HUSH-Bench shows that memory availability is not authorization: UIS reaches 83.0 despite retrieval reducing exact-target exposure to 23.0–30.3%.

Lingxiang Xu, Jiaoyun Yang, Min Hu et al.

Introduces RedditPersona framework using five community partition strategies to train parameter-efficient adapters, enhancing community behavior recognition.

Amirhossein Ghaffari, Ali Goodarzi, Huong Nguyen et al.

Proposes biomedical world models learning multiscale latent states and intervention-conditioned dynamics for future simulation and decision support.

Guangyu Wang, Jingkun Yue, Siqi Zhang et al.

SciVisAgentSkills enhances coding agents for scientific data analysis and visualization efficiency.

Kuangshi Ai, Haichao Miao, Kaiyuan Tang et al.

EpiEvolve adapts a frozen epidemic LLM through memory, reaching 0.629 accuracy and cutting post-shift recovery from 5 to 2 weeks.

Yiming Lu, Sihang Zeng, Zhengxu Tang et al.

ALE benchmark evaluates AI on long-term, high-value real-world industry tasks; current top models achieve less than 1% success rate on hardest tasks.

Yiyou Sun, Xinyang Han, Weichen Zhang et al.

Proposes trajectory-based safety alignment, significantly improving LLM robustness against mid-inference token injections.

Kyungmin Park, Taesup Kim

BiNSGPS uses bidirectional neuro-symbolic feedback, achieving 90.5% on Geometry3K and 90.1% on PGPS9K.

Qi Wang, Peijie Wang, Fei Yin et al.

LEAP uses blueprint-driven decomposition and iterative self-correction to boost formal proof success from below 10% to 70%, solving all 12 Putnam 2025 problems.

Po-Nien Kung, Linfeng Song, Dawsen Hwang et al.

Proposes an information-gain-based clarification framework for LLM agents, improving success rate by 3.7% with 0.3 fewer interaction steps.

Mengyi Deng, Zhiwei Li, Xin Li et al.

SafeSteer employs localized on-policy distillation focusing on safety tokens, reducing reliance on large datasets and auxiliary reward models, achieving a superior safety-capability trade-off.

Hao Li, Jingkun An, Zijun Song et al.

Iteris employs an explore-plan-execute loop with multi-agent collaboration to generate numerical evidence and proof drafts, verified through expert review, advancing open problems in computational mathematics.

Leheng Chen, Zihao Liu, Wanyi He et al.

Proposes structured post-retrieval assembly, boosting memory accuracy to 82%/93% single-hop and 27%/41% multi-hop, surpassing all prior results.

Vikas Reddy, Sumanth Reddy Challaram

Andes introduces a self-evolving tree routing framework to enhance weak models' instruction alignment, achieving SOTA results with 33.39% accuracy on PostTrainBench.

Zhengyang Zhao, Shengjie Ye, Lu Ma et al.

DAG-MoE improves MoE models by structural aggregation, enhancing language model performance.

Jiarui Feng, Hanqing Zeng, Karish Grover et al.

CARGO is a training-free routing framework using response agreement and Bayesian early stopping to decide when local models can answer reliably, supporting adjustable cloud offloading ratios.

Evan Chen, Shiqiang Wang, Kevin S Chan et al.

This work reveals that decoder attention capacity limits cause failure in deterministic reasoning; tool delegation helps surpass this barrier.

Dongxin Guo, Jikun Wu, Siu Ming Yiu

Introduces context-dependent argumentation frameworks (CDAFs) with perspective activation, analyzing the complexity of strategic attack manipulation, establishing NP-completeness bounds.

Albert Sadowski, Jarosław A. Chudziak

PAJAMA distills LLM evaluation into programs, matching 13B model accuracy while cutting API costs by 50×.

Tzu-Heng Huang, Shengqi Qiu, Frederic Sala

LISA combines linear attention and dynamic indexing to boost long-sequence reasoning speed by 50% with 5.6% accuracy gain.

Yu Zhao, Zekun Zhang, Fan Jiang et al.