AI Papers · Daily Picks

VBBR-Pro introduces a closed-loop testbed that enables scalable, verifiable, and controllable native visual reasoning through generation across diverse visual substrates.

2608.26105 · ▲ 254 · Code

Apodex 1.1 improves sustained, verifiable progress on complex real-world tasks by scaling executable environments and training agents to coordinate long-horizon work with state maintenance and recovery.

2608.23283 · ▲ 202 · Code

Game engines provide executable verification and long-horizon trajectories for reinforcement learning post-training of spatial world models, motivating a human-engine verification paradigm.

2608.25518 · ▲ 180

VGI-bench evaluates visual reasoning in video generation models through 27 tasks, revealing limited reliability and self-correction capabilities.

2608.19583 · ▲ 173 · Code

VoiceMem introduces a dual-brain streaming memory architecture for speech language models that improves retrieval accuracy, emotional personalization, and real-time efficiency.

2608.26005 · ▲ 168 · Code

FrontierChallenge evaluates end-to-end scientific workflows across domains, revealing that frontier models complete only about 20% of tasks despite high partial scores and frequent claims of completion.

2608.24979 · ▲ 141

Off-policy reinforcement learning stabilizers vary with data availability, motivating regime-aware algorithms that adapt normalization and Q-function clipping to improve efficiency across CPU and GPU-parallel training.

2608.24479 · ▲ 137 · Code

The study formalizes probabilistic alignment for world models, introduces PAWBench and PAWEval to evaluate video generators as stochastic samplers, and finds current models fail to match reference behavior distributions.

2608.27345 · ▲ 129 · Code

JIT-Agent is a trainable model that synthesizes adaptive agent harnesses for off-the-shelf LLMs, improving performance across diverse models and tasks.

2608.25593 · ▲ 109 · Code

OraRL improves reinforcement learning post-training for video multimodal language models by integrating oracle rollouts with decoupled advantage estimation and sign-balanced pruning, achieving higher sample efficiency and scalability without chain-of-thought generation.

2608.20492 · ▲ 108 · Code

GigaBrain-0.7 is a vision-language-action model that improves embodied generalization via a three-system architecture, large-scale heterogeneous pretraining, and joint alignment training.

2608.15875 · ▲ 102 · Code

UrbanGround evaluates whether multimodal language model agents can sustain reliable navigation and spatial reasoning in a realistic 3D city replica, revealing that local perceptual skills fail to compose into extended goal-directed behavior.

2608.27456 · ▲ 100 · Code

EchoWM is an omnimodal world model that generates synchronized high-resolution video, sound, music, and speech while following continuous 6-DoF navigation trajectories across first- and third-person views.

2608.23189 · ▲ 78

Test-Time Policy Optimization enables label-free test-time training for mathematical reasoning by asymmetrically distilling agreeing rollouts and penalizing disagreeing ones, matching supervised performance.

2608.27448 · ▲ 73 · Code

Self-OPD eliminates task-specific teachers in flow matching by using self-explored stochastic branches and normalized advantages to optimize the velocity field for multi-objective alignment.

2608.26872 · ▲ 71

WeMM-Embedding is a family of universal multimodal embedding models that align text, images, videos, and interleaved inputs in a shared space, achieving state-of-the-art retrieval and recommendation performance across public benchmarks and large-scale WeChat applications.

2608.24053 · ▲ 67 · Code

DiffusionOPSD uses on-policy self-distillation to turn image-level rewards into explicit intermediate targets for diffusion models, improving alignment efficiency and enabling separate analysis of target construction and policy fitting.

2608.24646 · ▲ 66 · Code

AutoSaddler automatically improves LLM agent harnesses via offline failure-driven optimization, boosting performance on long-horizon benchmarks.

2608.23041 · ▲ 62 · Code

Agentic data generation is framed as constrained distribution design over factorized experience tuples, emphasizing execution-grounded accuracy, learner-relative complexity, and diversity rather than scale alone.

2608.27260 · ▲ 61

TLive-Omni is an omni-modal model for live-commerce that unifies image, video, audio, and text via timestamped token grouping, staged supervised training, and reinforcement fine-tuning with verifiable feedback to enable accurate real-time understanding.

2608.20958 · ▲ 58 · Code

A hierarchical taxonomy and dense supervision strategy improve diffusion-based image editing through fine-grained concepts, large-scale paired data, and granular evaluation.

2608.16812 · ▲ 50 · Code

Prime Agent is an open-source harness that uses recursive subagents, persistent computation, and agent-to-agent coordination to extend language models' long-horizon capabilities across coding and reasoning tasks.

2608.23552 · ▲ 46 · Code

Harness-Aware Training enables compact models to adapt to evolving digital-avatar harness configurations with low latency and high accuracy.

2608.15763 · ▲ 46

GameWAM is a unified world-action model for native video-game control that jointly predicts future visuals and executable keyboard-mouse actions using block-causal flow matching, mode-specific distributions, and block-cycle replanning.

2608.26200 · ▲ 43 · Code

SecOPD improves defense against adaptive prompt injection by using token-level feedback during fine-tuning, sharply reducing attack success rates on language models.

2608.21500 · ▲ 41 · Code

MobilePA-Bench is an interactive sandbox benchmark that evaluates mobile planning agents on tool-calling, sub-agent collaboration, memory usage, and composite skill invocation under real runtime constraints.

2608.23035 · ▲ 41 · Code

Block3D accelerates text-to-3D generation by using block-wise diffusion with confidence-guided correction to reduce inference time while preserving geometric fidelity.

2608.19567 · ▲ 32 · Code

Recuris introduces a recursive memory architecture that tracks progress and guides skill selection to improve long-horizon agent success through localized, validation-gated updates.

2608.24876 · ▲ 26 · Code

RISE adaptively decides when to continue or stop imagination rollouts for planning by weighing expected benefit against cost, supported by a counterfactual driving dataset with expert annotations.

2608.20430 · ▲ 26 · Code

Agent-G² models hint depth as a Gaussian distribution estimated online from existing rollouts, improving reinforcement learning on long-horizon tasks without extra probing.

2608.23318 · ▲ 26 · Code

ReWorld separates short-horizon control and long-horizon memory during training, then bounds both at inference via mixed attention windows, a pose-indexed landmark bank, and distribution-matching LoRA distillation to enable real-time interactive world modeling with strong action fidelity and long-range recall.

2608.23565 · ▲ 24 · Code

Multi-teacher on-policy distillation suffers from token-level budget misallocation across domains, which is addressed by balancing, dynamic allocation, and reward refresh to recover most of the oracle ensemble's capability.

2608.19098 · ▲ 20 · Code

JoyAI-Echo-1.5 unifies long-form video and interactive world generation through cross-shot memory, geometry-aware camera control, and rollout-aware training to maintain identity and coherence over extended sequences.

2608.23383 · ▲ 19 · Code

Zero-WAM enables robotic manipulation of unseen tasks by conditioning a causal video-action model on in-context human video guidance, supported by an automatically generated dataset and a future-chunk prediction objective.

2608.26103 · ▲ 19 · Code

StreamPI enhances single-frame vision-language-action models with streaming temporal reasoning via instruction-anchored attention and randomized interval training, improving robot manipulation without extra parameters.

2608.26067 · ▲ 18 · Code

Visual Rubrics-Based Reinforcement Learning improves vision-language model grounding by scoring answers on visual faithfulness, reasoning consistency, and instruction following using structured partial credit.

2608.25580 · ▲ 15 · Code

LAION-BVD is a large-scale open video dataset enabling multimodal pre-training across video, audio, and image modalities with synthetic captions and strong benchmark performance.

2608.24845 · ▲ 14 · Code

The study introduces a benchmark for evaluating autonomous software migration by coding agents, finding that current models rarely complete migrations correctly.

2608.23564 · ▲ 14 · Code

Procedura is a 3D modeling agent that generates editable, part-structured procedural assemblies with sharp geometry and validated articulation from text prompts.

2608.26238 · ▲ 11 · Code

AutoResearch is a two-stage autonomous system that grounds research ideas through integrated generation and evidence-based execution to improve experimental reliability and measurable outcomes.

2608.17906 · ▲ 10 · Code

Agent skills package specialized knowledge and workflows into reusable resources that extend AI agent capabilities. Recent work automatically discovers such skills from agent experience, which enables agents to progressively adapt through interaction. However, the insights that guide skill development typically remain scattered across optimization histories, limiting their systematic reuse across iterations. We introduce WikiSkill, a framework that co-evolves agent skills with a persistent knowledge base (wiki). At a high level, WikiSkill separates raw execution experience, accumulated knowledge, and executable skills, while continuously consolidating experience into the wiki, which subsequent skill updates can build on. Across diverse benchmarks and models, WikiSkill consistently outperforms state-of-the-art skill-evolution methods and improves over no-skill baselines in most model-benchmark settings. We find that skill evolution complements model scaling: larger models generally benefit more from evolved skills, while smaller models with skills can outperform substantially larger models without them. We also find that evolved skills transfer effectively across models and model families, and skills evolved by other models can outperform self-evolved skills. Finally, our ablation studies confirm that persistent knowledge accumulation in the wiki is critical for effective skill evolution. These results demonstrate the benefits of systematically accumulating and refining agent experience for developing reusable and transferable skills.

2608.27454v1

To improve large language models' ability to resolve real-world software issues, prior work has focused on constructing large-scale agent trajectory datasets and performing supervised fine-tuning (SFT) on successful trajectories. However, task success does not guarantee high-quality supervision: successful trajectories may still contain ineffective, redundant, or risky steps. Directly using such trajectories for SFT can introduce noisy supervision and encourage models to imitate undesirable problem-solving behaviors. Therefore, we propose SWE-Prime, a multi-granularity, two-stage SFT data selection method that progressively filters training data at the trajectory and segment levels. Specifically, the first stage performs trajectory-level screening based on process quality, result quality, and data representativeness, selecting a high-quality and representative subset of successful trajectories. The second stage performs segment-level selection by grouping consecutive steps into semantic segments and assessing each segment based on its contribution to the final solution, learnability, and potential risks. During SFT, all segments remain in the sequence to preserve context, while only selected segments contribute to the loss computation. Experiments on SWE-Bench Pro and SWE-Bench Verified show that training on the 10% trajectory subset selected by SWE-Prime outperforms training on the full resolved dataset, yielding relative performance gains of up to 12.2% and 24.2%, respectively.

2608.27449v1

In real-world software development, code review typically involves iterative interactions between developers and reviewers to improve software quality, making the process costly and time-consuming. Although recent work explores large language models (LLMs) for automated code review, most approaches oversimplify code review into a single-round, static decision task, which fails to capture the multi-round interactive nature and the complex problem-solving processes inherent in realistic review scenarios. To bridge this gap, we introduce MCR-Bench, the first defect state-aware benchmark designed for realistic multi-round code review. MCR-Bench covers five commonly-used programming languages and consists of 2,269 real-world multi-round code review tasks, each of which is annotated with fine-grained defect information and cross-round state labels. Each task in MCR-Bench is equipped with fine-grained defect metadata (e.g., description, type, severity) alongside dynamic state annotations, capturing the complete evolutionary trajectory of a defect throughout the multi-round process. We obtain several findings through extensive experiments on MCR-Bench with mainstream LLMs. (1) Limited overall capability: experiments reveal that mainstream LLMs exhibit limited overall performance in defect detection and defect lifecycle state tracking, with performance degrading significantly as the number of interaction rounds increases; (2) Defect-sensitive performance: LLMs' performance varies substantially across different defect types and severity levels, with semantically complex or low-salience defects being significantly more likely to be missed; (3) Underlying Failure Mechanisms: our in-depth error analysis dissects the distinct drivers of false positives and false negatives, revealing critical weaknesses such as cross-round temporal misalignment and inadequate long-range memory.

2608.27442v1

LLM-based agents are increasingly deployed in product-level execution harnesses, where jailbreaks can trigger harmful tool use and persistent state changes, creating greater risks than unsafe text generation alone. Existing automatic red-teaming methods often rely on fixed attacks, while recent agentic attackers coordinate multiple jailbreak tools and show stronger potential through trajectory-based retrieval. However, such retrieval can reuse misleading experiences due to retrieval bias and unclear tool credit, and full trajectories add context overhead while reducing interpretability. We propose RedEvoAgent, a black-box red-teaming agent that distills cross-case attack trajectories into a concise, human-readable attack skill. The attack skill adaptively evolves through tool-effectiveness profiling and Deciding-Tool Attribution for skill updates, and a validation ratchet that retains only updates improving validation performance. Experiments on multiple benchmarks, target models, and target execution harnesses show that RedEvoAgent outperforms fixed and agentic baselines, improves tool efficiency, and transfers across attacker models and target execution harnesses.

2608.27439v1

Recent advances in inference-time scaling have significantly improved the reasoning performance of large language models (LLMs). However, these methods typically rely on repeated generation or external verification. To address this limitation, we introduce CritICL, a novel inference-time framework that improves reasoning while maintaining high efficiency. Our key insight is that LLM failure modes exhibit structured patterns across model scales within the same family. Instead of treating failures as undesirable outputs, CritICL leverages them as a source of guidance. Specifically, we utilize failure modes derived from weaker models and incorporate them into inference through critique-based in-context examples. We propose two variants: CritICL-dynamic, which adaptively predicts input-specific failure modes and retrieves critiques, and CritICL-static, which uses a global failure mode profile to provide stable guidance. Experimental results show that CritICL consistently outperforms standard in-context learning and achieves performance competitive with or superior to test-time scaling methods, while requiring significantly fewer generations and lower token cost. Code available at: https://github.com/umwyf/CRITICL

2608.27455v1

Recent prominent post-training methods, such as Reinforcement Learning (RL) and On-Policy Self-Distillation (OPSD), have driven rapid progress in mathematical reasoning for large language models, yet their reliance on ground-truth labels precludes test-time training (TTT). Replacing ground truth with majority-vote pseudo-labels is a natural alternative, yet it is fragile: an incorrect vote corrupts the teacher and misleads every token. We observe that this failure mode is asymmetric: rollouts that disagree with the pseudo-label are typically wrong regardless of whether the vote itself is correct. Building on this observation, we propose Test-Time Policy Optimization (TTPO), an asymmetric objective that distills agreeing rollouts via OPSD and penalizes disagreeing rollouts with Grouped RL. Token-level selection further refines both branches: distillation down-weights already-converged positions, while RL penalizes only confident errors. Both updates remain well-grounded even under frequent pseudo-label errors, and majority-vote routing yields tighter self-supervision as the model improves. Without any labels, TTPO matches label-supervised OPSD on five competition-level benchmarks, raises Qwen3-1.7B from 38.0% to 45.2% in TTT, yields +25.2% to +36.4% without thinking, and shows strong cross-task generalization.

2608.27448v1

Transduced language models (TLMs) compose a pretrained \emph{source} language model with a functional finite-state transducer to induce a language model over \emph{target} strings. Computing the probability of a target prefix under a TLM amounts to summing the source-model probabilities of all source strings that the transducer maps to target strings beginning with that prefix. This set can be exponentially large or infinite. Prior work uses a computational shortcut based on source prefix probabilities, then approximates the resulting sum with threshold-pruned beam summing. This produces a lower bound with unknown error. Instead, we resample source prefixes without replacement and reweight each selected prefix by the inverse of its inclusion probability. We show that applying this correction recursively gives an unbiased estimator of the target prefix probability and lets us estimate the mass lost by threshold pruning. Our beam-summing algorithm extends the retained source prefixes and samples which prefixes to keep, reducing their number as more probability mass is added to the running estimate. This can save computation and guarantees that the run halts with probability one. We evaluate the method on encyclopedic text and DNA against sequential Monte Carlo baselines that resample with replacement. It achieves a better compute--variance tradeoff on text and lower error at the same maximum number of particles on DNA. On a DNA-to-amino-acid transduction, it reduces runtime by several orders of magnitude relative to threshold-pruned beam summing and makes estimating prefix probabilities for long target strings feasible. Replacing threshold pruning with unbiased sampling in a published reading-time analysis substantially lowers the estimated corpus surprisal but leaves the published conclusions unchanged.

2608.27428v1

Reinforcement Learning with Verifiable Rewards (RLVR) significantly improves LLM reasoning but often causes a drop in policy entropy, leading to narrowed reasoning coverage and degraded pass@$k$ for large $k$. While existing methods mitigate this entropy collapse through algorithmic regularizations, cross-model non-parametric perturbation is also neglected. In this work, we propose a simple yet effective approach to preserve the generative diversity of LLMs during RLVR. Instead of relying solely on internal exploration, we force the target model to generate answers based on partial reasoning trajectories generated by a smaller, weaker language models. These unfamiliar prefixes effectively disrupt over-confidence and encourage the exploration of distinct reasoning paths. We empirically study the potential of outer prefixes, revealing the mechanism of the impact of distributional discrepancy to the exploration dynamics in RLVR training. Experiments across multiple mathematical benchmarks show that our method consistently outperforms vanilla RLVR. Notably, the performance gain becomes increasingly pronounced as $k$ scales up, demonstrating a substantial expansion of reasoning coverage. Furthermore, our approach efficiently mitigates entropy collapse without requiring additional SFT, intricate reward designs, or complex prompting.

2608.27420v1

Multimodal large language models (MLLMs) can interpret a street view, but urban agency depends on whether such local evidence remains useful after the agent starts to move. In this paper, we investigate how far current MLLM agents can turn local urban perception into reliable action in a complicated real-scale city. We propose UrbanGround, the first sandbox to make this question testable in a physically constrained replica of Hong Kong built from territory-wide 3D geospatial data. UrbanGround supports closed-loop interaction from a first-person view and provides an interactive map for navigation. Agents can directly enter the 3D city and explore from a first-person view. Our analysis follows the growth of the spatial problem through three research questions. We first test whether an agent can ground a local scene well enough to answer spatial questions after active observation. Then we ask whether that grounding supports navigation as destinations become farther away and less explicit. Finally, we examine whether the resulting behavior survives changes in route availability and pedestrian motion. Contemporary MLLM agents usually show useful atomic abilities in visual recognition and short-range spatial reasoning, while orientation and pedestrian-aware movement remain unreliable. Their central failure emerges over extended exploration, where local abilities do not compose into sustained goal-directed behavior and errors accumulate without effective correction. We hope UrbanGround will support broader study of how far current MLLM agents can explore reliably in complex, open-ended urban environments.

2608.27456v1

Vision-language models (VLMs) can locate an image region referred to by a text prompt and route the corresponding visual evidence to the output, yet the internal mechanism behind this behavior is not understood. Inspired by retrieval heads in large language models, we ask whether VLMs contain an analogous mechanism for visual retrieval. We answer affirmatively by introducing Visual Retrieval Heads (VRHs), a small subset of attention heads (about 1.7-2.6%) that are causally responsible for grounding text descriptions to image regions. To find them, we recast existing head-scoring methods under a unified design space over query tokens, key aggregation, and cross-sample aggregation. We then show that scoring attention from output prediction tokens with a sum over the ground-truth referent region most reliably identifies causal heads. Across eleven VLMs and five referring-expression benchmarks, masking only the top 20 VRHs reduces grounding accuracy by up to 80 percentage points, while masking the same number of random heads has little effect. Beyond replicating the causal-sparse-universal triad established for text retrieval heads, VRHs exhibit several properties not previously reported: they generalize across visual reference tasks, remaining causal on attribute, spatial, counting, and visual-math benchmarks despite being discovered through bounding-box prediction; they are functionally specific, preserving output format while corrupting localization; and they are architecturally shared, transferring causally across VLMs that share an LLM backbone but differ in vision encoder, projector, and instruction tuning.

2608.27417v1