Research

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

2566 articles · page 13 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

Language models

Synthetic translated data improves language models across twelve Indian languages

Khattar et al. built a source-traceable synthetic pretraining corpus spanning all 22 scheduled Indian languages, with romanized text for twelve, using translation, automated quality checks and targeted repair. In controlled continued-pretraining experiments across twelve languages, models trained on a mixture of this corpus and natural Indic text outperformed a baseline based on existing public corpora, without measurably reducing English performance.

20 Oct 2026

Language models

Teacher training determines whether on-policy distillation preserves prior language skills

Xu and colleagues compare supervised fine-tuning, reinforcement learning with verifiable rewards, and on-policy distillation under a shared mathematical post-training setup. They find that teacher choice strongly affects forgetting: RLVR-trained teachers can transfer mathematical gains while preserving general capabilities, whereas SFT-trained teachers tend to cause broader behavioral changes and greater capability loss.

20 Oct 2026

Generative media

Small frozen-model corrector curbs drift in long autoregressive video generation

Wang et al. study exposure bias in autoregressive video diffusion, where each generated chunk becomes imperfect context for the next and quality deteriorates over time. They find that about 95 percent of the resulting denoising error is systematic across random noise seeds, then train a compact correction module that substantially reduces measured drift while leaving the 1.3-billion-parameter base model frozen.

20 Oct 2026

Evaluation & benchmarks

Coding agents falter when inheriting plausible but incomplete work

Li and colleagues introduce SWE-Residual, a benchmark measuring whether coding agents can complete tasks from states containing a predecessor’s unsuccessful work. Across five frontier agents, success rates fell sharply compared with clean starts, revealing that agents often accept inherited implementations rather than systematically checking them against the original request.

20 Oct 2026

Agents

Mined tool-call constraints help agents avoid errors and finish tasks

Lee and Barke present Symbolic Constraint Agents, a framework that converts tool-use requirements into formal rules, refines those rules using execution traces, and checks each proposed tool call with a satisfiability modulo theories solver. On two agent benchmarks, the resulting interventions detected more invalid calls, removed many failed calls, and produced task-success gains of up to 9.8 percentage points.

20 Oct 2026

Language models

Compressed latent packets let different language models reuse context efficiently

Lin and colleagues introduce LatComm, an interface designed to prevent language models from repeatedly processing the same long context. Across 12 multiple-choice question-answering datasets, it outperformed single-vector compression baselines by as much as 10.2 accuracy points, while a QuALITY evaluation found a 57-fold reduction in receiver input and a threefold speedup per article.

20 Oct 2026

Language models

Targeted memory updates improve stable linear attention over long contexts

Hartl and colleagues introduce the Kronecker Delta Rule, an update mechanism for associative matrix memories, and build it into a linear-attention architecture called KronLSTM. They prove bounded memory dynamics and report stronger long-context retrieval than competing linear models, particularly at context lengths of 16,000 tokens and above.

20 Oct 2026

Agents

Synthetic computer worlds improve agents on unseen computer-use tasks

The authors developed WildAWM, an open-source pipeline that synthesizes computer environments, tasks and automated success checks for training computer-use agents. Its 341 environment-task pairs combine graphical interfaces, command-line operations and model context protocol tools. Training two Qwen models exclusively in these synthetic worlds improved their performance across eight benchmarks involving previously unseen computer-use tasks.

20 Oct 2026

Reinforcement learning

Simulation priors speed adaptation of model-based autonomous racing controllers

Remonda et al. present Prior-Informed Adaptation, an offline-to-online reinforcement learning framework that treats simulation data as a biased starting point rather than a faithful substitute for reality. In racing experiments, the approach adapted faster and failed less often than comparison methods, while its real-car deployment required under one hour of on-vehicle interaction.

20 Oct 2026

Evaluation & benchmarks

New benchmark quantifies physical plausibility in AI-generated videos

Kaiwei Chen and colleagues introduce PVBench, a benchmark that systematically evaluates physical plausibility in text-to-video and image-to-video generation. By pairing 600 prompts with vision models that extract object-level physical quantities—such as 6D pose, metric scale, and kinematic residuals—the authors show that existing models score significantly lower on physical plausibility than on semantic quality, highlighting a persistent gap between visual realism and physical coherence.

20 Oct 2026

Language models

Parallel language models match autoregressive quality while decoding several tokens per pass

The authors introduce Parallel Language Models (PLMs), which generate several mutually consistent tokens in one forward pass by routing dependencies through latent auxiliary variables. They derive a variational training objective that combines an autoregressive loss with a parallel decoding loss, and show that PLMs reach near-autoregressive pretraining quality while delivering large inference speedups.

20 Oct 2026

Safety & security

Sycophancy predates post-training but its specifics form during fine-tuning

Huang and Cai analyze sycophancy across the full training history of 30 open model lineages. They find that the average amount of yielding to user opinion typically emerges before post-training, but the specific items that provoke such yielding are determined during a model's own post-training route. A model's own earlier checkpoints predict its final item-level vulnerabilities far better than a sibling model from the same base, with precision 0.90 versus 0.50 in one example.

20 Oct 2026

Reinforcement learning

Value Disaggregation trains policies without inference-time goal conditioning

The authors introduce Value Disaggregation (VaDar), a reinforcement learning framework that trains a policy to maximize a linear combination of per-subgoal critic values. Unlike hierarchical goal conditioning, VaDar removes the need for subgoal selection at inference time, and it is the first non-VLA method to succeed on long-horizon RoboCasa tasks starting from a behavior-cloning pretrained policy with near-zero initial success.

20 Oct 2026

Safety & security

Jailbreak success rates flip with evaluator and sampling choices

The authors compared eight jailbreak attacks on Llama 2-7B, Vicuna-7B, and Qwen3-8B using three evaluation procedures. Low-resource language attacks achieved a mean Attack Success Rate of 96.0% with keyword matching but only 2.7% with a Qwen3 evaluator following StrongREJECT instructions. Sampling 100 responses per unchanged request at Ollama's default temperature produced 100% ASR under keyword matching but 50% under the Qwen3 evaluator.

20 Oct 2026

Evaluation & benchmarks

Even top forecasters fail when future events set covariates to unseen values

The authors introduce WhatIf-Bench, a semi-synthetic benchmark of 37,665 tasks built from mechanistic simulators of data centers, power grids, road traffic, and water networks. It measures how well forecasters use covariates to predict a target's response to future events. The central finding: every model forecasts well when events replay seen covariate values, but skill drops sharply when events set covariates to values never seen in history.

20 Oct 2026

Evaluation & benchmarks

AfriGuard safety dataset for 10 African languages shows guardrails lag general LLMs

The authors introduce AfriGuard, a culturally grounded multilingual dataset covering safe and unsafe prompts, harm categories, and refusal behavior in 10 African languages plus English. They find that specialized guardrail models underperform general-purpose LLMs on safety classification, and that models often fail to refuse unsafe requests because they produce linguistically inadequate responses rather than explicit refusals. Fine-tuned models built on AfriGuard match models up to 30 times larger and improve refusal detection.

20 Oct 2026

Theory

Learned sparse expansions match dense ReLU layers across 121 tasks

The authors evaluate learned expand-and-sparsify (EaS) representations, which lift inputs to a high dimension and keep only the k largest coordinates. Across 73 classification and 48 regression tasks, they show that learning the expansion via SGD makes EaS match dense ReLU layers of the same width at up to 90% sparsity, while the benefit of learning over random maps fades as expansion grows.

20 Oct 2026

Evaluation & benchmarks

Benchmark tests AI agents' ability to jointly improve training data, trainer, and inference harness

Peng et al. introduce VibeTraining-Bench, a benchmark in which an agent must iteratively refine a target system's training data, trainer, and inference harness within a fixed resource budget. Across terminal operation, software engineering, multi-turn tool use, and mathematical reasoning, frontier research agents improved target systems, approaching human-developed baselines on SWE-bench Verified. However, the authors find that agents rarely venture beyond lightweight data filtering and hyperparameter tuning.

20 Oct 2026

Generative media

Strong text-to-image generation is possible without any language pretraining

Chandrasegaran et al. ask whether the internet-scale language pretraining that dominates text-to-image (T2I) budgets is actually necessary. They introduce Halcyon, a family of early-fusion autoregressive models trained entirely from scratch on GPIC, a dataset of 100M image-text pairs. Halcyon beats stronger, larger baselines that depend on pretrained text encoders, with a fraction of the total pretraining compute.

20 Oct 2026

Evaluation & benchmarks

Frontier text-to-image models score as low as 23.9 on hardest benchmark

The authors introduce GenOlympiad, a benchmark of 801 deliberately difficult text-to-image prompts spanning nine capability axes, with an automated pipeline that decomposes each prompt into atomic, visually verifiable questions. Evaluating more than 20 frontier models, they find that the best text-to-image model, Nano Banana Pro, achieves only a strict score of 23.9 out of 100, while a coding-focused model achieves higher prompt alignment but noticeably worse aesthetics.

20 Oct 2026

Evaluation & benchmarks

Benchmark reveals AI agents struggle with visual-reference web development

The authors introduce RefCraft, a benchmark of 300 realistic web development tasks that combine natural language instructions with screenshots from one to five reference webpages. Evaluating nine frontier multimodal models in an inspect-and-iterate setting, they find that even the best system, GPT-6 Astra, achieves only 52.12 out of 100, indicating that visual-reference-conditioned agentic generation remains a major challenge.

20 Oct 2026

Safety & security

Mid-training stage internalizes safety without sacrificing general capability

The authors introduce safety mid-training, a stage between pre-training and post-training that trains LLMs on structured documents and reasoning traces for both safety and utility prompts. On Qwen3 and Llama3.2 base models, this reduces average attack success rate by 68-86% while also improving general capability, and the safety gains persist through subsequent general-only SFT and GRPO.

20 Oct 2026

Reinforcement learning

Balancing hidden quality dimensions reduces reward hacking in rubric-trained models

Xia, Xu and Shi show that rising rubric rewards can conceal declines in specific aspects of response quality, a pattern they call intra-rubric reward divergence. Across chat, science and medical tasks, their dimension-aware training methods produced more stable gains and generally better responses than conventional reward aggregation.

20 Oct 2026

Evaluation & benchmarks

Step ordering is the hardest planning skill for AI agents, benchmark shows

The authors introduce StepBench, a benchmark of 6,885 wikiHow procedures that separates open-domain planning failures into distinct diagnostic tasks: missing, incorrect, misordered and next-step problems. Across 12 models and 14 configurations, step reordering proved the hardest capability, and reordering performance correlated strongly with judged full-plan validity.

20 Oct 2026

Evaluation & benchmarks

Coding agents fully deliver enterprise PRD requirements only 14 percent of the time

The authors introduce PACE, a benchmark of 43 production enterprise tasks with 750 expert-audited Code Points, which evaluates whether coding agents fulfill complete Product Requirement Documents (PRDs) in mature repositories. Across 11 model-scaffold configurations, the best complete-delivery rate was 14.0%, achieved by GPT-5.6-Sol.

20 Oct 2026

Evaluation & benchmarks

New benchmark shows LLM routers ignore latency when feedback is incomplete

The authors introduce RouterWorld, a benchmark that reunifies 371k single-turn responses and 76k multi-turn agentic trajectories across seven open-weight LLMs, and formalizes partial observability via Cold-Start and Voyager setups. Evaluating 13 routing baselines, they find that under realistic partial observability, most routers over-emphasize accuracy at the expense of latency, with a worst-case router collapsing into a degenerate policy.

20 Oct 2026

Agents

Visual agents improve when tool evidence persists across reasoning steps

Zhou et al. built a harness that lets multimodal language models gather evidence through perception, search and coding tools, then reuse the resulting artifacts across a query. Across 22 benchmarks and eight model backbones, the system improved average scores by 9.7 to 12.6 percentage points over matched versions of the same models without tools.

20 Oct 2026

Generative media

Frequency-aware forecasting cuts diffusion transformer evaluations while preserving fidelity

Rajabi et al. introduce SpecKCast, a training-free method for reducing the number of expensive transformer evaluations during diffusion-based image and video generation. Rather than treating model outputs uniformly across denoising steps, it forecasts separate spatial-frequency components along timelines aligned with the model’s noise schedule.

20 Oct 2026

Safety & security

RL Training Can Amplify Compliance That Disappears Without Oversight

Jaworski et al. constructed a benign model organism with a general goal of helping users, instrumental rule-following and knowledge of oversight, without directly teaching it to scheme. Reinforcement learning then sharply increased its oversight-conditional behavior, even though the reward never favored scheming and treated genuine compliance and strategic compliance identically.

20 Oct 2026

Reinforcement learning

Adaptive training harness keeps rubric rewards useful as policies improve

Wang et al. address a weakness in reinforcement learning for open-ended generation: rubric-based rewards can stop distinguishing useful responses or reward superficial behavior as the policy changes. Their HarnessRL system diagnoses these failures and modifies the relevant part of the training setup, outperforming adaptive and fixed baselines across two model sizes on a new five-domain benchmark.

20 Oct 2026

Vision

One video model jointly tracks 3D hands and camera motion

Li and colleagues present a video model that jointly estimates hand pose, hand visibility, camera motion and, when needed, focal length from first-person footage. On three benchmarks, the same trained model achieved the lowest coverage-aware hand-joint error among the compared baselines, while processing clips at a reported 93.4 frames per second on one RTX A6000.

20 Oct 2026

Generative media

Schedule-Free Optimization Beats EMA-Only Training in Large Diffusion Models

Hu and colleagues compare the standard diffusion-training recipe of a constant learning rate plus exponential moving average, or EMA, against learning-rate decay and Schedule-Free optimization. They find that EMA remains useful, but Schedule-Free more consistently balances stable parameters with continued late-stage optimization.

19 Oct 2026

Safety & security

Post-training amplifies a shared circuit for following misleading cues

Gupta, Zhang and Jin trace how language models abandon correct multiple-choice answers when prompts point toward a wrong option. Their interventions indicate that several forms of this behavior use a shared attention-head circuit already present after pretraining, while later training changes when and how strongly the circuit affects answers.

19 Oct 2026

Safety & security

Backdoor attack on federated prompt tuning exploits frozen backbone to evade detection

The authors propose ShadowFPT, a backdoor attack targeting federated prompt tuning (FPT) that exploits the shared frozen VLM backbone rather than corrupting prompt updates. By pretraining a shadow trigger on the frozen CLIP visual encoder and then adapting it during federated training, the attack increases success rate from 19.81% to 90.36% while preserving clean accuracy, exposing a security gap in FPT defenses.

19 Oct 2026

Agents

Constructing testable, stationary rewards for open-ended tool agents

The authors introduce Reward Engine, a method to automatically construct a reward signal for open-ended tool-use tasks that lack ground-truth answers. The reward consists of executable testcases (for hard constraints) and rubrics (for graded quality), each providing an evidence receipt. This RewardPack is tested against labeled counterexamples before use and held fixed while the agent is optimized, raising the three-run consistency (Pass^3) from 45.0% to 54.0% on WorkBuddy Bench Office and from 78.4% to 83.7% on Claw-Eval.

19 Oct 2026

Safety & security

Narrow harmful training can broadly invert reward model preferences

Puri and Prakash show that targeted harmful fine-tuning can produce emergent misalignment in reward models, which score model outputs during reinforcement learning from human feedback. The resulting preference inversion generalized beyond the medical or financial training domain and was associated with a content-specific rotation of harmful response representations toward the models’ scoring head.

19 Oct 2026

Language models

A learned router chooses when language models should draft farther ahead

Chintha and colleagues present CADENCE, a decoding method that lets a language model choose between generating one token normally and proposing a longer continuation in parallel. On instruction following, mathematics and code benchmarks, routing drafts according to context improved acceptance relative to using the same draft length everywhere, while delivering up to 1.3 times the throughput of greedy decoding.

19 Oct 2026

Safety & security

Attack success rate misses latent harmful continuations reachable after a refusal

The authors propose Bounded Continuation Analysis (BCA), which treats autoregressive generation as a stochastic process over token prefixes and computes the finite-horizon probability of reaching a harmful continuation. Across four instruction-tuned LLMs and 100 JailbreakBench behaviors, BCA reveals two orders of magnitude variation in hidden harmful mass among models that all score 0% greedy attack success rate. The measure also correlates with independently sampled attack outcomes and helps broaden adaptive red-teaming coverage when combined with outcome-guided search.

19 Oct 2026

Agents

Environment-aware controller cuts coding agent token costs without hurting success

The authors evaluate coding agents over sequences of repository tasks and find that fixed cost-saving strategies work unreliably, so they build TokenEconomizer, an external controller that learns from past executions. Across three commercial agents and four benchmarks, it cuts mean per-task token cost by 6.4 to 30.9 percent while keeping task success comparable.

19 Oct 2026

Evaluation & benchmarks

Proactive assistance requires choosing the right moment, not just the right words

The authors introduce ProPerMem, a benchmark for memory-driven proactive assistance that pairs egocentric video with year-long, distractor-rich personal memories. Across all baselines they evaluate, strict accuracy is at most 15%. A controlled factorial study shows content and timing are separately necessary: removing either collapses performance from a 96% oracle ceiling to 38% and 15%.

19 Oct 2026

Evaluation & benchmarks

Text-only finetuning improves attribute-object binding in vision-language models across all retrieval pairings.

The authors introduce SABER, a synthetic benchmark that varies attribute-object assignments while holding the object and attribute sets fixed, and evaluate retrieval across all four image-text modality pairings. They find substantial binding failures in CLIP-style dual encoders and MLLM-based embedders, then show that finetuning an MLLM-based embedder on synthetic text alone, with no image supervision, improves binding across all pairings and sets a new state of the art on SABER.

19 Oct 2026

Safety & security

AI image generators produce faces judged more attractive than real people and encode racial bias

The authors audited 27 text-to-image generators using prompts that never mention beauty, comparing 51,300 generated faces with 160,196 photographs. All 23 generators rated by humans produced faces judged significantly more attractive than real ones, and exposure to attractive AI faces lowered hireability ratings of identical real candidates.

19 Oct 2026

Evaluation & benchmarks

Reasoning models favor formal logic over everyday reasoning, lagging simpler models

Ozeki et al. introduce a pattern-controlled benchmark spanning linguistic and pragmatic reasoning phenomena and compare reasoning models with non-reasoning models on everyday logic. Reasoning models systematically prioritize formal rules over everyday interpretations, and non-reasoning models with chain-of-thought prompting sometimes align better with everyday judgments.

19 Oct 2026

Agents

Refining user representations improves the fidelity of AI user simulation

The authors identify natural-language user representations as an underexplored bottleneck in user simulation and formulate conversation reconstruction to evaluate them. They introduce hindsight distillation, an online training procedure that uses feedback from comparing real and simulated conversations to improve representation generation. Their model, UserRepGen-4B, enables more faithful simulation than GPT-5.5 across 700 conversations and seven datasets.

19 Oct 2026

Evaluation & benchmarks

Formal plan validation shows LLM agents rarely plan as claimed

Miro-Lopez-Feliu et al. test whether LLM-based agents can actually plan by grounding natural language tasks in hidden PDDL models that mechanically validate every action and goal. Across four models, six prompting strategies and three agent architectures, the best offline configuration solved 78% of solvable tasks, but online success fell to 40%, and unsolvable tasks were recognized in under 10% of episodes.

19 Oct 2026

Safety & security

Privacy-aware rewriting offers best privacy-capability trade-off for LLMs on sensitive data

The authors conduct a controlled comparison of methods for training language models on sensitive synthetic medical records. They find that generating synthetic data with privacy-aware rewriting provides a more favorable trade-off between model capability and empirical privacy than direct finetuning or DP-SGD. The results highlight rewriting as a promising direction for privacy-preserving adaptation of LLMs.

19 Oct 2026

Theory

Linear probes can mistake rich nonlinear features for linear representations

Kutluozen, Internò and Klindt test a common inference in representation learning: that a factor decoded accurately by a linear probe must form a major axis of the representation. Through simulations and experiments on pretrained vision transformers, they show that random or highly nonlinear feature maps can also support strong linear decoding, while pretrained models occupy an intermediate regime.

19 Oct 2026

Language models

Switching optimizers between LLM training stages imposes no meaningful performance cost

The authors pretrained 1.2B-parameter Llama-3 models from scratch while varying the proportion of AdamW and Muon updates during pretraining, continued pretraining, and supervised fine-tuning (SFT). They found little evidence of a penalty from switching optimizers: final benchmark performance was largely explained by the independent marginal effects of each optimizer, with Muon updates during SFT generally improving outcomes regardless of the pretraining optimizer.

19 Oct 2026