Proposes lightweight clip selection combined with large language models to extract key moments, achieving near-reference summary quality with less than 6% video content.

Galann Pennec, Zhengyuan Liu, Nicholas Asher et al.

The FACTS Leaderboard evaluates large language models' factuality using four sub-leaderboards, with an average score of 68.8.

Aileen Cheng, Alon Jacovi, Amir Globerson et al.

Fine-tuning on narrow datasets causes broad, unpredictable behaviors and hidden backdoors, exposing security risks in LLMs.

Jan Betley, Jorio Cocola, Dylan Feng et al.

SAGE benchmark evaluates LLMs' cross-cultural understanding via 9 dimensions and 4,530 tasks, exposing systematic deficiencies.

Shiwei Guo, Sihang Jiang, Qianxi He et al.

DeepSeek-V3.2 employs sparse attention and large-scale reinforcement learning to match GPT-5's reasoning and agent capabilities, significantly narrowing the open-source gap.

DeepSeek-AI, Aixin Liu, Aoxue Mei et al.

Introducing reasoning overconfidence in LLMs; Long-CoT reduces it via iterative reflection, based on the cognitive rigidity hypothesis.

Jiannan Guan, Qiguang Chen, Libo Qin et al.

Decentralized peer-to-peer framework Matrix boosts multi-agent synthetic data generation throughput by 2-15× via message-driven asynchronous scheduling.

Dong Wang, Yang Li, Ansong Ni et al.

Task Arithmetic is the only method that reliably improves LLM performance in 'in-the-wild' settings.

Oğuz Kağan Hitit, Leander Girrbach, Zeynep Akata

Proposed retrieval-aware PMP enhances sarcasm detection, achieving 9.87% macro-F1 improvement on Twitter Indonesia dataset.

Michael Iskandardinata, William Christian, Derwin Suhartono

Evo-Memory evaluates self-evolving memory in LLMs during test-time, achieving an average performance gain of 0.65 across diverse tasks.

Tianxin Wei, Noveen Sachdeva, Benjamin Coleman et al.

Introduces RLER, combining search and evolving rubrics, training DR Tulu-8B to outperform open-source models with 15.6% gain on benchmarks.

Rulin Shao, Akari Asai, Shannon Zejiang Shen et al.

SMILE integrates semantic and lexical precision, achieving efficient and accurate QA evaluation.

Shrikant Kendre, Austin Xu, Honglu Zhou et al.

Nemotron Elastic achieves efficient many-in-one reasoning by embedding nested submodels, reducing training costs by 360x.

Ali Taghibakhshi, Sharath Turuvekere Sreenivas, Saurav Muralidharan et al.

Poetic style as a universal single-turn jailbreak, achieving 62% attack success across 25 LLMs, with high cross-domain transferability.

Piercosma Bisconti, Matteo Prandi, Federico Pierucci et al.

Agent-R1 introduces step-level trajectory abstraction and flexible context management, enabling multi-turn reinforcement learning with diverse optimization strategies, achieving state-of-the-art results.

Mingyue Cheng, Shuo Yu, Daoyu Wang et al.

Spark-Prover-X1 enhances formal theorem proving via diverse data training, solving 27 problems on PutnamBench.

Xinyuan Zhou, Yi Lei, Xiaoyu Zhou et al.

LLMLagBench uses PELT change-point detection to identify training data cutoffs in LLMs, revealing many models' knowledge ends earlier than declared.

Piotr Pęzik, Konrad Kaczyński, Maria Szymańska et al.

Introduces GAD, an adversarial framework for black-box LLM distillation, with Qwen2.5-14B approaching GPT-5-Chat performance.

Tianzhu Ye, Li Dong, Zewen Chi et al.

ADI-20 dataset expands ADI-17 for Arabic Dialect Identification using ECAPA-TDNN and Whisper models.

Haroun Elleuch, Salima Mdhaffar, Yannick Estève et al.

InTRO employs token-level exploration and self-feedback via KL divergence, boosting reasoning accuracy by up to 20% and producing more concise rationales.

Mingye Zhu, Yi Liu, Zheren Fu et al.