Research

New AI papers from arXiv, picked daily and explained with their claims, benchmark results and code.

2566 articles · page 19 of 54 · The Research front page

All Agents Language models Reinforcement learning Vision Generative media Speech & audio Robotics & embodied AI Safety & security Evaluation & benchmarks Efficiency & systems Code & software AI for science & health Theory

Viewgrid list compact

Safety & security

Dormant backdoor can lie hidden in one model and activate on merging

The authors introduce OMCB, a one-sided merging-conditional backdoor that remains dormant and undetectable in a single specialized LLM, but activates when that model is merged with other models. They provide theoretical grounding for when activation occurs and propose SMOKE, a meta-learning attack that implants such backdoors. Experiments across multiple architectures and merging methods confirm the attack's effectiveness and its evasion of existing backdoor detectors.

15 Oct·3 min

Evaluation & benchmarks

MLLMs struggle to perceive human physical states quantitatively from video

The authors introduce HumanPhys, a benchmark and training corpus for evaluating multimodal large language models (MLLMs) on physically grounded human perception. The benchmark spans 13 sub-tasks covering spatial state, kinematics, blink events, deformation, illumination, and camera pose, with ground truth derived programmatically from sensor calibration or renderer state rather than human annotation. They find that all tested MLLMs perform poorly, with overall scores ranging from 19.74 to 46.94, and that training on the corpus improves performance on held-out benchmarks such as SPAR-Bench.

15 Oct·3 min

Language models

Reward-guided token weighting makes language-model distillation more effective

Chen et al. introduce Reward-Aligned Distillation (RAD), which tests whether token-level guidance from a stronger model is likely to improve the student’s eventual reward. At matched training-response budgets, RAD outperformed standard distillation and token-selection baselines on competition mathematics and coding benchmarks, although it requires additional optimization and reward-gradient computation.

15 Oct·2 min

Efficiency & systems

Near-square experts and depth-aware allocation improve MoE models

The authors study expert shape in LatentMixture-of-Experts (MoE) models, showing that when a fixed parameter budget is split between an expert's input dimension and intermediate width, a near-square balance maximizes the rank of the expert's Jacobian. Guided by layerwise intrinsic dimension (ID) estimates, they further allocate per-expert capacity across depths while keeping per-layer parameters and compute constant, achieving lower validation loss and stronger downstream performance than homogeneous baselines.

15 Oct·3 min

Language models

Multimodal training needs different scaling recipes for text, vision and audio

Khandelwal et al. fit scaling laws to 268 multimodal pretraining experiments, spanning 457 million to 8.3 billion parameters and as many as 150 billion tokens per run. They find that text, vision and audio respond differently to additional model capacity and data, while interactions between modalities depend strongly on direction. Their framework also predicts losses for several larger, unusual-mixture and continual-pretraining runs with low error.

15 Oct·3 min

Evaluation & benchmarks

New benchmark tests whether LLM agents can actively investigate causal systems

Chen et al. introduce ACED-Bench, a neuro-symbolic benchmark comprising 180 tasks in which LLM agents must actively collect evidence from a hidden probabilistic graphical model and return causal answers. The best contemporary models approach ceiling on a qualification check but leave substantial room for improvement on the core benchmark, with evidence-ledger analysis isolating failures in experiment design, evidence retention, and answer commitment.

15 Oct·2 min

Generative media

Calibrated depth reward improves 4D consistency without harming visual quality

The authors propose GeoFix, a framework that uses a single geometric reward (FOCAL) to improve geometric consistency and visual quality in 4D generation from video diffusion models. FOCAL measures cross-view depth consistency in static regions and calibrates it against the expected depth agreement for a given camera field-of-view overlap, preventing larger overlap from being mistaken for better geometry. GeoFix achieves a 39% improvement in WorldScore 3D consistency and a 4% increase in VBench Quality Score compared to baselines.

15 Oct·2 min

Evaluation & benchmarks

Benchmark evaluates AI agents' sustained progress on long-horizon tasks with process rewards

Li et al. present LHTB 2.0, an extended benchmark for evaluating AI agents on long-horizon terminal-based workflows. The authors add 31 tasks (30 via a systematic reconstruction pipeline) to the original 46, correct defects, and record checkpoints every 30 minutes over three-hour trials. Across 1,145 scored runs from 15 frontier models, they find that dense progress ladders capture substantial partial progress that binary pass/fail verifiers miss: on reconstruction tasks, binary scoring averaged 0.221 while dense ladder scoring averaged 0.653.

15 Oct·3 min

Agents

Forecasting future actions makes language-agent training more sample-efficient

Yu and colleagues introduce CoPE, an auxiliary training objective that extracts more supervision from successful agent trajectories by asking the policy to forecast future action sequences. Across simulated and realistic task environments, the authors report better aggregate performance than standard reinforcement learning baselines, including up to twice the sample efficiency on ScienceWorld.

15 Oct·3 min

Reinforcement learning

Hindsight assigns token credit from one verified language-model rollout

Chen and colleagues propose Hindsight-Privileged RLVR, a reinforcement-learning method that extracts token-level credit from a single completed attempt. Across mathematics, code generation and embodied-agent tasks, it reduced rollout requirements and training time relative to eight-sample GRPO, although neither method was consistently more accurate across domains.

15 Oct·2 min

Efficiency & systems

Tree-trained draft branches make speculative decoding accept longer continuations

Zhao and colleagues extend speculative decoding, where a small drafter proposes tokens that a larger language model verifies, from a single continuation to a budgeted tree of alternatives. Across two Qwen3 targets, TreeSpark accepted 23.4% to 31.6% more tokens per verification round than DSpark and improved single-request throughput by 15.1% to 25.0%.

15 Oct·3 min

Reinforcement learning

Hyperbolic geometry beats extreme network depth in goal-conditioned maze navigation

Klein et al. modify contrastive reinforcement learning by measuring distances between learned representations along hyperbolic geodesics rather than Euclidean straight lines. On hard maze tasks, this geometric bias let comparatively shallow agents represent corridor-following routes more effectively than much deeper Euclidean networks.

15 Oct·2 min

Evaluation & benchmarks

Coding agents’ reports omit much of their recoverable task state

Xu and colleagues introduce CTXGAP, an evaluation that tests how much information about a coding agent’s execution remains recoverable from its accumulated user-facing reports. Across 4,799 valid task-state questions, the authors find a 40.7 percentage point gap between report-only and full-context review, including after repairs passed hidden tests.

15 Oct·3 min

Agents

Translation agents learn to retrieve evidence only when it improves results

Li and colleagues present Selective Retrieval-Augmented Translation, an agentic reinforcement-learning framework that treats retrieval as an optional translation action rather than fixed preprocessing. Across idiom, knowledge-intensive and document-level translation, the authors report that this selective policy consistently beats systems that either always or never retrieve.

15 Oct·2 min

Robotics & embodied AI

Action-aware embeddings retrieve better robot experiences for future actions

Liu and colleagues trained a vision-language-action embedding model that represents both a robot’s present state and the future continuation of prior experience as searchable vectors. On a new retrieval benchmark, native actions substantially improved matching, while retrieved demonstrations and evolving task-specific memory increased success for several frozen or conditioned robot policies.

15 Oct·3 min

Agents

Progress-aware training helps small language model agents escape repetitive loops

Cao et al. examine progress stagnation, a failure mode in which distilled small language model agents repeatedly pursue unsuccessful strategies during long interactions. They propose Progress-Aware Distillation, or ProD, which compares student and teacher trajectories and applies a stronger training penalty when the student shows more stagnation. Tests across eight models and both in-domain and out-of-domain tasks indicate that this approach reduces looping while improving success and efficiency.

15 Oct·3 min

Generative media

Spatially addressed memory preserves scenes across long generated videos

Cetin et al. introduce SMART, a memory system for video generation that retrieves observations according to where they were captured rather than how recently they appeared. On long trajectories in Memory Maze and Minecraft, the authors report that it preserves scene-specific details beyond its training horizon while keeping memory size and per-step computation constant.

15 Oct·3 min

Generative media

World model learns where to predict and where to generate

Arapin and colleagues divide future-frame prediction into two components: predictable changes caused by an action, and uncertain content such as newly revealed surfaces or dynamic textures. Their model learns a pixel-level uncertainty map that gates diffusion, improving action-conditioned prediction across robotic manipulation and camera-motion tasks.

15 Oct·2 min

Generative media

One LiDAR model generates, forecasts and resimulates driving scenes in real time

Che and colleagues present LidarVerse, a conditional world model that generates new LiDAR scenes, predicts future point clouds and resimulates recorded environments from new sensor trajectories. By distilling a diffusion teacher into a four-step causal model, they report persistent rollouts at 11 frames per second on one NVIDIA H100 GPU, exceeding the native 10 Hz rate of common automotive LiDAR.

15 Oct·3 min

Reinforcement learning

Context-aware credit assignment improves language agents on long-horizon tasks

Han et al. introduce Contextual Credit Policy Optimization (CCPO), a reinforcement learning method that assigns step-level credit from sparse terminal rewards without training a separate value model. Across ALFWorld and WebShop, two simulated long-horizon agent benchmarks, CCPO reached up to 97.4% and 79.4% success, respectively, using Qwen2.5-Instruct models. The authors report gains over GRPO of as much as 21.6 and 18.5 percentage points.

15 Oct·2 min

Evaluation & benchmarks

Benchmark exposes separate static and dynamic memory failures in video worlds

Zhou and colleagues introduce a 260-case benchmark that evaluates memory in video world models using controlled camera departures, repeated revisits and off-screen events. Across nine models, the authors find that retaining visual content does not guarantee coherent state evolution, and that memory often deteriorates over successive returns.

15 Oct·3 min

Language models

Iterative reasoning states help multimodal models refine their own answers

Yu and colleagues recast multimodal chain-of-thought reasoning as an iterative state-update process rather than a sequence of separately trained stages. Their ReState framework learns from empty, correct, erroneous and model-generated reasoning states, then repeatedly feeds its current reasoning back into the model at inference. Across several model backbones and public benchmarks, the authors report improvements ranging from 0.18 to 13.07 percentage points.

15 Oct·2 min

Language models

A single transformer trains nested model sizes in one pass

Jin and colleagues introduce an elastic transformer whose smaller sub-networks are exact prefixes of the widest model’s hidden states. A single full-width pass can therefore train the whole model family at roughly the per-step cost of its largest member, while their parameterization transfers tuned hyperparameters across widths and keeps updates well-scaled for every supported sub-network.

15 Oct·3 min

Agents

Predicting tool outputs speeds reinforcement learning without corrupting rollouts

Montgomery, Zhang and Wang introduce ObsSpec, a speculative execution pipeline for reinforcement learning of tool-using agents. On terminal-use tasks, it reduced agent latency by up to 14.2% and increased asynchronous rollout throughput by up to 11.5%, while ensuring that retained training rollouts contained only verified observations.

15 Oct·3 min

Language models

UltraBERT balances masking strategies to improve classification and retrieval

Park, Lee and Kang investigate bidirectional encoder pretraining at scales of up to 8.8 trillion tokens, focusing on masked language modeling, where models reconstruct hidden tokens. They report that downstream gains level off after roughly 1 to 2 trillion tokens, while the proportion of masked tokens determines whether an encoder is better suited to classification or retrieval. Their UltraBERT recipe combines both regimes in one model.

15 Oct·3 min

Generative media

Spatial rewards and staged training improve image layout accuracy

Tan et al. address two weaknesses in reinforcement learning for spatial text-to-image generation: unreliable reward signals and training rollouts that attempt complex compositions without first establishing basic skills. Their framework uses an agentic evaluator called SpatialCritic and a curriculum that progresses from foundational spatial relations to combinations of those relations, producing gains on both SD3.5-Medium and BAGEL.

15 Oct·2 min

Safety & security

Enterprise agent risks emerge when local decisions break shared constraints

Ma and colleagues present EAIR, a framework for evaluating risks that emerge when enterprise LLM agents with different roles, permissions and information interact. Across five controlled micro-worlds, the authors report recurring cases in which individually permissible actions collectively violate organization-wide rules, then use workflow traces to locate contributing interaction boundaries and guide repairs.

15 Oct·2 min

Evaluation & benchmarks

Agentic image checks test depicted facts, not AI provenance

Wang and colleagues define Image Factuality Assessment, which asks whether the situation shown in an image is supported by external evidence without providing the system a textual claim to verify. Their trained 9-billion-parameter agent reached 79.05% balanced accuracy on a new 1,527-image test set, although only 52.26% of its outputs met the benchmark's strict standard for sufficient evidence.

15 Oct·3 min

Language models

Query-adaptive neuron steering improves frozen language models across benchmarks

Pan and colleagues test whether an LLM’s response to a query can itself identify useful internal interventions. Their Q-NeuS framework trains a lightweight gate to modulate selected neurons while keeping the base model frozen, improving results across seven models and transferring from MMLU-Pro training to unseen benchmarks.

15 Oct·2 min

Agents

Intermediate snapshots improve long-horizon coding agents at lower cost

The authors present SPROUT, a reinforcement-learning rollout framework for software tasks that can take many steps or hours to complete. By checkpointing the filesystem and agent context after every assistant turn, then using verifier feedback to choose where to retry, SPROUT produced more successful trajectories with fewer tokens and improved training efficiency over standard GRPO.

15 Oct·2 min

Agents

Terminal agents can build adaptive environments for their own training

The authors turn reinforcement-learning environment design into a task performed by a terminal agent. Their proposer inspects a solver’s rollouts, builds complete executable tasks in a sandbox and validates them before training, producing environments more efficiently than prompt-only generation and improving terminal-agent benchmark scores.

15 Oct·3 min

Evaluation & benchmarks

Frontier AI agents solve most expert-level theoretical physics calculations

Chandra and colleagues tested autonomous AI agents on research-level calculations spanning theoretical physics, using problems whose answers were not publicly available during evaluation. GPT-6-Astra in Codex solved 98 percent in a realistic tool-enabled setting, while two leading systems reached 92 percent without internet access.

15 Oct·2 min

Generative media

Models can isolate restricted image concepts during pretraining

Rajan and colleagues redesign text-to-image pretraining so that knowledge associated with designated unsafe concepts is directed into a removable component rather than erased afterward. In experiments on ImageNet and CC12M, detaching that component reduced restricted-concept generation while retaining better image quality, prompt alignment and compositionality than excluding unsafe training images entirely.

15 Oct·2 min

Agents

Agents search more effectively by choosing which reasoning branch to revisit

Li et al. replace an agent’s single expanding context with a tree of summarized checkpoints that the same model learns to navigate. Under matched active-context limits, their trained system achieved higher accuracy than flat ReAct and FoldAgent on both deep-research questions and software bug fixing.

15 Oct·3 min

Evaluation & benchmarks

Base models can reveal agent skills before instruction tuning

Qin et al. introduce ASAP, a framework for evaluating pretrained base models on end-to-end software-engineering and deep-research tasks without fine-tuning them or converting them into chat agents. The authors report that it distinguishes models and checkpoints while producing execution traces that reveal changes in exploration, editing, testing and recovery.

15 Oct·2 min

Code & software

Verifier-based rewards train coding agents toward formally correct programs

Ma et al. train coding agents against a live program verifier, rewarding partial progress on formal proof obligations rather than only successful verification. Their approach substantially increased held-out verification rates and also improved repository patches evaluated with withheld tests.

15 Oct·3 min

Robotics & embodied AI

Human demonstrations scale whole-body learning for humanoid loco-manipulation

The Lambda-humanoid Team presents HumanVerse, a 500-hour dataset pairing first-person video with synchronized body and reconstructed hand motion, plus λ₀, a vision-language-action policy trained to convert that experience into humanoid behavior. On Unitree G1 evaluations, increasing whole-body human data consistently improved task progress, and the full system reached 90.0% success on SIMPLE and 62.5% across four real-world loco-manipulation tasks.

15 Oct·3 min

Efficiency & systems

Verified retries let smaller code models recover lost capability

Galanti, Susskind and Malach test whether repeated, verified attempts can compensate for using a smaller or more heavily compressed language model. In code generation, they find that correct solutions often remain reachable within a modest number of samples, and parallel generation can exploit otherwise underused hardware without multiplying latency in proportion to the number of attempts.

15 Oct·3 min

Reinforcement learning

Symplectic latent dynamics prevent rollout contraction in visual world models

Yalovetzky, Balestriero and LeCun replace an unconstrained JEPA predictor with a learned, action-conditioned Hamiltonian flow over position-like and momentum-like latent variables. On conservative and actuated visual pendulums, the resulting model preserves latent volume by construction and predicts long-horizon velocity more reliably than the baseline.

15 Oct·3 min

Efficiency & systems

Two-bit scalar quantization better preserves language model predictions

Chen and colleagues present FidelityQ, a framework for compressing language-model weights to two bits, giving each weight only four possible reconstruction values. The method reshapes weight distributions, chooses reconstruction levels row by row, and calibrates on sequences where the quantized model most strongly disagrees with its full-precision teacher. On LLaMA-2-7B, it approaches stronger codebook methods while using a regular scalar representation.

15 Oct·2 min

Evaluation & benchmarks

Coding agents produce functional code that often fails security tests

Wu and colleagues introduce LSCBench, a benchmark that tests whether coding agents address security risks when users specify only desired functionality. Although agents completed most tasks correctly, their code passed both functional and security tests in no more than 25.1% of cases.

15 Oct·2 min

Generative media

Persistent memory keeps generated 4D worlds consistent across camera paths

Zhu and colleagues present Demiurge, a system that turns one monocular video into a navigable 4D representation, meaning a 3D scene that evolves over time. Rather than generating every novel camera path independently, the system follows a generate, recover and update cycle built around persistent geometric memory. The authors report more consistent appearance and geometry across trajectories, including views with extreme parallax.

15 Oct·2 min

Generative media

Diffusion transformer depth profiles reflect prediction targets, not just feature granularity

Meng and colleagues re-examine why intermediate features in diffusion transformers improve and then deteriorate on downstream vision benchmarks as network depth increases. Their experiments support a three-phase account centered on whether the model predicts noise, clean images or a mixture, and they use this account to anticipate where REPA representation alignment should be applied.

15 Oct·3 min

Generative media

One Bitwise Diffusion Framework Generates Across Three Multimodal Domains

Batzolis and colleagues test whether diverse modalities can be represented as binary streams and modeled through the same continuous diffusion protocol. Across three separately trained systems, a single checkpoint per domain supports joint generation and multiple conditional tasks, with strong cross-modal consistency despite having no modality-specific parameters inside the denoiser.

15 Oct·3 min

Agents

Recursive harness edits help agents adapt without answer keys

Xu and colleagues introduce Harness^2, a framework that modifies the prompts, memory, tools and orchestration around a frozen language model while it tackles previously unseen tasks. Across three professional benchmarks, the authors report that one refinement step consistently improved tested GLM-5 and Gemini 3.5 Flash configurations, without relying on ground-truth answers to select edits.

15 Oct·2 min

Agents

Open training stack scales mobile agents across apps and tools

Yan and colleagues built an open training stack for agents that operate smartphones, including a controllable environment spanning more than 110 apps, nearly 12,000 demonstration trajectories and over 2,000 reinforcement-learning tasks. Models trained on the stack achieved competitive benchmark results and transferred to real-device evaluation, while also supporting tasks that mix screen interaction with structured app functions.

15 Oct·2 min

Language models

Sparse replay memory extends linear attention to million-token contexts

Li Nanbo and colleagues augment linear-attention language models with a sparse episodic memory that preserves selected text outside the model’s fixed recurrent state. On six RULER retrieval tasks, NoteBank increased Gated DeltaNet’s mean accuracy at one million tokens from 1.7% to 70.3%, while adding at most 0.025% to estimated layer-level computation.

15 Oct·2 min

Reinforcement learning

Advantage guidance makes supervised policy training competitive with offline RL

Zhou and Levine present Advantage Guidance, a policy-learning method designed to retain supervised learning’s stability while improving beyond the behavior in an offline dataset. The method trains a policy to respond to estimated action advantages, then uses contrastive guidance at inference to favor high-advantage actions under a constraint that keeps the policy close to the data distribution.

15 Oct·2 min