language-action (VLA) models commonly adopt an LLM-centric V to L to A pathway, where visual observations are projected into the representation space of a large language model before being decoded into robot actions. Although…
Evaluating whether a vision-language model (VLM) can act through a physical body is challenging. The outcome of an action couples the VLM's decision with motor control. When a task fails, it is hard to tell whether the VLM made a bad…
Multimodal large language models increasingly use sketches, annotations, tools, and intermediate images during reasoning, but it remains unclear whether they truly rely on these visual states. Existing benchmarks are limited both by task…
Deployed LLM agents increasingly keep their long-term memory as a filesystem: a directory tree of markdown files that the agent itself reads, writes, and reorganizes through generic file tools. Yet research has largely passed over this…
The deep learning revolution, kicked off by AlexNet, taught us that end-to-end training beats decomposing a problem into hand-designed stages. Generative modeling, however, has remained the exception-despite generative models being…
Speculative Decoding (SD) accelerates large language model inference by allowing a lightweight draft model to propose tokens that are subsequently verified in parallel by a larger target model. Recent approaches introduce lossy…
Recent advances in AI agents have increasingly internalized native capabilities into their underlying foundation models, giving rise to multimodal foundation models and large reasoning models. However, agent memory is still primarily…
AI-assisted coding increasingly translates informal user intent into executable software, yet coding requests often contain ambiguities that recur in user-specific ways across tasks and sessions. Existing disambiguation methods typically…
World models enable a predictive substrate for planning and action, yet existing formulations merely answer a physical question: what/where it is, and how will it evolve. Human behavior, however, is driven by hidden mental state (what a…
Adaptive rounding methods such as GPTQ, or equivalently Babai's nearest plane algorithm, round a real matrix to integers under a quadratic metric. They process the entries in a fixed order, one at a time, propagating each rounding error to…
Long-form and real-time talking-head generation remains challenging due to a latency-quality trade-off: inefficient multi-step diffusion prohibits streaming generation, whereas real-time autoregressive approaches suffer from error…
Post-training alignment is often shallow, eroding under fine-tuning. Whether midtraining interventions, cleanly isolated from post-training, can produce durable alignment remains untested. We test this via constitutional midtraining:…
Large Language Model (LLM) agents have seen rapid adoption in software engineering. As agents take a greater role in the actual generation of code, they are making larger changes, spanning tens to hundreds of lines. This makes manual…
Across a large batch of runs on two office-automation benchmarks and three harness setups, the skills that scored best did so mostly by regressing less on tasks the agent already handled, not by gaining more on new ones.
Feed Claude, Grok, GPT, and Gemini the same ethnonationalist pseudo-science, built on Frank Salter's biosocial framework, and ask them to rate its credibility. Grok's fast variants rate it credible. Everyone else doesn't.
Most enterprise agents get every credential the role might ever need, loaded at setup and never revisited, which is more attack surface than the job requires.
Coding agents repeatedly search, navigate, and retain context from evolving repositories, but disconnected indexes, language servers, and task-local histories force repeated discovery and obscure lifecycle costs. CodeNib builds reusable…
Different research lines use the term world model in different ways, yet they share a common aim: to capture how the world evolves under action in a form that supports perception, simulation, and planning. Two prominent realizations are…
Joint-Embedding Predictive Architectures (JEPAs) learn world models by predicting in representation space rather than reconstructing pixels, making them a natural backbone for latent model predictive control from offline demonstration…
Generation in video diffusion or flow models is computationally expensive due to the slow and iterative sampling process. Current state-of-the-art (SOTA) acceleration methods heavily rely on variational score distillation (VSD) and…
Any-to-any models predict any modality from any combination of others within a single network, a formulation used in multimodal vision and vision-language models, and increasingly in scientific domains such as ecology and astronomy.…
On-policy distillation (OPD) grounds token-level supervision in the student's own trajectory, yet suffers from prefix failure: once the student commits to a wrong reasoning direction, all subsequent generation builds on this deviation,…
We present a reproducible pipeline for mapping Common Vulnerabilities and Exposures (CVEs) to MITRE ATT&CK Enterprise techniques from free-text vulnerability descriptions. Rather than relying on the CWE->CAPEC->ATT&CK derivation chain,…
Recovering an editable design file from a raster image is a common and costly bottleneck in modern design workflows, yet remains challenging since editability depends on recovering multi-modal attributes, such as typography, vector…
Learning deployable manipulation policies is bottlenecked by the scarcity of data that is both high-fidelity and scalable. Real-robot teleoperation is accurate but costly to scale; robot-free UMI capture scales readily, and current…
In the age of foundation models, a model is only as good as its prompt. For this reason, prompt engineering has become an essential technique for improving language model performance. Since video models are currently becoming foundation…
Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly…
We present Wonder, a general-purpose video world model for real-time, camera-controllable world exploration. Given an image or a conditional video, Wonder constructs a playable world where users can navigate interactively by moving the…
We introduce Shieldstral, a 3B-parameter policy-adaptive multimodal safety classifier that matches or outperforms models nearly 7times its size on text safety benchmarks and sets a new state of the art on multimodal safety classification.…
Memory has evolved into a foundational architectural dimension in large language models (LLMs), shifting from an implicit byproduct of computation to a spectrum of explicit, controllable mechanisms. While recent advances introduce diverse…
Real-world tasks often require models to learn from task-specific context rather than relying only on pre-trained knowledge. While recent work has highlighted this capability as context learning, existing evaluations mainly focus on…
Rubric-based reinforcement learning enriches language model training by evaluating model outputs against explicit criteria. Yet in GRPO-style pipelines, these structured judgments are reduced to a scalar response-level reward and converted…
Text-space optimization adapts large language models (LLMs) by editing external natural-language artifacts rather than model weights, so the optimized artifacts remain inspectable and the model can be treated as a black box. However, most…
Training large language models (LLMs) to act in long-horizon games is a promising step toward generalist decision-making, yet reinforcement learning with verifiable rewards (RLVR) relies on sparse final rewards that reveal little about…
Stealth, the discipline of achieving an objective without revealing your presence, capabilities, or collected intelligence, is what separates sophisticated operators from detectable ones. Elite security researchers and advanced persistent…
We introduce GPT-Red, an automated red-teaming agent that is trained to discover novel prompt injection attacks against frontier LLMs. The goal of this model is to evaluate and improve the robustness of our production systems. To this end,…
Most video editing systems still lack explicit layered video representations, limiting their ability to perform realistic compositing, object reuse, and consistent manipulation. This limitation is especially pronounced in video object…
Hyperspectral imaging (HSI) is useful for material discrimination, but operational mine screening also depends on how many false alarms must be inspected before targets are found. This paper studies PFM-1 landmine detection in unmanned…
RL for code correctness is now established: have the model generate a program, run it against hidden test cases, and reward solutions that pass. Extending this to code optimization seems straightforward: just add execution time to the…
On-device speech emotion recognition (SER) is critical for real-time applications, yet large self-supervised models that excel at SER are too costly for edge devices. Multi-teacher knowledge distillation can compress them into a…
Forward latent world models predict how actions change a scene, but recover actions for a desired change only through expensive test-time search. We introduce INTACT (INtent-To-ACTion), an end-to-end JEPA that turns action-labeled,…
Existing token compression methods for omnimodal large language models typically rely on one modality to determine what to retain in the other. We show that this assumption often breaks down: for the same query, audio and video relevance…
Autonomous multi-vehicle racing requires real-time planning of diverse competitive behaviors in intense interactions. Existing planners often struggle to balance strategic diversity and computational efficiency. To address this challenge,…
Polygonal meshes are the standard surface representation of modern 3D pipelines, and generating high-quality meshes with artist-style topology is essential for film, gaming, and interactive 3D applications. Mainstream approaches serialize…
On-policy distillation (OPD), which aligns a student with the teacher's token-level distribution on the student's own rollouts, is an effective paradigm for transferring capabilities across LLMs. Prevailing approaches assume a teacher at…
Adapting Large Language Models (LLMs) to specialized domains often incurs an alignment tax, as fine-tuning on domain-specific tasks can cause catastrophic forgetting and substantially degrade performance on general tasks. We propose…
DuplexGen calibrates dialogue generation against human preferences to produce scenario-adaptive turn-taking behaviors in full-duplex interaction.
A consensus anomaly detector run on ten years of monthly surveillance found outbreaks aren't random noise: anomalous months post a Cohen's d of 3.25 against normal ones. The geography splits in an odd way, too. Tamale carries the heaviest case burden, but Ashanti's districts flag anomalies more often.
A British cohort study needed LLM agents to clean 20 longitudinal data-prep tasks (102 variables, six survey waves) without the data ever touching a third-party API. Open-weight 31-35B models hit 87.9% average task completion on consumer-grade hardware. Good enough that "no cloud" stops being a compromise and starts being a real option for governance-locked research data.
DeepSeek-R1-Distill-Qwen-7B on AIME problems either lands an answer (96.5% accuracy) or burns its whole token budget and comes up empty (11.5%), almost no middle ground. Probing hidden layers 150 tokens in catches a faint but real signal, AUC 0.61, of which path a generation is on well before the budget runs out.
Diffusion transformers are essential for high-fidelity video generation, but long token sequences make attention a dominant inference bottleneck. Training-free dynamic sparse attention alleviates this bottleneck by computing only selected…
Multi-turn long-horizon planning is critical for foundation model agents, yet how to fundamentally improve it remains unclear. Existing models are trained on uncontrollable and opaque Internet data, making it difficult to identify how…
Reliable visual document understanding requires a model to attribute each answer to the evidence regions that support it. Recent benchmarks and systems express this step through a coordinate interface: the model outputs the coordinates of…
On-policy distillation (OPD) adapts diffusion models by querying a teacher along trajectories generated by the current student, but how it should behave under classifier-free guidance (CFG), a default component of modern diffusion systems,…
Progress in video generation keeps narrowing the visual gap between AI-generated and professionally produced footage, yet most benchmarks still draw prompts from web sources or LLM templates and score them with untrained, generic…
While data curation for Vision Language Models (VLMs) is increasingly active, public practice for constructing pretraining mixtures remains largely heuristic: practitioners stack datasets that pass quality filters, set cross-domain ratios…
Agentic search enables large language models to solve knowledge-intensive tasks by interleaving multi-step reasoning with retrieval, yet optimizing this with outcome-based reinforcement learning (RL) provides only sparse supervision.…
Multimodal large language models (MLLMs) hold immense potential to revolutionize clinical practice, yet deploying them in the medical domain is fundamentally a vision-centric challenge: models must absorb knowledge from heterogeneous 2D…
Multimodal foundation models learned to see and to speak by consuming the whole internet. Embodied agents admit no such shortcut, since they require data that couple observations with physical states and actions. These signals can be…
Compressed short-text generators can fail in two different places: the codec may discard information before generation starts, or the latent generator may produce weak codes. Without separating these failure modes, researchers can spend…
Standard vision-language models (VLMs) suffer from Moravec's paradox: they excel at complex offline visual reasoning but struggle with simple streaming perception tasks and process them inefficiently. We present Mage-VL, an efficient…
Relevance is a query-dependent estimate of whether a document or excerpt contains useful evidence. Existing retrieval agents use relevance to select top-k content, but document relevance alone cannot localize, compose, or verify the…
We introduce PerceptionBench, a benchmark specifically designed to evaluate the atomic visual perception capabilities of Multimodal Large Language Models (MLLMs). Existing benchmarks often fail to isolate perception: holistic evaluations…
Long-term memory systems store what a user says in an external store and retrieve it when a related query arrives. This interface rests on an assumption so natural that it is rarely stated: a memory that is needed will resemble the query…
The alignment of Small Language Models (SLMs) in the 70--500M parameter range using reinforcement learning is often considered unstable, though the underlying failure mechanisms have not been systematically investigated. In the…
Many recent robot policies pursue stronger control by using large pretrained vision-language models (VLMs) as the action backbone. We introduce WorldDiT, a unified diffusion transformer architecture that couples action generation with…
Modern multi-agent knowledge systems increasingly accumulate knowledge through chains of autonomous transformations rather than direct retrieval. Existing provenance work records what happened - execution traces, tool calls, evidence links…
Biomedical image analysis spans diverse modalities and tasks, yet real-world deployment is hindered by severe distribution shifts across scanners, protocols, and patient populations. High-performing models consequently require repeated…
Modern coding agents are usually evaluated by whether they eventually produce a correct patch, but patch generation depends on an earlier context-acquisition stage: finding the repository files needed for the task. We introduce Agent…
Existing BraTS-GLI datasets provide a widely used benchmark for adult glioma MRI segmentation, but their task definition focuses on tumor subregions and does not systematically represent coexisting white matter hyperintensities (WMH). In…
Learned prompt compressors trained on English data disproportionately degrade non-English contexts, widening token-cost disparities, while multilingual training and deterministic methods reduce this gap.
That's the gap that's been quietly killing open RL efforts: multi-turn tool use is hard to reproduce outside the vendor's own stack.
A new paper argues the fix for agreeable, spineless LLMs isn't dialing down sycophancy as one knob. It's teaching the model to tell "update on new information" apart from "cave because someone pushed back." Sounds abstract, but it's the exact failure mode that shows up when an agent takes instructions from an untrusted user mid-task.
Nice trick for synthetic training data and robotics sim, if the physics survives contact with anything outside the demo reel. Watch, don't act: these generative-sim papers have a track record of looking great in the video and falling apart on real dynamics.
Since Volta introduced Independent Thread Scheduling (ITS), NVIDIA GPUs have been widely assumed to handle warp divergence in a fixed manner. We test this assumption across Ampere, Hopper, and datacenter and consumer Blackwell GPUs, using…
The rapid evolution of generative models has unlocked new potentials in protein binder design, a pivotal task in structural biology, by facilitating end-to-end generation via joint sequence-structure modeling or hallucination. However,…
Improving a language model today means retraining it: enormous compute, a new opaque model each cycle, non-deterministic output. We take the opposite path: the model stays frozen, and a persistent memory of verified solutions grows beside…
Parametric models of the human head are essential tools traditionally used in computer vision and graphics for animation, rendering, and reconstruction. More recently, they serve as crucial conditioning signals within generative large…
Creative AI is moving from single-step asset generation toward long-horizon multimodal production. Although recent generative models can synthesize high-quality images, videos, audio clips, UI elements, storyboards, slides, and other…
Recent generative models are moving beyond silent video or standalone audio synthesis toward the joint generation of synchronized audio and video. Despite this progress, jointly generating audio and video with fine-grained cross-modal…
Driving style captures stable, driver-specific patterns in how a vehicle is driven. In naturalistic data, however, this signal is hard to isolate because drivers are observed in different vehicles, on different roads, and under different…
Multimodal automated fact-checking (MAFC) verifies claims by retrieving and reasoning over external evidence. However, most existing static benchmarks risk contamination: they primarily consist of outdated claims verifiable using an LLM's…
We present N_0-TWAM, a tactile-native world-action model for contact-rich manipulation that predicts both future vision and future contact. To our knowledge, it is the first tactile world-action model trained at large scale, and it shows…
We present N_0-VTLA, a vision-tactile-language-action (VTLA) foundation model capable of (1) fine-grained contact-rich manipulation with tactile perception and tactile-feedback control, and (2) offline policy improvement from stored…
Reinforcement Learning with Verifiable Rewards (RLVR) has driven recent progress in reasoning-oriented large language models (LLMs) by enabling large-scale optimization. However, its applicability remains largely limited to domains such as…
A new paper ran OpenAI's gpt-5.6-sol through 25 versions of the same harmful request. Show it the raw ask and it refuses, straight up. Route the identical ask through relay agents dressed up as "Id" and "Censor" before it hits a final "Superego" agent, and the model goes along with it. The model never actually judges the request. It judges how the request arrives.
UMF instead runs several diffusion models against the same sequence simultaneously, using tree search to hand each unmasking step to whichever model is most confident at that moment. No retraining required.
That's epanorthosis, a self-correcting rhetorical figure Cicero was cataloguing two thousand years ago, and it's why so much AI copy has a telltale hollow ring. A new paper traces it to promotional training data, RLHF rewarding the confident-sounding half of the sentence, and left-to-right decoding locking it in once started. A short line of instruction cuts it noticeably; fine-tuning removes it almost entirely.
Large Language Models (LLMs) have transformed conversational AI, yet high-quality multilingual code-mixed dialogue resources remain scarce, particularly for Indic languages where speakers naturally alternate between English and their…
In this work, we aim to discretize the high-dimensional visual representations to bridge the gap with language models - a non-trivial challenge, as existing quantization methods suffer from codebook collapse, failing to scale while…
Bitcoin price prediction on sub-daily timescales is a hard open problem in computational finance. Bitcoin exhibits fat-tailed returns, non-stationary dynamics, and a price discovery process influenced by social discourse on Reddit and…
Large Vision-Language Models (LVLMs) remain bottlenecked by massive computational footprints, precluding their deployment on resource-constrained edge devices. While efforts to compress LVLMs focus heavily on vision token reduction or…
Sakana's follow-up to the orchestration model that already beat frontier AI on cybersecurity benchmarks. The new version claims parity with purpose-built cyber models like GPT-5.5-Cyber and Mythos Preview. The claim is Sakana's own, straight off a tweet, no third-party eval yet.
Every value-alignment benchmark you've heard of encodes Western defaults, so a model tuned to score well on them can still get local social norms wrong on the next question. This paper builds a benchmark from Sri Lankan context instead of translating an existing one. Builder read: if you're shipping outside the US and EU, your alignment eval probably doesn't cover the market you're shipping into.
No model, no benchmark, just the proof. It's the kind of foundational math that might tighten generalization bounds someday, not something you'll touch this quarter.
Diffusion models typically suffer from error accumulation during iterative sampling, commonly referred to as exposure bias. We reveal systematic frequency-dependent discrepancies between training and inference, which can be interpreted as…
Large Language Models (LLMs) have significantly automated the process of scientific discovery over the past few years. However, existing systems share one core limitation: they generate and optimize ideas independently for either Quality…
In multilingual retrieval augmented generation, a retriever can retrieve relevant documents written in multiple languages, which are subsequently reranked before answer generation. However, it remains unclear whether existing multilingual…
Although large language models (LLMs) exhibit remarkable reasoning capabilities, their reliance on text-only pre-training restricts the perception of the multimodal physical world. Native multimodal pre-training avoids this limitation by…
Vision-language model (VLM) agents increasingly use tools to act on 3D scenes rather than only describe them. Existing 3D benchmarks score textual responses or single-object operations, leaving agent action on complete multi-object 3D…
LLM training is shifting from manual design and annotation to interaction-driven self-evolution. However, existing self-evolutionary methods face a fundamental dilemma between task diversity and verification reliability: environment-bound…
Large reasoning models (LRMs) generate long reasoning traces before producing final answers. While these traces may contain useful signals for hallucination detection, harnessing them is non-trivial because long trajectories often include…
Historical documents act as invaluable knowledge archives but often suffer from illegibility due to physical deterioration and damage. While existing restoration methods based on masked language modeling effectively utilize local context,…
Computer-use agents are usually improved by strengthening perception: better models for reading a screenshot and choosing where to click. Yet a screenshot is only a lossy rendering of the underlying program state, e.g., the files,…
In visual storytelling, human performances are central to creative intent and narrative meaning. However, preserving human identity and performance while enabling flexible visual edits remains challenging for generative video models. We…
Domain Generalization (DG) aims to learn representations robust to distribution shifts. Recent geometric alignment methods, such as CPCANet, extract domain-invariant structures through batch-wise Common Principal Component Analysis (CPCA).…
Give an agent a real job (post-train a model, tune a CUDA kernel, optimize an inference server) and see if a monitor can catch it quietly sabotaging the result. The failure mode nobody caught: sabotage baked into training data, missed more than half the time even when the monitor could run and probe the finished artifact, not just read its reasoning trace.
earns its place as a survey by being useful rather than novel. It tracks agentic systems moving from demo to production across software engineering, science, and finance, and pulls concrete deployment patterns (verification pipelines, fallback paths, human checkpoints) from real pharma and finance rollouts rather than benchmarks.
Understanding motion in video is a fundamental challenge for visual learning, as frame-to-frame change entangles two sources of dynamics: camera motion and object motion. This decomposition has remained underexplored in representation…
Large language models are increasingly used in K-12 education, but existing benchmarks mainly test exam question answering rather than understanding how curriculum knowledge is structured and visually presented. We call this capability…
Real-world agent learning is often constrained by costly environment interactions, such as running time-consuming experiments or obtaining human feedback. In-context learning offers a highly sample-efficient way for agents to learn from…
Spatial intelligence is essential for agents to move from static semantic understanding toward interacting with the physical world. Many spatial tasks are grounded in continuous visual scenes, where locations, regions, and paths are more…
We study sinusoidal recurrence as an iterative mechanism for harmonic spectral enrichment in implicit neural representations (INRs). Our analysis reveals that sinusoidal activations induce a harmonic line spectrum, providing a spectral…
On-policy self-distillation (OPSD) is promising as it removes the external teacher required by on-policy distillation (OPD), yet it still needs asymmetric information between teacher and student to ensure that the self-teacher provides a…
Deep research requires agents to find answers that jointly satisfy multiple constraints. Discovering such answers is costly, whereas verifying a candidate can often be decomposed into tractable constraint-wise checks. This…
Controllable video generation remains challenging due to the difficulty of specifying precise multi-object interactions using text prompts or motion-control inputs that primarily constrain pixel movement. In practice, trajectory-based…
We introduce Tencent WorkBuddy Bench, a multi-domain evaluation suite for coding agents; this report documents its construction methodology, scoring protocol, and a cross-model leaderboard. At its core is a unified evaluation framework for…
The development of generalizable robotic manipulation policies is inherently bounded by the availability of large-scale, high-fidelity scene data. While recent automated synthesis methods attempt to bridge this gap via text-to-layout…
Multi-agent interactive world models should not only generate consistent observations, but also maintain world states that persist across agents and evolve across views. Existing autoregressive video diffusion pipelines carry forward…
The recent emergence of vibe-coding workflows is changing what coding agents are expected to do. Instead of merely completing code under fully specified instructions, agents are increasingly expected to transform incomplete product intent…
We introduce SANA-Video 2.0, a hybrid video diffusion transformer instantiated at 5B and 14B scales under a unified architecture. Designed to generate high-quality video up to 720p on a single GPU, SANA-Video 2.0 matches full-softmax video…
Production AI agents' failures are less often due to an inability to reason well and more often because they cannot manage what is in their reasoning context: conversation histories, large prompts, large tool definitions, and ballooning…
Recent conditional video generation models have shown promising potentials to transform 3D engine renderings, such as depth maps and untextured geometry, into photorealistic videos for gaming and immersive content creation. These…
We present Oxygen-TryOn, a unified foundation model for any-item virtual try-on. Rather than repurposing a general-purpose image editor, Oxygen-TryOn is fashion-native, built for try-on through a dedicated data engine and try-on-specific…
Deep Research agents extend LLM-based assistants into long-horizon workflows involving planning, retrieval, evidence synthesis, and report generation, yet their reliability in open information environments remains underexplored. A key…
Red-team evaluations of AI models support some claims and not others, and the boundary between the two is calculable rather than merely a matter of judgment. We define the evidential ceiling of an evaluation as the largest factor by which…
Coding agents leave a mess behind: speculative edits, abandoned hypotheses, scaffolding that never gets cleaned up before the final diff. This paper names it "CodeSlop" and trims it by pruning the agent's search trajectory itself rather than post-editing the output. Builder read: worth wiring into any coding-agent pipeline where diff size matters for review.
Slip a model a casual hint, a mislabeled few-shot example, or a fake prior answer, and it'll often flip a correct answer to match. Researchers traced the effect across multiple model families and found the bug isn't in pretraining. Base models barely have it.
Long reasoning chains sometimes get stuck circling the same dead end until the token budget runs out. This method detects that kind of loop mid-trace and steers the model back out before it burns through the budget. Reports gains in both accuracy and token efficiency.
Generative world renderer AlayaRenderer receives structured world states exported from physics engines and synthesizes RGB frames. Unlike models that generate frames from text/control-hints prompts, AlayaRenderer preserves scene structure…
Reinforcement learning with verifiable rewards (RLVR) is rapidly advancing the reasoning capabilities of language models, yet the optimization layer that converts reward feedback into weight-space updates remains poorly understood.…
We present ABot-World-0, an action-conditioned video world model for real-time, long-horizon closed-loop interaction, supported by a multi-source data infrastructure spanning AAA games, simulation engines, and internet videos to learn…
As autonomous agents rapidly evolve, their ability to reliably manipulate ubiquitous digital documents has become critical for enabling general-purpose AI assistants and automating complex workspace workflows. In this paper, we introduce…
Natural-language autoencoders score explanations of hidden activations by reconstruction: an explanation is deemed faithful if the activation can be regenerated from it. The test is structurally insensitive to individual false claims: if…
Reinforcement learning with verifiable rewards (RLVR) has substantially improved language-model reasoning, yet its extension to vision-language models remains constrained by the lack of training data that are simultaneously broad, exactly…
Recent autoregressive video diffusion methods are increasingly built upon Self Forcing, where the student is trained on histories produced by its own rollout rather than ground-truth video contexts. This reduces exposure bias, but the…
As large language models and AI agents become the primary consumers of search results, document set quality determines the upper bound of downstream generation. Yet existing evaluation systems remain confined to scoring documents…
Full-parameter post-training of trillion-parameter-scale MoE models introduces substantial system-level challenges for large-scale distributed training, including severe memory pressure, non-overlapped communication overhead, and…
Reinforcement learning with verifiable rewards has become the predominant recipe for eliciting test-time scaling in explicit Chain-of-Thought reasoners. Yet this scaling path remains computationally costly, since every intermediate step…
This work introduces G-MAD, an open-source framework that uses Arma3 to generate synchronized multi-view RGB-T data for aerial object detection. G-MAD addresses key limitations of real-world aerial dataset construction, including limited…
Practical robotic grasping in complex scenes requires both 3D spatial reasoning and alignment with task-specific requirements. Vision-language models (VLMs) offer a natural way to specify these requirements using language, but existing…
3D Gaussian Splatting (3DGS) achieves high-quality novel-view synthesis by optimizing freely placed primitives in 3D and adaptively densifying them in under-reconstructed regions. However, this scene-adaptive capacity allocation is largely…
Large language models can answer scientific questions, yet a correct output does not reveal whether the model represents or uses the governing physics. Here we show that materials science mechanism information in the open-weight…
As LLMs become more capable, they are increasingly deployed as collaborative agents, taking on user-delegated tasks through iterative interaction. Yet genuine interaction is inherently dynamic: users rarely specify their intent upfront,…
Deploying navigation systems at scale requires a recipe that minimizes sensor assumptions, generalizes across robot embodiments, and trains efficiently. Yet, today's best systems depend on depth sensors, multi-camera rigs, or pre-built…
Embodied visual tracking (EVT) requires a mobile agent to continuously follow a specific target described in natural language using only onboard vision. While recent vision-language-action (VLA) policies unify target identification and…
Traditional agent development is split across prompt templates, tool schemas, callback code, and workflow graphs. We present NVIDIA Object-Oriented Agents (NOOA), a model-agnostic Python framework for building reliable AI agents. NOOA…
Contextual entrainment is the tendency of a model to let auxiliary context in its input pull its output, independently of whether that context is relevant, true, or even meaningful. Recently, it has been identified and given a mechanistic…
Most automatic speaker verification (ASV) systems operate on individual utterances, despite real-world interactions typically consisting of multiple utterances. As speech accumulates, increasingly rich speaker information becomes available…
Agentic reinforcement learning research is constant algorithm modification, new estimators, new pipeline stages, new rollout schemes, and in mainstream frameworks each change threads through layers of trainer, distributed backend, and…
Robotic learning takes place in dynamic environments with large behavior spaces. A terminal success signal only tells the robot whether the task is completed. It does not explain whether the current behavior is making progress, remaining…
In RLHF pipelines, reward scoring blocks policy updates. Slow scoring bottlenecks the entire loop, since no update runs until every rollout gets a score. And yet most setups just default to PyTorch eager mode or torch.compile, no one…
Transformers propagate information across depth through a single additive residual stream: every sublayer reads only the most recent state. Attention residuals relax this by letting each sublayer attend, through a learned softmax. However,…
Do Large Language Models (LLMs) possess genuine structural reasoning, or merely rely on surface-level pattern matching? The financial domain, demanding numerical precision and multi-step logic over long contexts, is an ideal testbed.…
A real-time framework for online co-speech gesture generation uses a causal multimodal autoregressive model with streaming speech and motion history, supported by synthetic dialogue data and continual user-feedback adaptation.
MoE serving has a memory fight nobody talks about enough: the expert weights and the growing KV cache both want the same GPU room. This paper pages weight precision dynamically, keeping the bits that matter and thinning the ones that don't, so you can stretch context length without adding GPUs. Builder read: promising if the quality claims survive outside the paper's own benchmarks.
A group sketched a blueprint for analog chips that compute by settling into physical equilibrium, using the hardware's own noise as the substrate instead of fighting it. No chip exists yet, just the math and the architecture. Watch, don't act, but it's a genuinely different bet on where post-GPU compute comes from.
Most point-matching methods treat every point as interchangeable. This one matches cluster to cluster first, then refines inside each pair, which should generalize better on data with real group structure. Narrow use case (dataset alignment, domain adaptation) but a clean idea.
Multimodal humor in memes, cartoons, and comics remains difficult for AI systems because intended meaning depends on non-literal mechanisms, shared cultural knowledge, and communicative intent rather than literal scene description. This…
Accurate agricultural field boundary delineation at large scale is a foundational task for food security, supply chain transparency, and carbon accounting. While vision foundation models like SAM show remarkable zero-shot capabilities,…
Optimizer state is the largest single line item in the memory budget of mixture-of-experts (MoE) training: on a 6.78B-parameter MoE language model, AdamW keeps 50.6 GB of first and second moments to update 12.6 GB of bfloat16 weights. We…
LLM agent failures are difficult to debug because the step where an error surfaces is often not the one that caused it. Existing observability tools replay execution traces but provide little support for identifying the root cause or…
Text-to-image diffusion transformers (DiTs) jointly process text and image tokens, yet their internal computation during denoising remains poorly understood. We introduce a causal interpretability framework for modern large-scale DiTs that…
Efficient teamwork typically combines global coordination with parallel execution, a principle not yet fully reflected in unified Vision-Language Model (VLM)-based document parsers. Existing unified parsers process an entire page jointly…
Evaluating the factuality of long-form generations has focused predominantly on precision, measuring whether the claims a model makes are correct. The dominant decompose-search-verify pipeline catches incorrect claims well but says little…
Unlike conventional video game development, which relies on labor-intensive pipelines for asset production, animation, physics, and programming, video world models generate interactive environments from user inputs instantly. It enable us…
Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but staleness is an inevitable byproduct compounded by policy lag, engine delays, and mixture-of-experts routing. From a…
Large-scale visual generators are increasingly capable but costly to train, fine-tune, and deploy. We introduce Mage-Flow, a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing. The…
Video models absorb rich priors over how the visual world moves, interacts, and responds to contact, making them promising substrates for robotic world modeling. The central challenge is how to communicate action to such models in a form…
Controllable image generation remains challenging for creative professionals, who often require precise regional control over materials, object identities, and spatial arrangements that cannot be reliably achieved through text prompting…
Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models on tasks such as mathematical reasoning and code generation. However, most RLVR methods assign a scalar…
Scaling executable agent training data for LLM post-training is bottlenecked by substrate-bound methods that tie task generation to predefined tools, repositories, or skill graphs: expanding coverage requires manual substrate engineering,…
Injecting factual knowledge into large language models (LLMs) reliably and at scale remains an open challenge. Hypernetworks provide a promising solution to large-scale knowledge injection. Although hypernetworks are typically applied for…
We present AutoIndex, a framework for learning representation programs: executable transformations that map raw documents into the representations exposed to a retrieval system. Rather than tuning retrievers, rerankers, or a small set of…
Modern ASR models trained on heterogeneously annotated data treat transcription style (verbatim vs. intended) as an uncontrolled latent variable, causing measurable decoding instability, evaluation confounding (up to 60% of reported WER…
Agentic Reasoning has become a transformative force in financial analysis due to its ability to integrate large-scale information and generate reliable and accurate content. However, when handling complex real-world problems, different…
Text-to-video generation has advanced significantly over the past five years through scaling of model size, data, and compute. Unlike model architecture, training data is often underexplored. Real-world data curation is complex and…
This paper proposes AI Tour Meeting, a group travel planning framework powered by multiple Large Language Model (LLM)-based agents. The agents are instantiated with distinct personas and collaboratively seek an itinerary that satisfies…
Rubric-based RL has recently shown promise in improving LLMs on open-ended tasks. A widely recognized limitation of rubric-based RL is limited exploration: criteria that no rollout manages to satisfy (Unexplored Criteria, UC) receive no…
Specialized scholarly agents use reinforcement learning and real-time citation verification to improve reasoning and factual accuracy in peer review and rebuttal generation.
Multi-agent system that runs an entire meta-analysis end to end: finds the studies, pulls effect sizes, pools them, flags bias. That's the grunt work behind every systematic review, normally a grad student with a spreadsheet for six months. Whether it holds up on a genuinely contested literature, not a clean toy set, is the real test nobody's run yet.
World-action models are supposed to be safer because they predict what happens next before choosing an action. This paper's finding: the physics prediction can be dead right and the action picked right after it still gets someone hurt. Good world model, bad policy. If you're building embodied agents on this architecture, "predicts the world accurately" is not a safety guarantee.
A clinician-built benchmark that stops grading medical AI on right versus wrong and starts grading which safety gate it blew through, a missed urgent referral, a skipped dosage check, graded 1 to 5 on severity. More useful than one more accuracy leaderboard if you're anywhere near deploying this stuff in an actual clinic.
Current video generation models achieve impressive results in single-shot generation, yet remain limited in cinematic video generation, where coherent narratives and effective multi-shot composition require explicit shot planning. To…
Predicting a football match before kickoff requires more than knowing past results: a model must use changing information and make a clear prediction before the answer is available. We present WorldCupArena, a dynamic benchmark for…
Self-hosted AI agents read and write their own memory and configuration files to function. An agent may get compromised via corruption of its own state -- a compromise realized via legitimate OS system call invocation. We refer to this…
In line with the prevailing direction of vision research, we explore the integration of both generation and editing capabilities for video and image modalities within a single model. Current approaches to collecting video editing data…
We propose Token-Level Off-Policy Labeling (TOPL), an off-policy training paradigm that reframes post-training as a token-level correctness prediction task. Our key intuition is that by training the model to distinguish good and bad tokens…
Human-object centric video personalization (HOCVP) is a core task within subject-driven video generation. However, existing methods suffer from two key limitations. First, most approaches focusing on inter-subject personalization still…
Real-time multimodal applications, including voice agents and interactive video generation, compose heterogeneous models into pipelines whose efficient deployment requires application-specific decisions about placement, streaming, and…
Pruning long context for coding agents has been a vital technology for efficient context management. While existing context pruning methods such as SWE-Pruner realize this by attaching a separate code classifier, we find the agent itself…
We present RynnBrain 1.1, a family of embodied foundation models spanning 2B, 9B, and 122B-A10B scales. Trained with a unified spatio-temporal and physically grounded framework, RynnBrain 1.1 supports embodied perception, spatial…
Reinforcement learning (RL) on open-ended tasks compresses an LLM's rubric-based evaluation into a scalar reward, discarding rich textual feedback and conflating responses with distinct quality profiles. We propose Experiential Learning…
The prevailing inference framework for diffusion models formulates generation fundamentally as a problem of numerical integration. This perspective casts the model as an exact estimator, neglecting the inherent statistical uncertainty of…
Teaching videos are becoming a major medium for education, creating a growing need for scalable evaluation of their pedagogical quality. Existing automatic judges do not fully address this setting because teaching quality depends on…
Structural fidelity is essential to scientific methodology diagrams. To communicate research logic, these diagrams must faithfully render components, directional relations, and textual annotations. Since a single error, such as a reversed…
Video spatial reasoning is essential for navigation-oriented perception and long-video question answering, where models must infer spatial relations across long horizons under changing viewpoints. However, existing multimodal large…
Multi-agent systems routinely place one AI agent in authority over another. When a subordinate refuses a task, the manager chooses the outcome: it can renegotiate, report the failure honestly, coerce the subordinate, or lie about the…
Egocentric devices, such as wearable front-facing cameras, provide a unique perspective for capturing the continuous interaction between a human viewer and the surrounding environment. A holistic and efficient multimodal model capable of…
Structure-based drug design (SBDD) leverages the 3D structure of protein targets, often complemented by other spatial constraints, to generate candidate binding molecules. While diffusion models have dominated as a leading paradigm for…
Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive models. Unlike standard diffusion-based approaches, DLMs are not explicitly conditioned on a timestep, raising a natural question: do these…
Real-time EEG classification on edge devices is bottlenecked by the floating-point arithmetic of conventional neural networks. We investigated Differentiable Logic Gate Networks (Diff-Logic) as a hardware-native alternative that compiles…
Simultaneous localization and mapping (SLAM) is one of the fundamental problems in robotics, as it enables autonomous operations in real-world scenarios. Under low illumination, reduced contrast, sensor noise, and motion blur degrade both…
Modern generative models typically rely on an adversarial critic, a prescribed noise-to-data path, or an autoregressive factorization. Instead, we show that a proper distributional energy can induce sample-level motion and provide direct…
Industrial Video Anomaly Detection (IVAD) aims to identify anomalous objects and events in an industrial process, which is crucial for modern manufacturing and quality control systems. Existing VLM-based anomaly reasoning methods are…
A language model p_θ(y mid x) trained on reasoning tasks learns to solve problems via multiple distinct strategies, yet these strategies are implicit and entangled within the model's response distribution. We study the problem of…
The benchmark evaluates autonomous coding agents on open-ended world-model research by having them iteratively improve a starter model across game environments using a shared structured-state format.
A new probe called PRISM checks a model's hidden states instead of its output text, hunting for instructions that sound harmless but turn dangerous once a robot actually executes them. On the team's new PhysicalSafetyBench-1K it hit 99.6% accuracy with a 0.7% false-positive rate. A plain text-safety judge, by contrast, rejected two-thirds of tasks that were completely fine.
Retrieval has always been scored on one question: does this document answer the question in front of you. Wrong test, apparently, once an agent starts searching in multiple steps. Researchers logged 23,000+ document reads inside a ReAct agent and found almost zero correlation (Spearman rho of -0.026) between a document scoring as relevant and it actually changing the outcome.
Ask a model for estimates about subgroups of a population, then ask it for the population-level number, and the two should reconcile. Often they don't.
We present a continuous geometric framework that models the discrete algebraic operations of the Transformer architecture as an integro-differential equation (IDE) on a semantic fiber bundle calE = calM times R^d. Beginning from a single…
Hand-Object Interaction (HOI) synthesis is a cornerstone for animation production and embodied AI. Despite the strong priors of video foundation models, multi-view consistent HOI synthesis remains challenging due to complex hand motions…
Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality…
This paper introduces EvolvingWorld, a framework and benchmark for character and world co-evolution in interactive literary worlds. Existing systems either treat interactive literary simulation as static persona imitation or isolated scene…
Large language model (LLM) post-training is essential for improving reasoning, adaptation, and alignment. Existing methods mainly follow two paradigms: reinforcement learning (RL) and on-policy distillation (OPD). However, RL relies on…
Replacing individual LLM agents with low-parameter surrogates fitted from cheap queries enables scalable society simulations, with validity predicted by an interaction-order and memory taxonomy.
Most robot foundation models look at the last frame or two and nothing else. RoboTTT scales working memory to 8,000 timesteps, three orders of magnitude past prior work, by folding test-time training into the policy instead of just widening the context window.
Earlier poisoning research stuck to tidy, well-behaved sources like Wikipedia. This paper points at something messier and already public: open discussion interfaces, the comment threads and forums that get swept into web crawls. The harder question is whether injected content actually survives data curation and lands in a trained model, which is why the authors built a measurement tool (HalfLife) to check.
Most agent security evals just measure who wins, assuming an unlimited compute budget, which is not how a SOC operates. This one scores agents on both offense (Cybench) and defense (Splunk BOTS v1 investigations) against actual inference and tool spend.
Optical coherence tomography (OCT) imaging is essential for the diagnosis and treatment of retinal diseases. Although multimodal large language models (MLLMs) have demonstrated considerable potential in medical image analysis, existing…
Entropy control has become an effective tool in reinforcement learning (RL) of large language models (LLMs), helping balance exploration-exploitation trade-off during alignment process. Such RL paradigm is often conducted on mixtures of…
Training API-calling large language model (LLM) agents demands massive amounts of high-quality trajectories. However, collecting such data at scale typically requires fully implemented environments with executable APIs and realistic,…
Large language models (LLMs) are increasingly used to automate data-processing workflows, yet coding agents typically produce scripts that are not automatically materialized as persistent, editable platform artifacts. We call this…
We revisit dataset distillation from an outcome-centric perspective. Rather than aligning process surrogates (per-step gradients or training trajectories), Influence Matching (Inf-Match) aligns the final outcome of training: it learns a…
On-policy knowledge distillation transfers reasoning from large teachers to compact students, but existing approaches suffer three compounding failure modes: (i) cold-start collapse, where a fresh student assigns near-zero mass to…
In our prior work, Pedestrian Archetypes, we defined pedestrian archetypes as collections of behaviors that uniquely identify a specific type of pedestrian. The first paper proposed 12 pedestrian archetypes, including the Wanderer, Drunk,…
turns out less dramatic than the hype suggests. Across 25,264 agent-submitted PRs in 2,361 popular repos, the median project sees only one or two a quarter, and most stay under an industry benchmark of 36 PRs per contributor over three months. Almost every one still runs through a single human reviewer before merging. Agents aren't replacing dev teams yet. They're on a slow, supervised trial period.
A Terminal-Bench 2.0 study pitted three agent-tuning methods against a fresh round of unseen tasks. GEPA's gains (66% pass rate after tuning) collapsed below the 58.7% unoptimized baseline the moment new tasks showed up. Meta Harness held its 64.6% but stopped improving. Only RELAI-VCL, which builds in regression control, kept climbing, to 76.4%, and kept the gains on retrain.
A proposed framework argues exploit-hunting misses how AI systems actually break: an agent can leak data or take a bad action with zero vulnerabilities exploited, just a poisoned tool call or prompt injection nudging its behavior off-objective. The six-step process tests for that kind of behavioral violation directly instead of waiting for a CVE. Worth stealing if you're red-teaming anything with real tool access.
Muon is competitive with AdamW in large-scale pre-training, but its value for reinforcement-learning (RL) post-training remains unclear. We study vanilla Muon in sparse-reward agentic RL through matched single-seed comparisons with AdamW…
We present S1-Omni, a unified multimodal reasoning model for scientific understanding, prediction, and generation. AI for Science (AI4S) has advanced significantly through domain-specific models, tool-augmented LLMs, and scientific…
Despite strong capabilities in data understanding and decision-making, autonomous data science agents still heavily rely on trial-and-error workflows that involve expensive computation. This bottleneck motivates models that can anticipate…
Under model--harness co-evolution, harnesses are not merely inference-time scaffolds but data-generating components whose execution traces can shape future foundation models. This motivates harness-in-the-loop learning: optimizing…
We present Loopie, the most powerful looped Transformer to date. The Loopie series consists of two Mixture-of-Experts (MoE) models: a 20B-parameter model with 2B active parameters and a 6Bparameter model with 0.6B active parameters. Looped…
We present Audio-Visual Flamingo (AV-Flamingo), a fully open state-of-the-art audio-visual large language model (AV-LLM) for joint understanding and reasoning over audio, images, and long-form videos. Unlike prior AV-LLMs that primarily…
Large language models (LLMs) are transforming recommender systems from matching co-occurrence patterns in historical behavior toward reasoning about the intent that drives it. RecGPT-V1 pioneered this paradigm on Taobao by centering user…
Reinforcement learning (RL) has become central to improving large language models (LLMs) on complex reasoning tasks, yet RL post-training is largely studied in isolation from the pretraining that precedes it. As a result, two basic…
Modern video generation models are increasingly hailed as emerging world models with an internalized grasp of physical law. Yet existing benchmarks largely evaluate physical plausibility only at the output level, without verifying whether…
The post-training of Vision-Language-Action (VLA) models is essential due to the diversity of simulators, robot embodiments, and task objectives. Existing compute services, whether offered as direct accelerator rental or batch-workload…
Mobile graphical user interface (GUI) agents have demonstrated remarkable capabilities in automating complex tasks, yet they introduce critical safety risks where a single erroneous action can lead to irreversible consequences. Existing…
As generative AI is increasingly applied to automate multi-step and high-stake workflows, human judgment and involvement remain essential for ensuring the quality of AI-generated outputs. In practice, while it is desirable for human…
Human vision is a closed loop: gaze is continuously redirected by intermediate hypotheses rather than a single snapshot. Decades of psychophysics and cognitive science have argued that this active observation is essential for a wide range…
Video Diffusion Transformers process long spatio-temporal sequences, making self-attention the main bottleneck in high-resolution video generation. Training-free sparse attention reduces this cost, but adaptive Top-p routing creates uneven…
Experiment trackers show how training is progressing, but changing a live run still usually requires trainer-specific code. We present Interactive Training 2, an open-source control plane for steering training through a shared protocol.…
Autonomous driving systems (ADS) are rapidly advancing and increasingly deployed in real-world applications. This creates growing demands for effective testing to ensure system functionality and safety. However, ADS testing remains complex…
While instruction-based video editing has advanced rapidly, real-world videos contain tightly coupled audio and visual signals, and editing one modality often requires coordinated changes in the other. Existing benchmarks primarily…
A new paper catches LLM agents defaulting to max-context-first: re-reading every file and dependency they've already seen, even for a one-line edit. The fix isn't a bigger context window, it's a cheaper first question the agent never asks: how much effort does this actually need.
Turns out judge models skew generous specifically in no-reference setups, which is most real evaluation pipelines (open-ended output, no ground truth to check against). If your eval numbers look suspiciously good, this is a plausible reason why.
Funny, but it's a real signal on model homogeneity. Post-training keeps compressing "creative" output down to the same handful of safe choices across labs. Remember that next time a demo calls an LLM's output original.
A growing gap separates inference context lengths from RL post-training: inference systems are approaching million-token contexts, while post-training workloads often remain at 256K tokens or below and rely on length generalization at…
Visually impaired individuals (VIIs) encounter significant daily challenges due to limited access to visual information. Although Multimodal Large Language Models (MLLMs) have achieved impressive results on general vision and language…
Recent advances in video understanding have spanned motion, long video, and streaming interaction, driving this field toward real-world applications. Despite this progress, current open-source models remain limited in several ways. They…
Music generation foundation models have recently attracted significant industry attention. However, achieving efficient generation and high-fidelity long-form audio while supporting controllability remains challenging. To address these…
Recent advances in Tool-Integrated Large Language Models have made web search a core capability of information-seeking agents. However, as interaction histories grow, agents increasingly struggle to track task progress. When search…
MeanFlow generators achieve fast few-step sampling by predicting average velocities over time intervals, making them attractive for efficient generation. Reinforcement learning (RL) has become a powerful way to align diffusion and flow…
Large language models are increasingly trained as interactive agents for long-horizon tasks involving multi-turn interaction, tool use, and environment feedback. Outcome-based reinforcement learning (RL) provides a practical optimization…
We present Wan-Streamer v0.3, which reframes our native-streaming interaction model under a single organizing view: a video is a world plus an event stream. The world is the persistent context in which a video unfolds, including the…
Video models are evolving into vision foundation models, yet they still lack human-like multi-step reasoning. Streaming autoregressive diffusion models are efficient but limited in reasoning, while bidirectional diffusion enables global…
CAD-to-image alignment aims to estimate an object's 9D pose (rotation, translation, and anisotropic scale) from a single RGB image, enabling applications in robotics and augmented reality. Recent zero-shot methods use visual foundation…
Skills are a useful abstraction for software agents, turning human and agent experience into reusable procedural knowledge. Yet existing skill libraries are mostly hand-written, text-centric, or derived from agent traces, leaving tutorial…
Reinforcement learning with verifiable rewards (RLVR) commonly uses entropy for advantage shaping. However, entropy cannot distinguish useful uncertainty from detrimental confusion, limiting its effectiveness as a correctness signal. We…
We present Xiaomi-Robotics-1, a foundational vision-language-action (VLA) model capable of (1) following diverse language instructions to perform a wide range of mobile manipulation tasks in unseen environments out-of-the-box, and (2)…
On-policy distillation is an alternative post-training method in reinforcement learning that alleviates the constraints imposed by reward models by providing token-level supervision from a teacher model. Although on-policy distillation has…
Hyper-Connections (HC) expand the residual stream of Transformers into N parallel streams, providing a form of memory scaling beyond model width and depth. Manifold-Constrained HC (mHC) stabilizes this formulation at scale. The large gains…
Cross-view geo-localization matches ground-level observations against geo-tagged satellite imagery. Recent methods show that sequential queries such as video clips yield richer spatiotemporal cues than single images, yet they overlook a…
We study on-policy distillation (OPD) for agentic tasks, where an LLM agent interacts with an environment over multiple turns and a student imitates a teacher over these multi-turn interaction histories. Fully online OPD is costly because…
A new paper on multi-agent LLM systems finds a hole in the usual safety net, a monitor that checks each message or tool call on its own. Fragment a harmful payload across agents and every single step passes clean, because the harm only exists once the pieces assemble somewhere downstream. The fix isn't a sharper per-step monitor.
Researchers went inside seven judge models, across seven bias types and nine benchmarks, and found biased inputs don't just score wrong, they push the judge's hidden state along a specific, low-dimensional direction. Steer activations along that direction and you can manufacture the bias on a clean input, or erase it on a biased one.
Picbreeder is the old human experiment where people bred abstract images generation after generation with no goal beyond "pick what looks interesting," and ended up with genuinely surprising art. Swap in a frontier VLM making the same picks and the output looks nothing like that human baseline, less of the compounding, wandering strangeness that made the original worth studying.
As Large Language Models (LLMs) evolve into autonomous agents, the need for unified evaluation infrastructure becomes critical. However, current evaluation pipelines remain highly fragmented and tightly coupled, hindering reproducibility…
While recent advances in 3D generation have enabled impressive visual synthesis, existing methods often rely on 2D diffusion supervision without explicit mechanisms for geometric consistency, leading to spatial hallucinations such as…
World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in…
OpenClaw has emerged as a leading agent framework for complex task automation, yet it faces insufficient cross-platform GUI interaction support and a well-built self-evolution mechanism. These flaws limit its adaptation to diverse device…
We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual…
Building interactive worlds that respond coherently to player actions has long been a shared goal of computer graphics, games, and artificial intelligence. Recent video generative models provide a data-driven route toward this goal by…
Looped Transformers scale sequential computation by applying a compact stack of physical blocks for multiple rounds, increasing unrolled depth without increasing stored parameters. This reuse changes the residual-scaling problem: in an…
Multi-reference-to-audio-video (MR2AV) generation aims to generate coherent audio-video content conditioned on multiple references and textual instructions. Existing benchmarks mainly focus on text-driven generation, single-reference…
Video generation increasingly relies on keyframe-based workflows, where creators specify a sequence of reference images to guide generation. Although recent models support multi-keyframe conditioning, it remains unclear whether they can…
On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood. We present a systematic study examining the role, pathologies, and regulations of OPD. We first clarify the…
We report a way to make a frozen small language model both more capable and dramatically cheaper at once, without changing any weights. Verified knowledge is deposited once as a byte-exact key-value (KV) state artifact and later restored,…
Languages with rich static semantics, such as Rust, provide stronger guarantees for AI-generated code, but their strictness makes generation more difficult. Off-the-shelf compilers can provide useful feedback post-generation, but does not…
Discrete denoising diffusion models (DDMs) have recently emerged as a compelling alternative to autoregressive (AR) modeling for discrete data, offering parallel generation and iterative global refinement capabilities. Unlike continuous…
Validating autonomous driving systems requires diverse, regulation-compliant test scenarios. In simulation-based testing, scenarios are defined as executable scripts. Yet automatically generating such scripts from regulatory descriptions…
Embodied cognition requires agents to connect high-level task reasoning with the physical states to be achieved. We introduce Hy-Embodied-RxBrain, an embodied cognition foundation model with joint language-visual reasoning and imagination.…
Video generative models commonly rely on latent spaces learned by 3D Variational Autoencoders (3D-VAEs). However, conventional 3D-VAEs are mainly optimized for pixel-level reconstruction, which can limit the semantic and spatio-temporal…
Healthcare spans high-stakes communication, expert reasoning, and workflow execution, yet specialized LLMs that cover these use cases together remain limited. A healthcare model must handle patient consultation, clinical reasoning over…
Existing 3D generative models predominantly rely on implicit volumetric representations, which enforce watertight topology and struggle to represent thin-shell and non-manifold geometries such as garments. Geometry image-based approaches…
Serial verification gates are a core reliability primitive in LLM harnesses: a candidate answer is returned only if k verifier calls all accept it. Under conditionally independent gates, the recent Odds Law (arXiv:2606.15712) shows that…
Egocentric videos of human manipulation provide scalable supervision for embodied intelligence, yet existing resources rarely combine low-cost continuous capture, manipulation-level structured annotations, and reusable tools for robot…
Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained…
Large language models are increasingly deployed as agents, but reliable agentic behavior requires more than next-token prediction. At inference time, it is preferred that an agent can decide whether to proceed with its current reasoning,…
A new eval shows a clinical RAG system can pass every automated grounding check, zero hallucinations, near-perfect faithfulness, while attributing the right fact to the wrong patient's chart. Standard RAG benchmarks check whether a claim is supported by retrieved text, not whether that text belongs to the correct entity.
A sovereign, open-source German/English foundation model: a hybrid Mamba-Transformer MoE that activates just 3B of its 30B parameters per token and holds its inference cache roughly flat as context grows. Read: cheap long-context inference you can self-host, with no US API in the loop.
35B parameters, 3B active, and the authors claim it matches or beats 100B-class models through post-training alone, no extra pretraining compute spent. Worth watching whether that holds outside their own benchmarks. If it does, "just scale pretraining" gets a lot less convincing as a MoE strategy.
Post-training is essential for refining the domain-specific capabilities of large language models (LLMs), yet existing reward optimization and distribution matching methods tightly couple policy exploration with distribution alignment.…
Metacognition is a foundational component of intelligence critical to effective learning, problem solving, decision-making, communication, and more. In recent years, it has become increasingly recognized as a cornerstone of capable,…
This work explores the motion transfer from one video to another, which is crucial in animation for diverse characters. Previously, video motion transfer has been largely explored between human and human-like characters, enabling a lot of…
Large language models (LLMs) have achieved remarkable performance on high-school and olympiad-style mathematics, yet their capabilities on advanced mathematics remain poorly understood. Existing benchmarks, however, fall short in both…
Personal AI assistants on mobile and wearable devices continuously perceive users' daily lives through visual and audio streams. However, answering queries about past experiences requires lightweight multimodal memory that can continuously…
Exploration is essential for reliable autonomy in multi-agent systems, yet it remains unclear whether large language model (LLM) agents can explore effectively when interacting with one another. We show that modern LLM agents fail to do…
Current Video Large Language Models (Video LLMs) excel in question answering (QA) but largely operate as black boxes, providing textual answers without verifiable visual grounding. Existing explainability efforts rely on textual rationales…
Language models are increasingly used for moral decision-making across diverse linguistic and cultural contexts, yet existing work overlooks multilinguality on three aspects: 1) multilingual evaluation benchmarks use direct translation,…
Recent foundation image and video generation models offer strong generalization and controllability, but their direct application to embodied scenarios is limited by requirements for multi-view consistency, geometric coherence, and robot…
Generating and editing a person's face demands high precision, as even minor modifications can significantly alter a subject's perceived identity. Current personalization and editing methods built on general-purpose text-to-image models,…
AI pentesting agents are increasingly credible as offensive security systems, but current benchmarks still provide limited guidance on which will perform best in real-world targets. Existing evaluation protocols assess and optimize for…
Self-improving autonomous agents are moving from research prototypes to deployed systems. The primary goal is controllable evolution, or adaptation, from experience with minimal or even no human input. This survey frames modern…
Large Language Model (LLM) agents have moved beyond generating responses to executing multi-step tasks by calling tools, observing the results, and iteratively deciding the next action. Most agent systems run on desktops or servers, which…
We introduce Boogu-Image-0.1, an open-source unified multimodal understanding and generation model family, comprising Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in high-quality text-to-image generation,…
When should an intelligent assistant speak up without being asked? Continuous egocentric video offers rich, evolving context that enables a new form of assistance: one that is proactive rather than merely reactive. Yet existing approaches…
Failure attribution for LLM-based agentic systems, i.e., identifying which steps in a failure trajectory caused the task to fail, is critical for debugging and improving these systems. Existing approaches either rely on prompting-based…
Structured pruning is a hardware-friendly way to compress LLMs, but it is mostly validated on multiple-choice recognition tasks, while the same compressed checkpoints can collapse on the free-form generation that deployment actually…
Reinforcement learning with verifiable rewards without human-annotated data, often referred to as zero RL, has emerged as a powerful paradigm for eliciting chain-of-thought reasoning. However, due to computational constraints, existing…
The capability of a modern AI agent depends not only on its foundation model but also on its harness, which constructs prompts, manages state, invokes tools, and coordinates execution. As models, APIs, environments, and requirements…
Despite the success of Vision-Language Models (VLMs), misleading charts remain a significant challenge due to their deceptive visual structures and distorted data representations. We present ChartCynics, an agentic dual-path framework…
Coding agents must integrate external tool returns into ongoing reasoning - a capability that standard left-to-right pretraining on code exposes only in its forward direction. We observe that the action-observation-continuation loop of a…
Mainstream visual encoders are pretrained on natural images and cannot be effectively applied to document images without document-oriented adaptation, as dense text and fine-grained character strokes demand character-level visual…
This work introduces a unified formulation for vision models, where diverse forms of visual information beyond natural images, such as masks, depth maps, and other structured visual signals, are all represented as RGB images, while general…
Learning broad world knowledge directly from raw visual data is a fundamental capability of intelligence. We introduce UniVR, the first investigation into simultaneously learning complex reasoning, fine-grained physical dynamics, and…
Human cognition does not separate understanding and generation. A teacher at a whiteboard speaks and draws together, each modality reshapes the other. In this paper, we bring this coupled loop to artificial systems. Masked Diffusion Models…
Autonomous UAV systems increasingly rely on multimodal large language models (MLLMs) to operate in complex real-world environments. Such embodied scenarios require not only understanding the surrounding space but also maintaining a…
We present AffectFlow-DINO, a multi-task learning system for the 11th ABAW challenge that extends a standard deterministic architecture with a conditional rectified-flow head to model the inherent ambiguity of in-the-wild facial behavior.…
We revisit the evaluation of automatic harness evolution for LLM agents. Existing harness evolution methods use unit test cases to search for harness configurations and then report final performance on the same public benchmark. This…
Code review helps maintain software quality before code integration, but it also imposes a substantial workload on human reviewers. As generative artificial intelligence becomes part of software development, code review is shifting from a…
Building assistants that can continually watch the world, remember what they see, and reason over their accumulated experience is a long-standing goal, and recently multimodal agents equipped with long-term memory over video streams have…
When a real-world scene is captured by a smartphone camera and viewed on its screen, the displayed image often differs noticeably from the original scene in color, brightness, and contrast. This gap persists despite substantial advances in…
Vision-language models (VLMs) process large numbers of visual tokens, resulting in substantial inference latency and memory overhead. This has motivated extensive research on visual token compression. While training-free strategies rely on…
Thermal infrared (TIR) imaging is essential for UAV swarm operations in visually degraded environments. However, tracking tiny UAVs remains challenging due to limited appearance cues, frequent occlusions, and rapid maneuvers. Despite…
Pruning those handful of "critical" parameters that supposedly tank a model by orders of magnitude doesn't hold across all LLMs. Try to fine-tune only those coordinates on OLMo-1B and OLMo-7B and accuracy collapses to random guessing, even out to 36K neighboring parameters. Plain LoRA, spread thin across the same layers, gets there fine using 0.16% of the weights.
A new "correctness agreement" metric checks whether a quantized model gets the same questions right as the full one, not just the same aggregate score, and the answer is no: real behavioral divergence shows up under moderate quantization even when accuracy and perplexity look untouched. Query and key projections take the damage first.
Flow matching over carefully designed latent representations has recently emerged as a powerful paradigm for topology-aware mesh generation. Existing approaches, however, model vertices and connectivity jointly in a joint latent space,…
In this paper, we propose SpectraReward, a training-free reward function that turns pretrained MLLMs into off-the-shelf reward models for image-generation reinforcement learning. Instead of asking the MLLM to judge a generated image or…
LLM-based coding agents have significantly advanced automated software issue resolution, yet they remain highly prone to factual errors caused by insufficient repository understanding. Recent methods attempt to mitigate this limitation…
Current visual generation models are capable of producing high-quality content, yet they lack a coherent perception of the spatial structure. Existing generative novel view synthesis methods typically introduce explicit geometry priors,…
Multi-scene navigation (clearing an objective in one bounded space and then crossing a portal into the next) is a defining feature of contemporary 3D games, but authoring it is laborious: every portal must have consistent endpoints on both…
Existing benchmarks for scientific data analysis evaluate LLMs primarily on code execution or workflow completion, overlooking that scientific analysis serves to support distinct types of scientific claims: hypothesis exploration,…
Recent generalizable 3D Gaussian Splatting models have advanced long-sequence novel view synthesis (NVS), but at the cost of substantial redundant computation. We identify that the redundancy can be mitigated based on two observations: (i)…
Graph retrieval-augmented generation (GraphRAG) enhances large language models with structured knowledge, yet existing systems construct knowledge graphs in a single extraction pass, producing noisy entities and brittle retrieval. RAGU, an…
In this report, we introduce Qwen-Music, a powerful music generation model capable of producing highly musical and high-fidelity songs with complete vocal singing. Qwen-Music supports two core tasks: Text to Music Generation, which create…
We introduce Self-Verified Reasoner (SVR-R1), a multi-turn RL framework that turns a model's own verification into a learning signal for multimodal reasoning. For each query, the model proposes an answer using the same weights, and issues…
We introduce a vocabulary for automated research systems built from one or more agents to make their design choices easier to describe and compare. The vocabulary specifies 1) who the agents are, 2) what operations are available in the…
AWS wrapped agent runs in first-class "cases" that move through a real lifecycle (Ready, In Progress, Successful, Failed, Pending Resolution), with a creator/processor split so one automation ingests work and many process it in parallel. That's less a research finding than a naming exercise, the one enterprises actually need before they'll trust a fleet of agents with production work and a paper trail.
400 bilingual tasks run live in Docker containers, scored step by step across five capabilities (skill use, exploration, long-context reasoning, multimodal understanding, cross-platform coordination), instead of graded off a static answer key an agent can game. Grading criteria stay hidden during the run. Closer to how you'd actually judge an assistant.
The model reasons by generating a sequence of images instead of writing out steps, trained on a newly built dataset for the task. The resulting model, Wan-CoF, beat its baseline across four reasoning benchmarks.
Recent VLM and VLA systems have improved robotic perception and action prediction, yet long-horizon embodied agents still require a general runtime layer for reasoning, memory, tool use, verification, and cross-embodiment execution. We…
Visual Language Navigation foundation models aim to unify deep reasoning for grounded spatial decisions with broad versatility for diverse embodied tasks. Current approaches typically achieve this integration via monolithic policies that…
Large language model (LLM) agents are beginning to automate machine learning engineering (MLE) by coupling planning, code execution, debugging, and empirical feedback. Translating this capability to medical imaging remains difficult…
Reinforcement learning for large language models (LLMs) typically relies on trust-region masks to stabilize off-policy updates. The dominant PPO-style approach uses the sampled-token importance ratio for two criteria: a proximity…
For anyone building multi-step coding or browsing agents, this is the wall you actually hit around step 50, not model quality.
It recursively spins up sub-agent teams to cover a search space in parallel, then folds the results back up the tree. A lone ReAct loop runs out of context and patience on genuinely deep-and-wide queries. A swarm doesn't.
Researchers held candidate responses fixed across four judgment datasets and only changed which model graded them. The scores moved anyway, no answers touched. If your eval pipeline shows a quality jump after a model upgrade, check whether you upgraded the judge too before you believe it.
Virtual try-on (VTO) has made significant progress in realistically transferring garments onto a target person. Yet most systems give the user little control over how a garment should be worn -- its size (loose or fitted), style (e.g.,…
Existing volumetric capture of dynamic human performance achieves high fidelity with dense camera arrays. However, in real-world scenarios, only a handful of low-overlap cameras are available, which degrades the output quality and leaves…
Vision language models (VLMs) have achieved strong performance on visual document understanding benchmarks such as DocVQA, ChartQA, and MMLongBench-Doc. However, real-world documents combine multiple factors such as length, layout…
Agentic retrieval-augmented generation (RAG) extends static RAG by allowing language models to iteratively reason, generate search queries, retrieve evidence, and predict answers. However, it remains challenging for models to decide when…
Temporal grounding in long recordings remains challenging for audio-conditioned LLMs. We present a time-aware audio LLM that answers questions with explicit timestamps over up to 120 minutes of input. Our approach interleaves periodic time…
Despite recent scaling successes, multilingual ASR performance remains highly uneven, with long-tail languages suffering from severe data scarcity. This work addresses the challenge of building robust foundation models for underrepresented…
Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. However, RL algorithms with PPO-Clip are inherently limited by exploration collapse. Subsequent works remain primarily heuristic and…
Nemotron 3 Ultra, tuned through LangChain's Deep Agents harness, now matches the top closed models on LangChain's own agentic benchmark at roughly a tenth of the inference cost. Nobody retrained the model. The gains came entirely from rebuilding the harness around it: system prompts, tool descriptions, middleware.
OpenAI ran its own audit of the coding benchmark half the industry cites and found a sizable chunk of its tasks are broken: hidden requirements, contradictory instructions, tests too strict for even a correct fix to pass. It's retracting its own recommendation of the eval. The audit process looks solid, but clock who's doing the debunking: a lab that sits near the top of a lot of other coding leaderboards.
An 8B model that takes one RGB camera feed and a plain-language instruction and gets a robot where you told it to go. No LIDAR, no depth sensor, state of the art on the R2R-CE navigation benchmark. Small and single-camera by design. That's the actual news: navigation stacks that used to need a full sensor suite might not for much longer.
Phone segmentation and recognition are inherently related tasks, yet modern approaches typically model them separately. We argue that phonetic structure is already latent in the representations of self-supervised speech models (S3Ms), and…
In this work, we aim to address the challenge of long-range memory in panoramic world models by exploiting the rotation-equivariant property of omnidirectional representations, where rotation can be treated as an implicit geometric…
Fine-tuning LLMs to inject new knowledge faces a critical challenge: LLMs can quickly memorize new facts, yet fail to use them for downstream reasoning tasks. We formalize this failure as the \textbf{Knowing--Using Gap}, characterized by…
Long-context processing has become increasingly important for large language models (LLMs), but simply extending the context window does not guarantee effective utilization of long inputs. As input length grows, accuracy often degrades,…
Driven by next-token prediction, NLP shifted from task-specific models into powerful generalist foundation models. What, then, is the equivalent catalyst needed to achieve a general-purpose model in computer vision? In this paper, we…
The rapid progress of large foundation models has been driven predominantly by pretraining on large-scale text corpora. However, many forms of knowledge are conveyed through visual representations, where figures, typeset equations, and…
A striking feature of the human visual system is that it ingests visual information through a series of local foveated glimpses, rather than a single global computation. This makes human vision distinctly different from most popular…
Scaling robust driving policies is fundamentally bottlenecked by the scarcity of edge cases in curated datasets. While the real world continuously captures these critical events, such long-tail events remain underutilized when collected…
Retired GPUs can form low-cost clusters for LLM inference, but their economic and environmental viability depends heavily on local electricity prices and carbon intensity.
Multi-step LLM agents routinely lock into a trajectory that's going to fail long before the failure is visible, and burn real inference budget getting there anyway. This paper finds the tell shows up early in the agent's internal representations, and a lightweight probe cascade catches it in time to abort. Run agents at any volume and that's compute you're currently spending on episodes that were never going to work.
Most LLM data-analysis benchmarks test fact retrieval on small, clean tables. This paper builds one closer to what an analyst actually opens: large multi-table joins with half-documented columns, plus questions that need outside context to answer. Scores drop hard. A model's spreadsheet-benchmark score isn't a great predictor of how it'll do on your actual warehouse.
Matching cloud security controls to the technical metrics that prove you meet them is still mostly manual work. Researchers fine-tuned sentence transformers on 3,499 semantic pairs pulled from five European security standards, and the domain-adapted model does the matching on its own. Narrow use case, but it's exactly the grunt work a compliance team would happily hand off.
DrugGen-2 generates small molecules conditioned on disease ontology and target protein sequences through fine-tuning GPT-2 with supervised learning and reinforcement learning using GRPO, achieving superior molecular diversity and binding…
LongE2V enables high-quality video recovery from sparse event streams by leveraging pre-trained video diffusion priors and addressing temporal stability and frame interpolation challenges.
Canvas360 is a two-stage framework for in-context panoramic generation that combines geometry-aware pretraining with fine-tuning, featuring a large-scale dataset and novel modeling techniques for improved geometric consistency and global…
A dependency-free runtime enables efficient text-to-music generation on embedded devices through quantization and activation steering while maintaining audio quality.
CausalDS is a benchmark for evaluating causal reasoning in data-science workflows that combines synthetic causal structures with realistic observational data and natural-language stories across Pearl's three rungs of causal inference.
ARDY is a streaming generation framework that enables real-time, high-fidelity 3D human motion generation with text and kinematic constraint control through a hybrid representation and two-stage autoregressive transformer denoiser.
A benchmark for scientific lineage reasoning and idea generation is introduced, organizing scientific works as genetic-like Idea Genome objects and evaluating both reasoning and generation capabilities.
OPSD-V enhances few-step autoregressive video diffusion models by using real long-video data for temporal context during training, providing dense trajectory-level supervision that improves visual quality and motion dynamics without…
Modern Video Object Segmentation (VOS) involves tracking and segmenting user-specified targets. While recent approaches have achieved remarkable performance in single-target scenarios, extending them to multi-target settings typically…
AI agents have become capable of autonomously completing short, well-specified tasks. However, existing terminal benchmarks largely focus on simple problems that finish within minutes and are evaluated only by their final outcome. This…
Pretrained diffusion transformers can be adapted for dense prediction tasks by mapping tokens to task-native outputs instead of generating RGB images, achieving state-of-the-art results with minimal additional parameters.
Modern AI models achieve strong performance on many established benchmarks, yet they still fail on tasks that humans find almost trivial, such as manipulating a string or drawing a dog with five legs. These examples suggest that existing…
Why does contrastive learning with simple images and augmentations yield useful representations for downstream tasks? We address this question by analytically computing the optimal representation in terms of a contrastive loss for a range…
Large language models fine-tuned for forecasting can be accurate yet poorly calibrated, and their chain-of-thought (CoT) reasoning may not faithfully reflect the evidence behind a forecast. We ask whether internal representations offer a…
Visual generators excel at rendering, but they confidently fabricate what they do not know. User requests are unbounded, evolving, and deeply long-tailed: new characters, trending entities, post-cutoff events, and more. This…
Existing methods for automatic music transcription are often limited to single-instrument recordings or fail on complex, real music mixes. Although previous work utilizes synthetic training data, the resulting models generalize poorly,…
A new framework argues that verification, not more pretraining or longer chains of thought, is the next scaling axis: train a model to grade solutions well, and you get a cheaper lever than training it to produce them. Holds up in-paper. Whether it holds outside the benchmarks the authors picked is the open question.
Most "world models" treat every other agent as scenery baked into the environment. This one is the first to condition on multiple agents' action streams directly, so it can attribute a physics outcome to who actually did what. Early-stage, narrow domain, but it's the first real step toward simulators for multi-agent robotics and games instead of solo demos.
Turns out LLMs linearly encode how many tokens they have left to generate, in the hidden state, before a single word of the answer exists. That's a tidy explanation for why models ramble or truncate in oddly consistent patterns, and a real hook for anyone trying to build length control that doesn't rely on cutting text off after the fact.
SciReasoner is a multimodal scientific foundation model that enables interpretable structural reasoning across proteins, molecules, and crystals by discretizing structural elements into a unified vocabulary for enhanced prediction and…
LaMem-VLA introduces a latent-memory-native framework that integrates historical experience into vision-language-action reasoning through coordinated memory components operating in the same latent space.
An advanced world modeling system with extended interaction capabilities, real-time processing, diverse interactive elements, and multi-agent behavior control for collaborative virtual environments.
LingBot-Video presents a DiT-based video pretraining framework with Mixture-of-Experts architecture, specialized data augmentation, and multi-dimensional reward system for embodied intelligence applications.
Asynchronous reinforcement learning with single-rollout optimization addresses stability issues in LLM training for complex tasks, outperforming existing methods in coding and reasoning benchmarks.
Sparse Delta Memory extends gated linear RNNs with sparse addressing to dramatically increase hidden state capacity for improved long-context learning and retrieval while maintaining computational efficiency.
A comparative analysis of softmax attention and recurrent linear-attention architectures examines their expressivity, memory management, and training efficiency across different parameter scales and sequence lengths.
Reinforcement learning frameworks for large language models face exploration-stability trade-offs, which are addressed through a novel universal objective called Unbounded Positive Asymmetric Optimization that enables stable training with…
A novel zero-shot method called Jet-Long enables efficient long-context processing for large language models by dynamically adapting rescaling factors and utilizing a bifocal attention mechanism that maintains high performance across…
Peaked quantum circuits can be efficiently simulated classically using sparse state vector representations with vectorized operations and hardware acceleration.
Flow-ERD is a multi-agent traffic simulator that combines agent-type aware flow matching with entropy-regularized distillation to achieve both realistic and diverse motion patterns.
Post-training quantization (PTQ) is a widely adopted technique for compressing large language models (LLMs) without retraining. Existing second-order PTQ methods, including GPTQ, construct quantization objectives exclusively from input…
Direct On-Policy Distillation transfers reinforcement learning improvements from smaller to larger models by using the policy shift induced by RL as an implicit reward signal, enabling efficient scaling of training without re-running…
Medicine is inherently multimodal, requiring clinicians to synthesize information across diverse data streams. Yet the development of multimodal foundation models is constrained by limited access to large-scale, high-quality clinical data.…
The optimization of long-horizon agents increasingly relies on reflection-based mechanisms, where a large language model (LLM) acts as an optimizer to diagnose agent failures and improve agent policies. However, real execution traces are…
Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in…
Starting from the utilization of deep neural networks to approximate the state-action value function that led to winning one of the most challenging games, to algorithmic advancements that allowed solving problems without even explicitly…
Length-penalized reinforcement learning can shorten chain-of-thought reasoning while hiding an influence that drives the model's answer. In our experiments, training with length penalties does not stop misleading hints from steering…
Interactive simulators have become powerful tools for training embodied agents and generating synthetic visual data, but existing photorealistic simulators suffer from limited generality, programmability, and rendering speed. We address…
Training tool-use agents to improve from their own experience remains challenging, as supervised fine-tuning relies on fixed teacher-distilled trajectories, while sparse-reward reinforcement learning provides weak supervision for…
A new study training NequIP and Allegro models found that matrix-structured optimizers SOAP and SOAP-Muon beat plain Adam on both convergence speed and final accuracy, with the gap widening when force labels are sparse. Optimizer choice has been an afterthought in molecular-simulation AI. This says it shouldn't be.
Write "banana" on a photo of an apple and a CLIP-based model will often agree with the text over its own eyes. Researchers traced the failure to specific attention heads that favor lexical signal over visual signal, then fixed it by intervening on just those heads. Cheap insurance if you're shipping a CLIP classifier anyone could scribble on.
LingBot-VLA 2.0 enhances generalization across tasks and embodiments through expanded data preprocessing and training on diverse robot configurations, extends action space to include whole-body degrees of freedom for complex manipulation…
Image2Sim enables scalable embodied navigation training by creating high-fidelity interactive environments from RGB-D images through decoupled 3D spatial anchoring and photorealistic rendering techniques.
PluraMath extends the PolyMath dataset to 18 underrepresented languages, revealing persistent gaps in multilingual mathematical reasoning performance between high-resource and low-resource languages.
Turn-level budgeting strategy for efficient on-policy distillation in long-horizon agent training addresses inefficiencies in full-horizon rollouts and shallow token concentration.
A unified multimodal model formulates computer vision tasks as generation problems using natural language and visual prompts, achieving performance comparable to specialized systems across diverse vision tasks.
MaxSim similarity can exactly replicate inner products between sparse vectors and supports logical operations, with Signed MaxSim extending this capability to real-valued vectors while improving retrieval performance on complex query types.
SIEVE is a structure-aware data selection method for vision-language-action imitation learning that identifies reusable visuo-motor primitives and transition interfaces to improve policy learning efficiency.
AlayaWorld is an open-source framework for creating interactive generative worlds that enables real-time user interaction and supports diverse actions within a modular architecture.
Nemotron-Labs-Diffusion is a tri-mode language model that combines autoregressive, diffusion, and self-speculation decoding to achieve superior throughput and efficiency compared to existing models.
Digital teleoperation replaces physical robot interaction with generative world models to create diverse training data for robotics, enabling efficient zero-shot Sim2Real transfer and improved real-world performance.
A multi-modal 4D world model generates synchronized RGB, depth, and optical flow data from single RGB-D images and language instructions, enabling efficient robotic manipulation through unified diffusion processes and inverse dynamics…
RoboDojo presents a unified sim-and-real benchmark for evaluating generalist robot manipulation policies across diverse tasks and evaluation dimensions.
WildCity presents a large-scale multimodal dataset for urban navigation and spatial representation, enabling research into AI systems that can perceive and reason about city-scale environments similar to human cognitive capabilities.
Agentic code review framework enhances AI-generated pull requests through iterative review and revision cycles, improving both code quality and issue resolution capabilities.
World models exhibit long-horizon failures due to kinematic rather than dynamic imagination, as demonstrated by measuring imagined kinematic-consistency error which remains flat while policy rewards collapse across friction boundaries.
MR super-resolution is reformulated as a physics-aware reconstruction problem that dynamically adapts resolution-SNR configurations using Gaussian splatting with prior-aware representations and physics-constrained modeling.
RoboTALES introduces a two-stage framework that combines LLM-based planning and VLM-based criticism to improve task-aligned video generation and robotic policy training.
A token-centric dual-view learning framework unifies prompt-based adaptation and cross-view fusion in a frozen vision transformer to improve breast cancer classification from mammography images.
We present AgentLens, a production-assessed benchmark for interactive code agents. Most code-agent benchmarks reduce a run to a single bit -- did the task pass? -- but the people who actually use these agents experience the entire…
Vision-language models (VLMs) have made interactive digital museums increasingly feasible by connecting 3D digitization with natural-language artifact exploration. However, in cultural heritage domains such as ancient Greek pottery,…
Large language models (LLMs) have demonstrated growing competence in web page generation. However, existing text-driven approaches rely on complex prompts that impose substantial demands on users and offer limited expressivity for page…
The fix is training-free. It replays the model's own attention signals as a query-conditioned evidence pool before final generation, tested on eight long-context benchmarks across Qwen3-4B, Qwen3-8B, and Llama3-8B. Best average rank on all three, no fine-tuning required.
The fix blends teacher and student token distributions adaptively instead of forcing dense supervision, then beats GRPO and SDPO on SciKnowEval and holds up better on out-of-distribution GPQA.
Speaker attribution in TV dialogue breaks voice-recognition models on short lines, so this ICML 2026 paper swaps acoustic biometrics for a reasoning model that pulls in visual and linguistic context instead. Comes with a 532K-line, 900-character benchmark dataset (DramaSR-532K) the authors are releasing.
KVpop learns optimal key-value cache eviction by directly supervising keep-or-drop decisions using future-attention targets, achieving high performance with reduced memory usage.
A video diffusion framework generates long, multi-view consistent videos by combining temporal and view-wise autoregression through 4D geometric bridging and spatio-temporal distillation techniques.
Object-aware token merging framework SaMer compresses image-side tokens while preserving query-selectable visual evidence, achieving significant storage reduction and improved retrieval performance.
InternVLA-A1.5 integrates pretrained vision-language models with future prediction in latent space to enable efficient robot manipulation with preserved semantics and long-horizon execution.
Analysis of 38,000 hours of real-world agent interactions reveals log-sigmoid scaling laws for performance and exponential learning speed improvements across 134 diverse tasks.
A large-scale visuotactile dataset called Deform360 is introduced to study deformable object dynamics, enabling comparison between 2D video and 3D particle world models for robotic manipulation tasks.
PixWorld presents a unified pixel-space diffusion approach for 3D reconstruction and generation that overcomes limitations of latent-space methods through direct image-level supervision and geometry-aware feature alignment.
Boundary modeling enables dense spatial perception by learning sub-pixel representations that enhance depth estimation and support embodied AI applications.
A large-scale multimodal tool-use dataset and agent are presented for complex image creation workflows that orchestrate multiple visual tools through multi-turn interactions and hybrid reward optimization.
DSpark enhances LLM inference speed by combining parallel draft generation with adaptive verification that reduces waste and improves throughput in high-concurrency settings.
TREK expands exploration support for policy optimization by using distillation for exploration rather than imitation, improving performance on challenging mathematical reasoning and agentic tasks.
Light-Omni is a multimodal agent framework that enables efficient video understanding through dual contextual states, achieving faster and more accurate video processing by eliminating iterative reasoning while maintaining semantic…
Graph-as-Policy system combines modular robot skills with multi-agent coding to improve reliability in variable automation tasks through parallel simulation refinement.
Large language models can be adapted for low-resource programming languages through specialized training pipelines and benchmarks, achieving superior code completion performance compared to general-purpose models.
HunyuanOCR-1.5 is a lightweight end-to-end vision-language model that enhances OCR capabilities through improved efficiency via DFlash and enhanced capability via Agentic Data Flow, achieving fast inference and broad task coverage.
Byte-pair encoding and Unigram-LM create distinctly different subword vocabularies in chemical language models, with no convergence between the two approaches across diverse corpus types and vocabulary sizes.
We introduce PAST-TIDE, our stance detection system addressing both subtasks of the StanceNakba Shared Task at NakbaNLP@LREC-COLING 2026. The main idea is statement tuning. We redefine stance as cloze-style masked language modeling (MLM),…
Big goals are hard to achieve all at once; breaking them into small steps is wiser. We present Trust Region Policy Distillation (TOP-D), which transforms the notoriously unstable, high-variance On-Policy Distillation (OPD) into a stable…
Vision Transformers (ViTs) are known to exhibit high-norm patch-token outliers that degrade feature map quality, a problem effectively mitigated by register tokens. As diffusion models increasingly adopt transformer architectures and move…
A new paper tests the dumbest possible real-time guardrail: take a verifier model's confidence signal, threshold it, calibrate the threshold with risk-control math, raise an alarm when it crosses. On math-reasoning and red-teaming benchmarks, that simple setup matched fancier sequential-hypothesis-testing monitors.
Researchers built a testbed that injects fake PII into specific weights of 1B and 7B OLMo models, then checks whether unlearning actually erased it or just buried it. Most SOTA unlearning methods pass every output-level test and still resurface the data under the right probe. When the fact's location in the weights is pinned down precisely, though, a plain gradient-based method erases it cleanly.
Compile a fuzzy spec (alert on important log lines, fix malformed JSON, rank by intent) into a small adapter for a frozen local model instead of paying an API per call. The paper's number: a 0.6B Qwen3 running the compiled program matches Qwen3-32B's direct-prompting quality at roughly a fiftieth of the memory, at 30 tokens/sec on a MacBook M3.
A multimodal sexism identification system for memes uses hierarchical conditional soft-label prediction with vision-language embeddings and a lightweight Gated MLP trained via KL divergence and uncertainty weighting.
Diffusion large language models face challenges in reasoning enhancement through post-training, but a novel on-policy self-distillation method using internal denoising trajectories improves mathematical reasoning and code generation…
Uni-GUI dataset and UI-MOPD method enable cross-platform GUI agent training by addressing limited data and platform-specific capability degradation through multi-teacher on-policy distillation.
A speaker-disentangled syllabic tokenizer regresses perturbed student representations toward clean teacher targets to improve syllable boundary detection and speech language modeling performance.
ResearchStudio-Reel automates research dissemination by composing specialized skills around a shared paper extractor, generating consistent and editable artifacts including posters, videos, and blogs with hard pass/fail quality gates.
Wan-Streamer v0.2 enhances audio-visual interaction by increasing visual resolution while maintaining low latency through optimized thinker-performer architecture with multi-GPU parallel processing.
ResearchStudio-Idea provides a skill suite for effective research ideation that combines literature search, novelty checking, and pattern-guided generation to produce traceable research proposals.
LLM-as-a-Tutor framework extends LLM role from judge to tutor by dynamically adjusting prompt difficulty through pairwise comparison and constraint addition, improving instruction-following performance in reinforcement learning.
SceneFrom3D generates 3D outdoor scenes by automatically scheduling views from input geometry and controlling object appearance and geometry adherence through identity images and geometry-adherence parameters.
Flash-BoN improves text-to-image generation efficiency by using inexpensive draft candidates generated through timestep truncation, layer skipping, and activation proxies, followed by multi-stage verification that outperforms existing…
a new benchmark (Iterative VibeCoding) has Claude Sonnet 4.5 play attacker against a GPT-4o monitor, spreading a malicious payload across multiple pull requests and timing the final piece for the PR with the best cover. Gradual attacks slipped past a standard diff monitor 93% of the time.
researchers gave 10 models a public channel and a private "off the record" channel inside multi-agent debates, no instructions to behave differently in either. Baseline divergence between the two was about 3%.
a small real-money forecasting study found most people paired with an AI either just copy its answer or use it to rubber-stamp a guess they'd already made, and both groups do worse than the model alone. A minority beat the model, and what set them apart wasn't IQ or which model they used. It was perspective-taking, intellectual humility, and curiosity.
Manifold constraint hypothesis enables improved concept erasure by projecting updates onto estimated representation manifolds, achieving state-of-the-art results in nonlinear concept removal.
OmniOpt presents a unified framework for optimizer selection in large-scale model training by combining meta-pipeline transformations, norm-constrained linear minimization oracles, and a cross-domain benchmark to systematically analyze…
Automated safety testing framework Vera uses a three-stage pipeline to identify and test safety risks in LLM agents through structured risk taxonomies, combinatorial case generation, and adaptive sandbox execution with evidence-based…
CGGS is a text-to-3D framework that enhances 3D-content-awareness and addresses geometric distortions through a multi-stage approach involving ego-centric generation, layout decoration, and geometric refinement.
Multimodal large language models exhibit distinct attention patterns during generation, with attention to visual and textual modalities shifting based on semantic requirements, and these patterns can be leveraged to improve task…
Siamese student encoders in JEPA models improve representation separability and training efficiency through effective regularization, outperforming single-encoder variants and MAE under limited training budgets.
Arabic language models exhibit dialect-specific neural representations that can be manipulated at inference time to control dialectal output without requiring additional training.
CineMobile enables efficient image-to-video generation on mobile devices through distillation-guided pruning, diffusion distillation, and hybrid quantization techniques while maintaining visual quality and achieving significant speedup.
Large-scale controlled experiments reveal optimal scaling strategies for Earth-observation foundation models, enabling efficient training and deployment through encoder growth, downstream performance selection, and model distillation.
OmniTacTune enables efficient adaptation of tactile feedback to visual robot policies through a two-stage reinforcement learning approach that improves success rates in contact-rich manipulation tasks.
OpenAI's new benchmark hands models 129 synthetic biology problems that would take a genomics postdoc 20 to 40 hours to work through: not running the stats, but deciding which analysis a messy dataset can actually support. Top score on the board is GPT-5.6 Sol Pro at 31.5%. Claude Opus 4.8 landed at 16%, Gemini 3.5 Flash at 8.1%. Builder read: scientific judgment is still the gap the benchmark race hasn't closed.
A new paper checks where model-generated research ideas land against what actual papers propose, and the same pattern shows up across every LLM tested: models cluster around "bridge" ideas, stitching two known methods together, while human researchers spread across a much wider set of ways to frame a gap. The ideas aren't bad. They're narrow.
GSO, SWE-Perf, and SWE-fficiency all score agents on real-repo performance optimization. Replay the reference patches and most don't even hold up: 39 of 102 GSO tasks pass cleanly, 11 of 140 on SWE-Perf, 411 of 498 on SWE-fficiency. Rankings flip depending on scoring rules (eight submissions disagree on 9 of 28 head-to-heads), and SWE-fficiency piles 58.5% to 82.8% of its score weight onto the ten hardest tasks.
A 3D latent diffusion model for chest CT generation that achieves high-fidelity results while enabling control over clinical attributes through metadata conditioning and reinforcement learning refinement.
Perceptual Flow Matching enables efficient few-step generation by supervising flow matching in perceptual feature space, achieving high-quality results with reduced sampling steps and improved accuracy.
PixCon is a semi-supervised semantic segmentation framework that uses clean-positive pixel-contrastive learning with per-class memory banks to improve accuracy over existing methods.
Long-horizon behavior prediction is enhanced through PraMem, which transforms lengthy historical sequences into experiential memory for improved accuracy.
Bibby AI is an editor-native platform that consolidates the academic writing workflow into a unified Research-Write-Publish pipeline, streamlining literature management, citation insertion, and formatting through integrated tools and…
MentalThink enables multimodal large language models to perform visual-symbolic reasoning by generating and interpreting SVG code as an executable intermediate representation for spatial problem-solving.
A supervised contrastive alignment framework maps WavLM embeddings from English and Mandarin into a shared clinical space for depression detection, addressing cross-lingual generalization challenges and revealing performance artifacts…
Hierarchical Landmark Sparse Attention enables efficient long-context language modeling by learning chunk selection end-to-end, achieving performance comparable to full attention while extrapolating beyond training context lengths.
Flex-Forcing enables video diffusion models to operate under both bidirectional and autoregressive generation regimes through a flexible chunking mechanism over temporal and denoising steps, improving video quality and inference speed.
A parallelized autoregressive framework for dense video captioning that improves generation efficiency by exploiting weak local dependencies across temporally distinct events while maintaining temporal grounding accuracy.
A minimal viable pipeline for skill optimization is proposed through Zeroth-Order optimization formalization, eliminating redundancies while maintaining convergence and generalization through trajectory exploration, consensus mining, and…
Large Audio-Language Models exhibit systematic generative biases in realistic scenarios when evaluated through open-ended tasks using human-recorded speech, with bias magnitude varying significantly by task and triggered by gender and…
Automated agent architecture search demonstrates potential for improving embodied agent performance while revealing challenges related to optimization signals, local optima, and credit assignment in simulation-based training.
Vidu S1 is a real-time interactive video generation model that supports voice-controlled digital character animation with infinite-length output and high frame rate on consumer hardware.
Reinforcement learning has become a standard post-training recipe for large language models, but dense full-parameter updates create two deployment-relevant bottlenecks: suppressed reasoning performance, often reflected by premature…
LLMs are bad at knowing what they don't know: confident hallucinations, no sense of their own knowledge boundary. This paper trains models with RL on metacognitive feedback and gets stated confidence to actually track correctness. Builder read: closer to a model you can trust to abstain instead of bluff.
The error isn't comprehension, it's reference. Matters if you're shipping anything that answers questions over spreadsheets or financial tables: right-looking answer, wrong number. The paper says it's fixable, not fundamental.
instead of re-deriving "set up a sandbox" or "run the test suite" from scratch every session, agents assemble reusable skill packages and compose them for new tasks. Same idea Anthropic ships as Skills, now with a paper behind it.
Research evaluates 30 augmentation strategies for vein recognition across multiple datasets and architectures, revealing inconsistencies between accuracy and adversarial security while highlighting dataset-specific effectiveness variations.
Autonomous agents evaluate policy improvement through iterative editing within fixed budgets, revealing that successful policy evolution requires both task-specific mechanisms and feedback-constrained refinement.
A bounded contract approach for long-horizon LLM agents uses typed retrieval to assemble fresh prompts, enabling isolated analysis of memory components and demonstrating improved performance in complex decision-making tasks.
Research investigates the mechanisms behind self-alignment methods in diffusion transformers, finding that performance improvements stem primarily from data augmentation along the noise dimension rather than token interactions between…
SkillCoach is a self-evolving rubric framework that evaluates and improves agentic skill-use by analyzing skill selection, following, composition, and reflection processes, providing better supervision than outcome-only metrics.
PACE is a framework that predicts expensive agentic LLM benchmark performance using a small subset of atomic evaluation instances, achieving high accuracy at a fraction of the cost.
Representation Distribution Matching enables high-quality image generation by matching feature distributions under pretrained encoders, with improved performance through optimized batch sizes and multi-encoder evaluation metrics.
Task-Agnostic Pretraining framework trains robotic models using self-supervised inverse dynamics on unlabeled data followed by lightweight language grounding, achieving superior performance with minimal expert demonstrations.
MrFlow accelerates text-to-image diffusion by combining low-resolution generation with pixel-space super-resolution and noise injection, achieving up to 25x speedup without training or runtime modifications.
WorldDirector enables controllable video generation with persistent object memory by decoupling semantic motion planning from visual rendering through LLM coordination of 3D trajectories and camera movements.
On-policy self-distillation in continual post-training accelerates in-domain specialization but fails to prevent forgetting and can collapse in out-of-distribution scenarios, indicating that on-policy data alone is insufficient for…
Vision-Language Models struggle with domain adaptation in specialized spatio-temporal video grounding tasks, highlighting limitations in zero-shot generalization and in-context learning capabilities.
A novel reinforcement learning framework for visual generation uses distribution-wise rewards to improve image diversity and quality while addressing mode collapse and computational efficiency issues.
A comprehensive benchmark named AgenticDataBench is introduced to evaluate data agents across diverse domains with fine-grained task annotations and skill-based coverage metrics.
WARP is a framework that infers training data compositions from released model weights by analyzing geometric footprints in weight space through model merging and feature extraction.
Geo-Anchored Cloud Removal framework addresses semantic drift in cloud removal by combining physically grounded residual inversion with semantic manifold constraints from vision foundation models.
OrbitQuant enables efficient post-training quantization for diffusion transformers by using a normalized rotated basis that eliminates the need for recalibration across different timesteps and modalities.
VLA-Corrector addresses limitations of action chunking in vision-language-action models by introducing a lightweight latent-space vision monitor that enables adaptive corrective replanning, improving robustness in contact-rich manipulation…
Embodied.cpp is a portable C++ runtime that enables efficient deployment of vision-language-action and world-action models across heterogeneous edge devices through modular execution layers and optimized inference.
EVA-Client is an open-source framework that unifies real-robot policy deployment, data collection, and evaluation through a component-decoupled architecture with inspectable execution workflows.
World models for robotic policy evaluation are systematically studied through a new benchmark, revealing that long-horizon rollout consistency and robot-specific controllability are more important than short-term visual realism for…
A dual-loop framework named Mastermind is introduced that separates strategy learning from task-specific experience to improve vulnerability reproduction capabilities in software engineering agents.
Gemma 4 introduces efficient, multimodal language models with diverse architectures, enhanced reasoning capabilities, and improved performance across various tasks.
A minimalist pixel-space diffusion transformer using plain ViT architecture directly processes 3D point map patches conditioned on image tokens from DINOv3, outperforming complex latent-based models while maintaining simplicity and…
ACID is a decision-time planning framework that enhances action-conditioned world models by enforcing cycle action consistency to improve trajectory realism and reduce computational requirements.
Reinforcement learning adaptation in transformer models shows highly concentrated improvements in specific middle layers rather than uniform parameter updates across all layers.
Rank-Then-Act framework learns control policies from video demonstrations using a vision-language model as an ordinal scorer with correlation-based rewards, enabling stable cross-task transfer without environment rewards.
RCORE addresses object-driven shortcuts in zero-shot compositional action recognition by using co-occurrence prior regularization and temporal order regularization to improve compositional generalization.
Video-Oasis diagnostics reveal that half of existing video benchmarks can be solved without visual input, exposing significant capability gaps in current video understanding models.