Introduces Cartridges, a self-study trained lightweight KV cache that replicates large context models, reducing memory by 38.6× and increasing throughput 26.4×.

Sabri Eyuboglu, Ryan Ehrlich, Simran Arora et al.

Proposes a unified mathematical framework for LLM hallucinations, analyzing their roots, detection, and mitigation, with experimental validation.

Chaozhuo Li, Pengbo Wang, Chenxu Wang et al.

BioMol-MQA constructs a multi-modal drug interaction QA dataset integrating knowledge graphs, text, and SMILES, enhancing LLM reasoning in biomedical tasks.

Saptarshi Sengupta, Shuhua Yang, Paul Kwong Yu et al.

Introduces MMSU, a comprehensive benchmark with 47 tasks, integrating audio, semantics, prosody, and reasoning, evaluating 22 SpeechLLMs.

Dingdong Wang, Junan Li, Jincenzi Wu et al.

R-Search enhances LLM reasoning with search via multi-reward reinforcement learning, achieving a 32.2% improvement.

Qingfei Zhao, Ruobing Wang, Dingling Xu et al.

Pre³ transforms LR(1) grammar into DPDA, reducing decoding time by 40% and increasing throughput by 36%.

Junyi Chen, Shihao Bai, Zaijun Wang et al.

A prompt-based multi-agent framework raises GPT-4o-mini PrivacyQA performance from 0.394 to 0.587 while narrowing dialect gaps.

Đorđe Klisura, Astrid R Bernaga Torres, Anna Karen Gárate-Escamilla et al.

Fine-tuning OpenAI Whisper with limited MSA data significantly improves recognition of five major Arabic dialects, matching larger models' performance.

Ömer Tarik Özyilmaz, Matt Coler, Matias Valdenegro-Toro

STORYTELLER uses SVO nodes and dynamic modules to improve story coherence, achieving 84.33% human preference rate.

Jiaming Li, Yukun Chen, Ziqiang Liu et al.

Introduces Common Corpus, a 2 trillion token open multilingual dataset for ethical LLM pretraining.

Pierre-Carl Langlais, Pavel Chizhov, Catherine Arnett et al.

SealQA benchmark evaluates search-augmented LLMs' reasoning under noisy/conflicting info; models perform poorly, accuracy below 50%.

Thinh Pham, Nguyen Nguyen, Pratibha Zunjare et al.

AnnaAgent uses multi-agent collaboration and tertiary memory to realistically simulate seeker emotional evolution and multi-session memory in psychological counseling.

Ming Wang, Peidong Wang, Lin Wu et al.

Introduces TSGD-M, a sampling momentum method that scales textual gradients effectively within limited context windows, improving prompt optimization performance.

Zixin Ding, Junyuan Hong, Zhan Shi et al.

ComposeRAG employs a modular architecture with question decomposition, retrieval decision, and verification, achieving up to 15% accuracy improvement in multi-hop QA.

Ruofan Wu, Youngwon Lee, Fan Shu et al.

Voice conversion improves cross-domain robustness for Arabic dialect ID, achieving 34.1% accuracy increase.

Badr M. Abdullah, Matthew Baas, Bernd Möbius et al.

Proposes SCRIPT encoding based on Unicode script/category, with constrained BPE merging, achieving robust multilingual pretokens and reducing partial UTF-8 sequences.

Sander Land, Catherine Arnett

Proposed a cross-level attribution method revealing a 'mid-activation, late-amplification' pattern, improving layer efficiency by 37% in sparse MoE models.

Junzhuo Li, Bo Wang, Xiuze Zhou et al.

Proposes HARDTESTGEN, leveraging LLMs to synthesize high-quality test cases, improving verification precision by 11.3% and recall by 17.5%.

Zhongmou He, Yee Man Choi, Kexun Zhang et al.

DeepTheorem enhances LLM reasoning for theorem proving using natural language and RL, featuring a 121K informal theorem dataset.

Ziyin Zhang, Jiahao Xu, Zhiwei He et al.

Table-R1 enhances table reasoning via inference-time scaling using distillation and RLVR, surpassing GPT-4.1 with a 7B model.

Zheyuan Yang, Lyuhao Chen, Arman Cohan et al.