BRANE uses LLM-extracted query features to dynamically optimize retrieval configurations, achieving up to 89% cost savings on MuSiQue and others.

Melissa Z. Pan, Negar Arabzadeh, Mathew Jacob et al.

MOSS enables source-level self-rewriting in autonomous agents, boosting OpenClaw’s four-task mean grader score from 0.25 to 0.61 in one cycle.

Qianshu Cai, Yonggang Zhang, Xianzhang Jia et al.

LCGuard uses adversarially learned transformations on Transformer KV caches to reduce sensitive information reconstruction in multi-agent systems while preserving task performance.

Sadia Asif, Mohammad Mohammadi Amiri, Momin Abbas et al.

ToolCUA optimizes GUI-Tool path selection via staged training, achieving 46.85% accuracy.

Xuhao Hu, Xi Zhang, Haiyang Xu et al.

The study proposes a framework to diagnose reward hacking in rubric-based RL, finding that even strong verification does not eliminate reward hacking.

Anas Mahmoud, MohammadHossein Rezaei, Zihao Wang et al.

DR-Gym environment optimizes electric utility demand response using reinforcement learning, enhancing grid flexibility and energy affordability.

Jose E. Aguilar Escamilla, Lingdong Zhou, Xiangqi Zhu et al.

SciCrafter evaluates AI's discovery-to-application ability in Minecraft; current models achieve only 26% success.

Zhou Ziheng, Huacong Tang, Jinyuan Zhang et al.

Proposes an LLM-based evaluation framework to enhance math reasoning assessment accuracy beyond symbolic math limitations.

Erez Yosef, Oron Anschel, Shunit Haviv Hakimi et al.

AgentSearchBench improves agent search ranking quality using execution signals, bridging the gap between semantics and performance.

Bin Wu, Arastun Mammadli, Xiaoyu Zhang et al.

A-MAR framework enhances multimodal art retrieval explanation quality through structured reasoning plans.

Shuai Wang, Hongyi Zhu, Jia-Hong Huang et al.

SafetyALFRED evaluates safety planning in multimodal LLMs in kitchen settings, finding good hazard recognition but low risk mitigation success.

Josue Torres-Fonseca, Naihao Deng, Yinpei Dai et al.

Large language models exhibit normative conformity, revealing underlying mechanisms.

Mikako Bito, Keita Nishimoto, Kimitaka Asatani et al.

MathNet provides a global multimodal benchmark for mathematical reasoning and retrieval, covering 30,676 Olympiad-level problems from 47 countries.

Shaden Alshammari, Kevin Wen, Abrar Zainal et al.

BLF system achieves state-of-the-art binary forecasting performance on ForecastBench using sequential Bayesian updating of linguistic beliefs.

Kevin Murphy

ClawEnvKit automates environment generation for claw-like agents, reducing costs by 13,800x.

Xirui Li, Ming Li, Derry Xu et al.

Proposed DeepInsightTheorem framework enhances informal theorem proving by identifying core techniques, significantly outperforming baselines.

Yunhe Li, Hao Shi, Bowen Deng et al.

Using the CompCQ framework, this study analyzes LLM-generated competency questions across domains, revealing generation characteristics.

Reham Alharbi, Valentina Tamma, Terry R. Payne et al.

The study shows language models exhibit strong spatial transfer in shortest path problems but fail in length scaling due to recursive instability.

Yao Tong, Jiayuan Ye, Anastasia Borovykh et al.

Diagnosing LLM judge reliability using transitivity analysis and conformal prediction sets, revealing 33%-67% documents with at least one 3-cycle.

Manan Gupta, Dhruv Kumar

Study reveals LLMs and VLMs struggle with viewpoint rotation understanding without vision, proposes VRUBench dataset, and improves performance via selective fine-tuning.

Zhen Yang, Ping Jian, Zhongbin Guo et al.