RedBench consolidates 37 datasets with 22 risk categories and 19 domains, enabling comprehensive LLM vulnerability evaluation.

Quy-Anh Dang, Chris Ngo, Truong-Son Hy

MedDialogRubrics evaluates LLMs' multi-turn diagnostic capabilities using a multi-agent system with 5,200 patient cases.

Lecheng Gong, Weimin Fang, Ting Yang et al.

Proposed Deferred Commitment Decoding (DCD) improves diffusion language model accuracy by 1.73%.

Yingte Shu, Yuchuan Tian, Chao Xu et al.

Introduces Hidden State Poisoning Attacks (HiSPA) against Mamba models, using short triggers to irreversibly overwrite hidden states and impair long-context retrieval, with experimental validation on ROBENCH-25.

Alexandre Le Mercier, Chris Develder, Thomas Demeester

mHC projects residual mappings onto the doubly stochastic matrix manifold, restoring identity properties and enhancing large-scale training stability.

Zhenda Xie, Yixuan Wei, Huanqi Cao et al.

Proposes Encyclo-K, a knowledge statement-based dynamic benchmark, achieving contamination resistance and multi-knowledge assessment, with top model accuracy at 62.07%.

Yiming Liang, Yizhi Li, Yantao Du et al.

AutoForge automates environment synthesis and ERPO enhances multi-environment RL stability, achieving 73.1 in tau-bench.

Shihao Cai, Runnan Fang, Jialong Wu et al.

Proposes CAT, a structured context management framework achieving 57.6% success on SWE tasks, significantly outperforming baseline methods.

Shukai Liu, Jian Yang, Bo Jiang et al.

Proposes multiple tokenization penalty functions to quantify the impact of word splitting on LLM performance; experiments show significant correlation.

Sachin Pawar, Manoj Apte, Kshitij Jadhav et al.

Utilized Transformer-based models and transfer learning to enhance NLP for six Senegalese languages, achieving up to 78% accuracy in text classification and BLEU scores of 32 in translation.

Derguene Mbaye, Tatiana D. P. Mbengue, Madoune R. Seye et al.

Axlerod integrates LLM and RAG to achieve 93.18% accuracy in insurance policy retrieval, reducing search time by 2.42 seconds.

Adam Bradley, John Hastings, Khandaker Mamun Ahmed

Memory-T1 employs reinforcement learning for time-aware memory retrieval, boosting multi-session dialogue reasoning to 67.0% accuracy, outperforming baselines by over 10%.

Yiming Du, Baojun Wang, Yifan Xiang et al.

Seed-Prover 1.5 combines agentic RL and test-time scaling, solving 87.9% of PutnamBench.

Jiangjie Chen, Wenxiang Chen, Jiacheng Du et al.

Introduces MMRB2 benchmark, evaluating omni reward models across four tasks; Gemini 3 Pro achieves 75-80% accuracy, surpassing GPT-4o's 59%.

Yushi Hu, Reyhane Askari-Hemmat, Melissa Hall et al.

This study introduces JustRL, a simple fixed-hyperparameter RL training method for 1.5B models, surpassing complex approaches with more stability and less compute.

Bingxiang He, Zekai Qu, Zeyuan Liu et al.

This study benchmarks 6 SpeechLLMs across 16 datasets, revealing that while cascaded systems remain reliable, recent SpeechLLMs can match or outperform them, emphasizing LLM integration.

Sara Papi, Javier Garcia Gilabert, Zachary Hopton et al.

Proposes a unified 'forms-functions-dynamics' framework for agent memory, covering token-level, parametric, and latent types with detailed functional and dynamic analysis.

Yuyang Hu, Shichun Liu, Yanwei Yue et al.

FIN-bench-v2 evaluates Finnish LLMs using a unified benchmark suite, validated with a 2.15B parameter model.

Joona Kytöniemi, Jousia Piha, Akseli Reunamo et al.

AutoTool employs a dual-phase optimization with PL ranking to enable dynamic tool selection, improving reasoning accuracy by 6.4% across ten benchmarks.

Jiaru Zou, Ling Yang, Yunzhe Qi et al.

NL2Repo-Bench evaluates long-horizon repository generation of coding agents, with average pass rates below 40%, highlighting significant challenges.

Jingzhe Ding, Shengda Long, Changxin Pu et al.