Proposes InfiniteScienceGym, a seed-based, verifiable scientific benchmark for evaluating reasoning and tool use, with models achieving only ~50% accuracy.

Oliver Bentham, Vivek Srikumar

SD-Zero transforms binary rewards into dense token-level supervision via self-revision and self-distillation, boosting reasoning performance by over 10%.

Yinghui He, Simran Kaur, Adithya Bhaskar et al.

Proposes TCER (Triviality Corrected Endogenous Reward), leveraging relative information gain to mitigate bias in open-ended text generation, improving diversity and quality.

Xinda Wang, Zhengxu Hou, Yangshijie Zhang et al.

METRO uses large language models to autonomously extract strategies from expert dialogues, improving performance by 9%-10%.

Haofu Yang, Jiaji Liu, Chen Huang et al.

Proposes Polyglot Score to evaluate 10 multilingual teacher models; finds data quality outweighs model size in effectiveness.

Lester James V. Miranda, Ivan Vulić, Anna Korhonen

CARE framework evaluates LLM alignment with community linguistic behaviors through reaction tone analysis.

Nuan Wen, Xuezhe Ma

BLUEmed combines multi-agent debate with hybrid retrieval-augmented generation, achieving 69.13% accuracy in clinical terminology error detection.

Saukun Thika You, Nguyen Anh Khoa Tran, Wesley K. Marizane et al.

Proposes GIANTS, a reinforcement learning-based model that improves scientific insight prediction by 34% on 17k samples, outperforming proprietary baselines.

Joy He-Yueya, Anikait Singh, Ge Gao et al.

RecaLLM addresses the 'lost-in-thought' issue with explicit in-context retrieval, excelling on RULER and HELMET benchmarks.

Kyle Whitecross, Negin Rahimi

MedConceal uses an interactive patient simulator to evaluate hidden-concern reasoning in medical dialogue, featuring 300 cases and 600 interactions.

Yikun Han, Joey Chan, Jingyuan Chen et al.

By decomposing long-context reasoning into atomic skills, the method enhances LLMs' reasoning by an average of 7.7%.

Yanling Xiao, Huaibing Xie, Guoliang Zhao et al.

Unified framework for LLM post-training based on trajectory provenance, supporting expansion, reshaping, and consolidation roles.

Shiwan Zhao, Zhihu Wang, Xuyang Zhao et al.

This study evaluates multilingual embedding models' ability to encode general proficiency, revealing poor cross-corpus transfer with performance dropping from QWK≈0.7 to 0.38.

Laurits Lyngbaek, Ross Deans Kristensen-McLachlan

BOSCH employs black-box binary optimization via large neighborhood search for short-context head selection, outperforming static heuristics across models from 1.7B to 30B parameters.

Abbas Ghaddar, Ivan Kobyzev, Boxing Chen et al.

AutoSOTA automates AI model optimization using a multi-agent architecture, discovering 105 new SOTA models in 5 hours per paper.

Yu Li, Chenyang Shao, Xinyang Liu et al.

RTT framework maps response-level scores to token rewards via a Token Relevance Discriminator, improving instruction-following accuracy by 2.5% over baselines.

Tianze Xu, Yanzhao Zheng, Pengrui Lu et al.

Trivial vocabulary bans (e.g., 'very', 'just') outperform deep linguistic constraints like E-Prime in improving LLM reasoning, due to output regularization effects.

Rodney Jehu-Appiah

LLMimic uses role-playing LLM training to enhance AI literacy, reduce persuasion success, and improve social responsibility.

Qihui Fan, Min Ge, Chenyan Jia et al.

Using social media simulation, analysis reveals LLMs over-idealize disability, reinforcing biases and stereotypes.

Marco Bombieri, Simone Paolo Ponzetto, Marco Rospocher

Unified modular framework for agent memory; fusion model outperforms SOTA in long-term tasks with 8-15% improvements.

Yanchen Wu, Tenghui Lin, Yingli Zhou et al.