AI papers — 2026-05-28

Authors: Fangfu Liu, Kai He, Tianchang Shen, Tianshi Cao, Sanja Fidler, Yueqi Duan, Jun Gao, Igor Gilitschenski, Zian Wang, Xuanchi Ren

Org: NVIDIA

Problem: Existing interactive video generation world models are limited to two-player settings and cannot scale to arbitrary numbers of agents in a permutation-symmetric way.

Method: Gamma-World introduces two architectural components layered on top of a video diffusion transformer: Simplex Rotary Agent Encoding (SRAE), a parameter-free extension of 3D RoPE that places agents at the vertices of a regular simplex in rotary-angle space to enforce permutation-equivalence, and Sparse Hub Attention (SHA), which reduces cross-agent attention complexity from O(n²) to O(n) via learnable hub tokens that mediate inter-agent communication.

Key results:

• Real-time rollout at 24 FPS via causal student generation
• Generalizes from two to four players without additional training
• Improvements in video fidelity, action controllability, and inter-agent consistency over baseline approaches (no specific numeric deltas stated)

Why it matters / caveats: Scalable multi-agent world models are a prerequisite for training embodied AI in rich social and cooperative environments; the O(n) attention scaling is a meaningful engineering advance. Results are reported primarily qualitatively, and generalization beyond four players is not demonstrated.

Authors: Hongru Hou, Tiehua Mei, Denghui Geng, Jinhui Huang, Ao Xu, Hengrui Chen, Jiaqing Liang, Deqing Yang

Org: Fudan University

Problem: Applying reinforcement learning to proactive recommender systems suffers from length-dependent gradient bias that favors path extension and high gradient variance that destabilizes training.

Method: ProRL augments standard policy gradient training with two targeted corrections: Stepwise Reward Centering, which subtracts a per-step baseline to eliminate the length-dependent bias, and Position-Specific Advantage Estimation, which normalizes advantages at each position across trajectories to reduce variance. Both modifications are applied to the policy gradient objective without changing the reward structure.

Key results:

• Significantly outperforms state-of-the-art proactive recommender systems on three real-world datasets (specific numeric margins not stated in abstract)
• Accepted to ICML 2025

Why it matters / caveats: The two corrections are lightweight and could transfer to other sequential RL-for-NLP tasks that share the length-bias problem. Exact accuracy numbers are not reported in the abstract; full results require reading the paper body.

Authors: Minki Kang, Shizhe Diao, Ryo Hachiuma, Sung Ju Hwang, Pavlo Molchanov, Yu-Chiang Frank Wang, Byung-Kwan Lee

Org: NVIDIA

Problem: Standard RL fine-tuning of vision-language models for agentic tasks creates a Thinking-Acting Gap where the model defaults to reasoning tokens and rarely explores tool-use paths, leaving tool-call rollouts nearly always failing.

Method: AXPO (Agent eXplorative Policy Optimization) targets tool-using subgroups by fixing the thinking prefix and resampling only the tool call and its continuation, paired with uncertainty-based prefix selection that identifies which prefixes are problematic and most benefit from targeted exploration. This is applied on top of SFT-initialized models using a GRPO-style reward signal.

Key results:

• Tool use attempted on only ~30% of rollouts under standard RL; all-wrong tool-using rollouts occur ~40% of the time (baseline diagnosis)
• SFT+AXPO outperforms SFT+GRPO by +1.8 pp Pass@1 and +1.8 pp Pass@4 at 8B scale
• 8B model with AXPO surpasses a 32B base model on Pass@4 using 4× fewer parameters

Why it matters / caveats: Demonstrates that targeted resampling of failure-prone action types is more sample-efficient than uniform rollout RL. Gains are measured relative to SFT+GRPO baselines; comparison to other agentic RL methods is not fully detailed in the abstract.

Authors: Haiwen Diao, Jiahao Wang, Penghao Wu, Yuhao Dong, Yuwei Niu, Yue Zhu, Zhongang Cai, Weichen Fan, Linjun Dai, Silei Wu, Xuanyu Zheng, Mingxuan Li, Yuanhan Zhang, Bo Li, Hanming Deng, Huchuan Lu, Quan Wang, Lei Yang, Lewei Lu, Dahua Lin, Ziwei Liu

Org: Not stated

Problem: Current multimodal models rely on separate visual encoders and language models fused post-hoc, preventing end-to-end learning of pixel-level spatiotemporal representations alongside language.

Method: NEO-ov (Native One-Vision) is a single unified transformer trained end-to-end from raw pixels to language tokens, with no external vision encoders, auxiliary adapters, or post-hoc fusion modules. The architecture develops cross-frame and pixel-word correspondence internally, enabling unified spatiotemporal modeling for multi-image and video understanding at scale.

Key results:

• Competitive with modular encoder-decoder systems on standard multimodal benchmarks
• Excels at detailed visual perception tasks compared to modular baselines
• Specific benchmark scores not stated

Why it matters / caveats: Removing the encoder bottleneck is architecturally important for long-horizon video tasks where separate encoders impose fixed resolution or temporal granularity. Quantitative comparisons on named benchmarks are not available in the abstract.

Authors: Guowei Xu, Zhenting Qi, Huangyuan Su, Weirui Ye, Himabindu Lakkaraju, Sham M. Kakade, Yilun Du

Org: Harvard University

Problem: Self-improvement methods for language models that rely on expansion-only forward search are theoretically confined to a narrow entropy shell and require exponentially many samples to solve hard problems.

Method: Bidirectional Evolutionary Search (BES) couples forward candidate evolution — using recombination, mutation, and selection operators over solution trajectories — with backward goal decomposition that recursively splits a hard goal into checkable subgoals. A theoretical analysis proves that evolutionary recombination escapes the entropy-shell confinement and that backward decomposition exponentially reduces the sample complexity of finding correct solutions.

Key results:

• Outperforms existing open-source self-improvement frameworks on three open problem-solving benchmarks in both average and best-case performance
• Specific benchmark names and score deltas not stated in abstract

Why it matters / caveats: The theoretical grounding distinguishes BES from purely empirical self-play methods and suggests principled scaling properties. Evaluation is limited to three benchmarks; real-world deployment costs of the dual-direction search are not quantified.

Authors: Guijin Son, Seungyeop Yi, Minju Gwak, Hyunwoo Ko, Wongi Jang, Youngjae Yu

Org: Seoul National University

Problem: Language model mathematical reasoning benchmarks are saturated at competition level, and no large-scale dataset of genuinely unsolved or research-level problems with verifiable trajectories exists.

Method: The authors build ResearchMath-14K by sourcing 14,056 research-level problems from academic literature and generating 220K reasoning trajectories using teacher language model agents. They identify a hallucination scaling problem — newer model generations produce 5.6× more references and 5.0× more fabricated references per trace — and apply trajectory filtering before fine-tuning Qwen3 models of varying sizes.

Key results:

• Dataset: 14,056 problems; 220K teacher trajectories
• Newer model generations produce 5.6× more references and 5.0× more fake references per trace
• Fine-tuning yields +9.2 points on average across Qwen3 models from 4B to 30B parameters
• Dataset publicly available at Hugging Face (amphora/ResearchMath-14k)

Why it matters / caveats: Filling the research-level math gap is important for evaluating frontier reasoning. The hallucination-scaling finding is a cautionary result for any pipeline relying on newer models to generate training data without filtering.

Authors: Caijun Xu, Changyi Xiao, Zhongyuan Peng, Yixin Cao

Org: Fudan University

Problem: RL-based reasoning improvement depends heavily on costly external teacher models or carefully curated correct demonstrations and does not leverage the information in already-observed failure traces.

Method: DenoiseRL converts incorrect reasoning traces into training signal by treating them as noisy prefixes and training the model to recover from those prefixes using a failure-oriented optimization objective. The framework is on-policy in that it bootstraps from the model's own weak outputs, avoiding dependence on a superior teacher model or pre-filtered datasets.

Key results:

• Consistently outperforms on-policy RL baselines on mathematical and general reasoning benchmarks (specific margins not stated)
• Demonstrates stronger self-corrective behavior as training difficulty increases
• Improves training efficiency relative to teacher-dependent approaches

Why it matters / caveats: The failure-reuse strategy is data-efficient and removes the teacher-model dependency that limits RL scaling in practice. Concrete numeric comparisons against specific baselines on named benchmarks are not available from the abstract.

Authors: Xinle Deng, Ruobin Zhong, Hujin Peng, Xiaoben Lu, Yanzhe Wu, Guang Li, Buqiang Xu, Yunzhi Yao, Jizhan Fang, Haoliang Cao, Junjie Guo, Yuan Yuan, Ziqing Ma, Yuanqiang Yu, Rui Hu, Baohua Dong, Hangcheng Zhu, Ningyu Zhang

Org: Zhejiang University (ZJUNLP) / Alibaba

Problem: Debugging failures in LLM memory systems (long-context, RAG, Mem0, EverMemOS) is difficult because errors cascade across operations and existing tools lack fine-grained attribution of root causes.

Method: MemTrace transforms memory pipelines into executable memory evolution graphs that make information flow traceable at the operation level. An automated attribution algorithm identifies root-cause operations — such as information loss and retrieval misalignment — and a paired prompt optimization step applies targeted corrections without manual intervention. A benchmark, MemTraceBench, covers failure patterns across four representative memory systems.

Key results:

• Automatic fault correction and prompt optimization improve end-task metrics by up to 7.62%
• Systematic coverage of failure modes across Long-Context, RAG, Mem0, and EverMemOS systems

Why it matters / caveats: Systematic error attribution for memory pipelines is underexplored and practically important as RAG and agentic memory systems are widely deployed. The 7.62% improvement figure is the upper bound of a range; average gains are not stated.

Authors: Ruowen Zhao, Bangguo Li, Zuyan Liu, Yinan Liang, Junliang Ye, Fangfu Liu, Diankun Wu, Zhengyi Wang, Xumin Yu, Yongming Rao, Han Hu, Jun Zhu

Org: Tencent Hunyuan

Problem: Standard text-guided vision-language pre-training lacks low-level spatial and physical knowledge needed for precise manipulation in embodied environments.

Method: GEM adds an auxiliary depth-map generation head to a vision-language pre-training pipeline so that the backbone must learn spatially grounded representations during pre-training. The team releases GEM-4M, a dataset combining grounding, reasoning, and planning data with depth-map supervision, and builds GEM-VLA, an action model fine-tuned on top of this pre-trained backbone for robotic manipulation tasks.

Key results:

• State-of-the-art results across diverse embodied benchmarks (specific scores not stated)
• GEM-VLA demonstrates superior task execution in both simulation and real-world settings (numeric margins not stated)

Why it matters / caveats: Coupling generative depth supervision with language pre-training is a low-cost way to inject spatial priors without requiring 3D sensor hardware at inference time. Exact benchmark numbers and the scale of improvements over prior VLAs are not available from the abstract.

Authors: Suji Kim, Kangsan Kim, Sung Ju Hwang

Org: KAIST AI

Problem: Small computer-use agents underperform larger models in specific software domains because domain-specific training data is scarce and not targeted at their actual failure modes.

Method: LearnWeak uses a stronger reference agent to identify weak areas of the target small model, automatically generates domain-targeted tasks for those weak areas, and trains with an error-aware objective that distinguishes planning errors from execution errors — allowing the loss to be weighted toward the more informative error type.

Key results:

• +11.6 percentage points for EvoCUA-8B on OSWorld benchmark across eight domains
• +11.1 percentage points for OpenCUA-7B under the same evaluation

Why it matters / caveats: Automated weakness-driven specialization reduces the need for manual dataset curation and is applicable to any domain where a stronger reference agent is available. The approach is evaluated on OSWorld only; generalization to other computer-use benchmarks is not demonstrated.

Authors: Rui Meng, Bhavana Dalvi Mishra, Jiefeng Chen, Chun-Liang Li, Palash Goyal, Mihir Parmar, Yiwen Song, Yale Song, Rajarishi Sinha, Parthasarathy Ranganathan, Burak Gokturk, Jinsung Yoon, Tomas Pfister

Org: Google

Problem: Autonomous research agents hallucinate references and produce unverifiable claims, making their outputs unreliable for scientific use.

Method: ScientistOne wraps an agentic research pipeline in a Chain-of-Evidence (CoE) framework that requires every claim to be backed by traceable evidence links at generation time. A CoE Audit module then performs post-hoc integrity checks verifying citation accuracy, reported scores, and method-code alignment. The system was evaluated across 75 papers and five research systems on tasks including Parameter Golf and MLE-Bench.

Key results:

• Hallucinated references: up to 21% in baselines vs. 0/337 for ScientistOne
• Score verification: as low as 42% correct in baselines vs. 12/12 (100%) for ScientistOne
• Method-code alignment: 20–80% in baselines vs. 14/15 for ScientistOne
• Gold medals on MLE-Bench tasks; state-of-the-art on Parameter Golf

Why it matters / caveats: Near-zero hallucinated references is a strong result; the CoE framework offers a reusable pattern for verifiable AI research. Evaluation is over 75 papers, which is modest, and all tasks are ML-domain benchmarks rather than open-ended scientific discovery.

Authors: Yixuan Tang, Yi Yang

Org: Not stated

Problem: It is unclear whether AI research agents broaden or concentrate scientific exploration when used to generate research ideas at scale.

Method: The authors ran four AI research-agent frameworks across six large language models, generating 37,802 research ideas seeded from shared literature in AI/ML subfields. They compared the semantic concentration and citation impact of AI-generated ideas against human-authored follow-up papers using distributional similarity measures and citation analysis.

Key results:

• 37,802 AI-generated ideas produced across 4 frameworks × 6 LLMs
• AI-generated ideas are substantially more concentrated than human-authored papers
• AI ideas remain closer to source materials than human follow-up work
• Papers resembling AI-generated ideas receive lower citations
• Novel differences stem from recombining existing methods, not introducing new research questions

Why it matters / caveats: This is an empirical warning for the research community: heavy reliance on AI ideation tools may homogenize the literature and reduce citation impact. The study is restricted to AI/ML domains and does not test whether human-AI collaboration mitigates the concentration effect.

Authors: Jizhan Fang, Buqiang Xu, Zhixian Wang, Haoliang Cao, Xinle Deng, Baohua Dong, Hangcheng Zhu, Ruohui Huang, Gang Yu, Ying Wei, Guozhou Zheng, Feiyu Xiong, Haofen Wang, Huajun Chen, Ningyu Zhang

Org: Alibaba

Problem: Static or flat memory representations in LLM agents fail to capture evolving relationships between memories and degrade in complex, long-horizon environments.

Method: FluxMem models agent memory as a heterogeneous graph whose topology is progressively refined through three stages: link repair (adding missing connections), interference pruning (removing conflicting edges), and abstraction alignment (merging semantically redundant nodes). The graph structure dynamically updates as new observations arrive, maintaining a compact and accurate memory representation across tasks.

Key results:

• State-of-the-art performance on LoCoMo, Mind2Web, and GAIA benchmarks (specific scores not stated)
• Strong adaptation and generalization in complex agentic environments (numeric margins not stated)

Why it matters / caveats: Graph-based memory with topology refinement is a principled approach to handling contradictory or redundant memories over long horizons. Exact score improvements over baselines are not available from the abstract, limiting direct comparison.

Authors: Yunyang Ge, Xianyi He, Zezhong Zhang, Bin Lin, Bin Zhu, Xinhua Cheng, Li Yuan

Org: Peking University

Problem: Full-attention diffusion transformers for text-to-video generation are computationally prohibitive, with quadratic complexity in sequence length blocking practical high-resolution, long-duration generation.

Method: OSP-Next combines four techniques: Skiparse-2D Attention (a hybrid sparse attention pattern), Sparse Sequence Parallelism (distributed sparse attention that reduces inter-GPU communication overhead), HiF8 quantization (a novel 8-bit floating-point format applied to activations and weights), and reinforcement learning fine-tuning to recover quality lost during sparsification and quantization.

Key results:

• VBench total score: 83.73%, surpassing the Wan2.1 baseline
• Single-GPU speedup: up to 1.64× on NVIDIA H200
• Eight-GPU speedup: over 1.52× on H200
• With HiF8 quantization: 1.69× and 2.27× speedups on Ascend 950PR
• VBench score degradation from quantization: only 0.4%

Why it matters / caveats: The combination of sparse attention and HiF8 quantization with RL-based quality recovery is a practical deployment recipe for high-resolution video diffusion. Evaluations focus on VBench; perceptual quality across diverse prompts and human evaluations are not reported.

Authors: Ke Lin, Yiyang Luo, Zhaolong Su, Yunya Song, Anyi Rao

Org: Not stated

Problem: Autoregressive language models with linear-time inference (e.g., RWKV) are causally constrained and cannot natively exploit the bidirectional context that improves discrete diffusion language models.

Method: B³D-RWKV (Triplet-Block Diffusion RWKV) interleaves RWKV causal blocks with bidirectional diffusion blocks in a triplet layout, allowing the model to maintain O(L) inference efficiency for the recurrent component while enabling parallel, bidirectional denoising within diffusion steps. The architecture is trained as a discrete diffusion model and evaluated at 7.2B parameters.

Key results:

• 7.2B parameter variant reaches comparable accuracy to existing models on an 8-task evaluation suite
• Average 1.6× decoding throughput speedup compared to baselines

Why it matters / caveats: Combining linear-time recurrence with bidirectional diffusion is an architecturally interesting hybrid that could serve as a faster alternative to full-attention diffusion models. The 8-task suite is small and accuracy is described as "comparable" rather than superior; the quality-speed trade-off needs broader evaluation.

Authors: Hao Jiang, Shurui Li, Tianpeng Bu, Bowen Xu, Xin Liu, Qihua Chen, Hongtao Duan, Lulu Hu, Bin Yang, Minying Zhang

Org: Alibaba

Problem: Online RL training of large language models suffers from imbalanced exploration-exploitation trade-offs, and there is no principled metric to diagnose or optimize this balance during tree-based policy search.

Method: IB-TPO introduces IB-Score, a metric derived from Information Bottleneck theory that quantifies the exploration-exploitation balance of a policy at any training step. The score is incorporated directly as an optimization objective, and an IB-guided tree sampling strategy uses the score to allocate rollout budget toward nodes where the balance is most degraded, replacing uniform or reward-only sampling.

Key results:

• 2.9% to 3.6% performance improvement over GRPO baseline
• 50% greater trajectory efficiency under an equivalent token budget
• Accepted to ICML 2026 main conference

Why it matters / caveats: Grounding exploration-exploitation balance in Information Bottleneck theory provides a theoretically motivated diagnostic tool that can be plugged into existing tree-search RL pipelines. Gains over GRPO are modest (under 4%) and are reported on a single set of tasks; broader benchmarking is needed.

Authors: Kewei Zhang, Jin Wang, Sensen Gao, Chengyue Wu, Yulong Cao, Songyang Han, Boris Ivanovic, Langechuan Liu, Marco Pavone, Song Han, Daquan Zhou, Enze Xie

Org: NVIDIA

Problem: Diffusion-based vision-language-action models for autonomous driving are accurate but too slow for real-time deployment due to iterative denoising over long token sequences.

Method: Fast-dDrive is a block-diffusion VLA that applies bidirectional denoising within semantic token blocks while preserving causal ordering across blocks. It introduces structured token freezing (keeping high-confidence structural tokens fixed across denoising steps), Scaffold Speculative Decoding (using a lightweight draft model seeded by frozen tokens), and trajectory rollout averaging to improve robustness.

Key results:

• nuScenes: average L2 error reduced to 0.32 m, a 22% improvement over prior state of the art
• WOD-E2E test set: state-of-the-art ADE@3s, ADE@5s, and highest RFS among diffusion-based VLAs
• 12× throughput speedup over the autoregressive baseline when integrated with SGLang

Why it matters / caveats: A 12× throughput gain with a 22% accuracy improvement makes this a strong practical advance for onboard deployment of diffusion VLAs. Evaluation is on nuScenes and WOD; real-world closed-loop driving tests are not reported.

Authors: Boxiang Qiu, Liliang Chen, Yue Liao, Nan Wang, Lintao Wang, Jiayi Luo, Wenzhi Zhao, Shengcong Chen, Di Chen, Ye Li, Chen Gao, Shuicheng Yan, Si Liu, Maoqing Yao, Guanghui Ren

Org: AgiBot World

Problem: Existing video world simulators for robotic manipulation lack closed-loop evaluation capability, accurate state decoding, and fast enough inference for practical policy training.

Method: GE-Sim 2.0 is a 2B-parameter action-conditioned video generation model trained on real-world teleoperation and contact-rich interaction data. It adds three components to the base video generator: a state decoder that extracts proprioceptive information from generated frames, a world scoring mechanism that evaluates task completion from rollouts, and an accelerated inference framework supporting up to 4× frame skipping for extended-horizon evaluation.

Key results:

• 2B parameters total
• Generates a 25-frame rollout in 2.3 seconds on a single H100 GPU
• Up to 4× frame skipping for extended horizon evaluation
• Leads the WorldArena public leaderboard
• Outperforms dedicated robotic world models and proprietary video generators (numeric margins not stated)
• Policies trained with GE-Sim 2.0 rollouts show real-world effectiveness

Why it matters / caveats: Fast, closed-loop video simulation at 2B parameters is a practical milestone for data-efficient robot policy training without physical hardware. Leaderboard rankings are comparative snapshots; long-term sim-to-real transfer quality and failure modes under distribution shift are not characterized.