Research

Foundational LLM capabilities only partially predict agentic execution gaps

The authors develop a framework that links foundational LLM benchmarks to agentic operations by measuring capability realization gaps, deviations of observed execution behavior from predictions based on offline capability scores. Testing 15 model services on the τ²-Bench Banking tasks, they find that foundational profiles explain only part of the variation in operation scores; models with similar benchmarks can differ substantially in execution, and these gaps persist across task subsets. Adding operation profiles substantially improves held-out prediction of overall task reward.

Self-evolving AI agent generates cinematic multi-shot videos by iteratively directing and critiquing itself

The authors propose CineEvolve, a framework where a single vision-language policy acts as both Director and Critic to generate cinematic multi-shot videos. Trained on a dataset of 40K cinematic trajectories and stabilized by a novel Evolutionary GRPO algorithm, the model outperforms existing methods in narrative coherence and visual execution.

Speech & audio

Omni models struggle to track who speaks when in long videos

Ge and colleagues introduce OmniLingual-MsASR, a benchmark testing whether multimodal language models can transcribe speech while identifying speakers and timing their utterances. Across nine models, the authors find substantial weaknesses in temporal alignment and speaker identification, even when the recognized words are comparatively accurate.

Single-stage RL method unifies preference alignment and few-step image generation

The authors propose AdvFidelity, which integrates artifact-sensitive fidelity supervision into reinforcement learning (RL) through an online adversarial reward. This enables joint reward alignment and high-fidelity few-step generation within a single RL stage, without explicit distillation objectives. Across two model backbones and three task settings, AdvFidelity matches full-step LeapAlign in alignment, task performance, and overall generation quality using only 4 or 8 sampling steps without classifier-free guidance.

Continuing long-context pre-training keeps improving models, and gains are predictable

Ziqing Qiao, Xiaoyue Xu, and colleagues at Tsinghua University and OpenBMB ask whether long-context continual pre-training (LCCP) yields sustained, predictable improvements as its token budget grows. They fit an additive power law to 109 training configurations across five model scales, introduce a stable checkpoint-averaging protocol for monitoring open-ended LCCP, and show that benefits persist through context extension and supervised fine-tuning without clear saturation.

22 Oct·1 min

Masked diffusion language models improve by encoding position priority as hierarchy levels

Gupta et al. introduce Hierarchical Masked Diffusion Language Models (HMDLM), which extend masked diffusion with a hierarchy variable per position and derive the corresponding forward/reverse processes and NELBO objective. In two-stage decoding, the model first predicts which positions are high-priority, then fills tokens conditioned on that assignment. The approach reduces perplexity by up to 7 points on WikiHow and improves Sudoku solution accuracy by up to 8.6 absolute points over diffusion baselines.

Efficiency & systems

Reusing target probabilities in speculative decoding can improve tree selection but requires careful trade-offs.

The authors investigate when next-token probabilities from a target model's verification pass can be reused to select the next tree of draft tokens. They find that while exact reuse is only guaranteed under restrictive conditions that cost accepted tokens, using a separately computed probability distribution for the first token (root row) improves acceptance rates and decoding speed. In experiments across six language models, both root selection and root-draw selection outperform draft-based ranking.

Edge vision-language models are as physically attackable as classic object detectors

The authors benchmark eleven detection and grounding models against physical adversarial patches worn on a person, introducing a conditional attack success rate that only scores images where a model still localizes under a matched gray occluder. They find the compact VLMs are as vulnerable as classic detectors. Failure modes split into omission, substitution, and displacement, with the latter two producing confident, well-formed boxes that downstream planners cannot distinguish from correct detections.

Shaping teacher behavior improves self-distillation for language models

The authors propose TeachBack, a method for on-policy self-distillation that shapes the teacher's generation behavior by controlling the length of reference reasoning and prompt style. Using a bidirectional preference objective, the teacher guides the student while the student's behavior feeds back to improve future teacher supervision. Experiments on mathematical and logical reasoning tasks show that TeachBack consistently outperforms existing OPSD methods.

LLM agents often ignore safety refusals when using tools

The authors construct a benchmark of 2,304 prompts across four domains, comparing text-only, standard tool-enabled, and adversarially framed tool-enabled requests. They find that models refuse 44–76% of harmful chat prompts, but then perform unsafe tool calls on 10–52% of those same refused requests, showing a systematic transfer failure of refusal behavior to action contexts.

Vision

Balancing within-person and between-person variation improves clothes-changing re-identification

Researchers from Reichman University and Microsoft introduce MINER, a model-agnostic training framework that explicitly balances intra- and inter-person variance for Clothes-Changing Person Re-Identification (CC-ReID). By stochastically mining hard positive and hard negative examples and adaptively controlling the variance trade-off, MINER improves retrieval across four CC-ReID benchmarks and ranks among the top three in all three tracks of the TVRID 2026 challenge.

Every learnable distribution class is also differentially privately learnable

Afzali et al. show that any class of distributions learnable to arbitrarily small total variation error with finitely many samples is also learnable under (ε,δ)-differential privacy. The result settles a conjecture posed by Ashtiani in 2022 and provides private samplers for any learnable class.

Simulated users test coding agents as requirements unfold across conversations

Brehme and colleagues introduce SWE-DIVE, a framework for evaluating coding agents through conversations rather than a single complete task description. In a preliminary evaluation across 11 tasks, the authors report that simulated user personas produce distinct interaction behaviors and that unconstrained or changing requirements lead to more code regressions, changes that break previously working behavior.

Training on self-generated answers improves reasoning in low-bit language models

Zheng and colleagues introduce OnQ, a training pipeline that helps aggressively quantized language models recover reasoning performance by learning from their own generated responses. Across six model-size and precision settings, the authors report the highest average score over nine benchmarks, including a 10.26-point improvement over the strongest baseline for Qwen3-8B at 2 bits.

Coding agents often miss crashes and misread changes that prevent them

Huang and colleagues tested four coding agents on source-code versions that produce opposite crash outcomes for the same input. Even the best-performing agent missed 27.5% of real crashes, and examining relevant code did not consistently help agents distinguish crashing programs from their non-crashing counterparts.

Shifting denoiser inputs enables guided sampling from discrete diffusion models

Curtis McDonald and Peter Bartlett propose guiding a diffusion model by changing where its denoiser is evaluated, rather than adding a correction to its output. For discrete data represented as binary vectors, they derive exact relationships between base and tilted denoisers, then introduce training modifications intended to make those relationships usable with a neural network.

Robotics & embodied AI

World models calibrated to hidden physics via short interaction clips

The authors propose PhysiCalWorld, a method that calibrates action-conditioned world models to instance-specific dynamics by prompting them with a short exploratory interaction clip. The clip is compressed into a compact physics descriptor token that grounds subsequent predictions in the actual physical behavior of the scene. Without requiring physics labels, the method substantially improves rollout prediction accuracy and visual planning success across diverse environments.

Quasipolynomial algorithm finds approximate fixed points of ell_infty contractions

Soleymani and Farina give a deterministic algorithm that computes an epsilon-near fixed point of any d-dimensional ell_infty contraction in time log(1/epsilon)^{O(log^2 d)}. The method uses a boosting argument to shrink a search region, and implies that Condon's simple stochastic games can be solved in quasipolynomial time.

Benchmark tests proactive memory for wearable AI assistants in streaming egocentric video

The authors introduce ProactEgoMem, a benchmark for proactive memory in streaming egocentric video. It defines three object-centric tasks, Information Capture, State Tracking, and Location Tracking, at two levels of proactiveness (notification and intervention). Evaluating multiple proprietary, open-source, video-language, and streaming baselines shows that even the strongest models fail to capture memory both timely and accurately.

Time-resolved diagnostic separates irreducible velocity ambiguity from reducible training error

Mei et al. propose the Velocity Ambiguity Profile (VAP), a time-local Bayes-risk floor for velocity prediction in flow matching, and prove that the time-resolved loss decomposes as L_N(t) = A_v(t) + R(t;N), where only R(t;N) can be reduced by more data or a better estimator. In controlled Gaussian-mixture experiments the diagnostic behaves as predicted, and on CIFAR-100 DINOv2 feature flow guided sampling cuts uniformly evaluated velocity loss by 6.78% and global feature Fréchet distance by 23.6%, though class-conditional KID shows no improvement.

Teaching data agents to use charts inside their reasoning loop improves analysis

The authors probe how data agents use charts during reasoning, find a persistent gap between chart-reading capability and actual chart use, and propose ChartMind, a method that distills chart-selection behavior into a reusable, frozen experience package. On three held-out benchmarks, ChartMind improves accuracy and visual evidence quality across four backbones.

Formatting loss in document ingestion makes agents confidently wrong

The authors show that mainstream document converters silently discard visual cues like highlight and colour that authors use to mark obligations in policy documents. On 76 real documents, recovering author-flagged requirements dropped from F1 0.690 to as low as 0.075 when formatting was stripped, and agents gave fluent but wrong answers with a fourfold increase in false positives.

Future chunk prediction improves physical plausibility in autoregressive video generation

The authors introduce Future Chunk Prediction (FCP), an auxiliary objective for training autoregressive video diffusion models. In addition to standard teacher-forced denoising, FCP predicts a future chunk without access to intervening visual context. They report improved physical plausibility across model scales and conditioning settings, with a lower overall training compute budget, suggesting that future-chunk supervision efficiently models visual evolution.

Tree-based method reveals fine-grained LLM capabilities hidden by aggregate scores

The authors propose TIDE, a method that induces a decision tree from benchmark samples and model scores to diagnose fine-grained capabilities. Applying TIDE to five models across four benchmarks, they find that aggregate scores conceal substantial variation: for example, Claude Opus 4.6 is perfect on microeconomic reasoning while GPT-5.4 scores 83.3%, but GPT-5.4 outperforms on contract law reasoning (88.9% vs 77.8%).

Projecting conflicting gradients during joint RL and distillation improves agent training

The authors introduce CORD, a conflict-aware gradient combination framework for joint reinforcement learning and self-distillation. It detects when RL and distillation gradients oppose each other and projects the non-protected gradient onto the orthogonal complement of the protected one before combining. Across three joint-training recipes, CORD improves ALFWorld success by 1.5 to 15.9 percentage points, and one WebShop recipe gains 3.9 points.

Benchmark tests LLM agents on recreating published computer science contributions

The authors introduce FrontierCS 2, a benchmark of 48 open-ended tasks across 13 computer science domains. An agent inherits a repository with a paper's contribution deleted and must submit patches to beat the original code on every metric and workload, judged privately. The best agent, Claude Opus 5.5, passed 58% of runs, but no agent passed any of 16 tasks, and most high-scoring runs still failed on some metric.

Coding agents struggle with system-level performance optimization in microservice systems

Zhang et al. introduce MicroPerf-Bench, a benchmark for evaluating coding agents on end-to-end performance optimization in microservice systems. They injected empirically grounded performance anti-patterns into seven real-world open-source systems, creating 65 optimization instances. Evaluating nine agent-model configurations over 1,755 runs, the best configuration achieved a normalized score of only 42.0%, with the dominant failure being functionally correct changes that failed to improve performance.

Differentially private online prediction achieves regret proportional to square root of best expert's loss

Bo Li and Peng Ye present the first differentially private algorithm for online prediction from experts that attains regret Õ(√L*), where L* is the cumulative loss of the best expert. This improves over the standard Õ(√T) regret when L* is much smaller than T, showing that privacy incurs only a small additional cost in the low-loss regime. The algorithm works under an oblivious adversary and its regret matches the non-private lower bound up to poly-logarithmic factors for constant privacy parameters.

Vision inputs expose greater social bias than text in matched tests

Rezaei and colleagues built Melange, a benchmark comparing how multimodal language models answer socially sensitive questions presented through text or images. Across 23 open-source models, the authors report an average fairness-score decline of 0.291 when switching from text to vision, despite high accuracy at recognizing the depicted demographic attributes.

Shared embeddings let smaller query models search larger models’ document indexes

Krimmel and colleagues introduce 0.6-billion- and 9-billion-parameter retrieval models trained to reproduce representations from the same 18-billion-parameter teacher. The authors report that indexing documents with the larger model and encoding queries with the smaller one improves retrieval over using the smaller model alone, without increasing query-encoding cost.

Reinforcement learning

Training models to recover regions from descriptions improves visual grounding

Zhang et al. train multimodal models to describe a selected image region or video event, then locate it again using that description. Rewarding successful reconstruction improves localization scores and encourages video captions that distinguish a particular event from similar occurrences.

Learned attention proxies help streaming video generators retain historical context

Streaming video generators often discard older context to keep memory and computation manageable, sacrificing information that could guide later frames. Gao and colleagues introduce Attn-Forcing, which learns compact attention proxies for that history, and report better generation quality than baseline methods without retaining an ever-growing context.

Preserving camera-motion information helps video generators maintain more consistent geometry

Zhu and colleagues examine why video generators can produce convincing images while losing track of how the camera moves through a scene. Their CausalAlign framework identifies where camera-motion information weakens inside the model and reinforces it, reducing reprojection error by 22.6% relative to the base model in the reported evaluation.

Jointly learning what to store and retrieve improves agent memory

Xia and colleagues present Uni-Mem, a system that learns both how to build an AI agent’s long-term memory and how to retrieve information from it. The authors report better performance than the strongest reinforcement-learning-based memory manager in their comparisons on LoCoMo and LongMemEval, using training signals designed to distinguish storage failures from retrieval failures.

AI text detectors can expose readable features without sacrificing performance

Dugan and colleagues present a pipeline that turns a neural AI-text detector into a classifier using features with natural-language descriptions. They report that the resulting detectors retain the underlying model’s performance while improving generalization to unfamiliar text, making their decisions easier to inspect.

Coding agent rankings change when debugging becomes a longer conversation

Wu and colleagues introduce UserSWE, a benchmark that tests coding agents through conversations with simulated developers whose communication styles change between turns. Across 72 tasks from 42 repositories, the authors find that the strongest one-turn agent is not the strongest after eight turns, showing that single-turn scores miss important differences in interactive debugging.

Reinforcement learning

Shared hidden-state drift can collapse language generation while preference rankings improve

Li and colleagues show how preference training can improve response-ranking metrics while pushing a language model toward uncontrollable repetition. They identify a shared shift in the model’s internal representations that relative ranking losses barely register, then use interventions to show how it can make one token dominate generation.

Scalable flow language model rivals autoregressive quality at 2.5x speed

The authors introduce Lafa, a family of continuous flow language models scaled up to 8B parameters using fewer than 30B training tokens. By adapting pretrained autoregressive (AR) models to a blockwise asynchronous flow formulation, Lafa achieves quality competitive with AR models and enables a self-speculative decoding scheme that delivers 2.54x the throughput of a strong AR baseline.

Reinforcement learning

Self-Credit rescales self-assigned credit to improve long-horizon agentic RL

The authors introduce Self-Credit, a method that uses self-generated summaries with citations to assign credit in long-horizon agentic RL, refined by comparing actions across successful and failed rollouts. On BrowseComp-Plus, it improves by up to 37 points over base models and beats strong baselines by 5-8 points. Similar gains appear on GAIA and xBench-DeepSearch.

Vision

Cross-validation between scene maps reduces drift in pose-free 3D reconstruction

Wu et al. introduce CrossFit-GS, a method for training 3D Gaussian Splatting models from image sequences without camera poses from COLMAP or another Structure-from-Motion pipeline. Its central idea is to prevent a scene model from validating camera poses using images it has already fitted, which the authors identify as a source of self-reinforcing trajectory errors.

Vision

Point cloud aware language model enables controllable fine-grained 3D object placement

The authors introduce GALA, a multimodal language model that consumes point clouds to predict 6D object poses for 3D scene arrangement. It significantly outperforms baselines that rely only on oriented bounding boxes, and they release a curated dataset and agent pipeline.

Computing approximate correlated equilibria is as hard as computing Nash equilibria.

The authors show that computing an approximate correlated equilibrium of a polytope game is PPAD-complete, making it polynomial-time equivalent to finding a Nash equilibrium. They introduce fixed sets, a relaxation of fixed points, and prove that computing such a set is PPAD-hard at inverse polynomial precision. This result overturns the classical view that correlated equilibria are computationally easier than Nash equilibria.

One model edits people, voices and movements in sync

Yuan and colleagues present AVE, an end-to-end generative system designed to edit a person’s visual presence and speech together rather than passing outputs through separate video and audio tools. The authors report better realism, temporal stability and lip-speech alignment than cascaded baselines across person swapping, insertion and deletion.

Compression preserves broad LLM skills while weakening context-sensitive abilities

Ning and colleagues evaluate more than 35 compressed checkpoints using ten benchmarks designed to probe four broad ability dimensions, rather than relying only on aggregate task accuracy. They find that compression reshapes models unevenly: coarse and surface-level behavior often survives, while fine-grained, long-tail and context-dependent capabilities are more vulnerable.

Larger vision-language models can be worse on specific subtasks

Saxena et al. evaluate 19 vision-language models from four families on eight benchmarks, using per-question comparisons to reveal inverse scaling at the subtask level. They identify 375 subtask-model pairs where smaller models outperform larger ones, spanning 4,342 unique questions, with 13 subtask types showing this pattern across multiple families.

Translation bypasses safety alignment in low-resource languages across multiple LLMs

The authors test whether English safety alignment transfers to low-resource languages by translating harmful prompts that models refuse in English. They evaluate 10 LLMs across 78 low-resource languages from 20 language families, using a conditional metric that isolates translation bypass from existing English unsafety. Mean bypass rates range from 21.9% to 44.4%, risk concentrates in specific language families, and models do not maintain a stable ranking across conditioning schemes.

21 Oct·1 min

Compaction tuned for one query fails to aid later queries

The authors evaluate context compaction methods on later, unseen questions across four multi-query corpora. They find that query-aware sentence selection, which keeps sentences relevant to the current question, provides no measurable improvement over random deletion for subsequent questions. Meanwhile, word-level compressors such as LLMLingua-2 achieve significant gains on later questions by retaining more content words.

Undergraduate research problems remain difficult for formal coding agents

Zheng and colleagues introduce ReuF2F, a benchmark designed to test whether coding agents can settle entry-level open research problems with machine-checked proofs. Using a shared agent harness and independent verification, the best evaluated model solved 42.42% on its first attempt and 109 of 244 problems across two attempts.