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.

C$^3$TG achieves multi-dimensional text generation with conflict-aware control, significantly improving attribute accuracy.

Yu Li, Zhe Yang, Yi Huang et al.

Proposed Supervised Rational Attention (SRA), aligning model attention with human rationales, boosting interpretability and fairness.

Brage Eilertsen, Røskva Bjørgfinsdóttir, Francielle Vargas et al.

Llama-Embed-Nemotron-8B employs contrastive learning and multi-model data synthesis, achieving SOTA performance on multilingual and cross-lingual tasks, based on 16.1M query-document pairs, with open-source weights.

Yauhen Babakhin, Radek Osmulski, Ronay Ak et al.

OckBench jointly evaluates accuracy and token efficiency, revealing large disparities in reasoning density among models.

Zheng Du, Hao Kang, Song Han et al.

BACO framework uses token-level model collaboration with routing strategies to optimize diversity and quality, achieving 21.3% joint improvement.

Yichen Wang, Chenghao Yang, Tenghao Huang et al.

Batch prompting effectively suppresses overthinking in large reasoning models, reducing reasoning tokens by 76%.

Saurabh Srivastava, Janit Bidhan, Hao Yan et al.

MUTANT employs a two-stage BPE training with language-aware preprocessing, reducing fertility by 39.5%, boosting inference throughput by 44%.

Souvik Rana, Arul Menezes, Ashish Kulkarni et al.

SpecDiff-2 combines discrete diffusion models with alignment techniques, achieving up to 5.5× speed-up in LLM inference without accuracy loss.

Jameson Sandler, Jacob K. Christopher, Thomas Hartvigsen et al.

Toolathlon benchmarks 32 real-world apps with 604 tools; top model Claude-4.5-Sonnet achieves only 38.6% success in complex long-horizon tasks.

Junlong Li, Wenshuo Zhao, Jian Zhao et al.

MDP-Agent enhances AI search by transforming documents into LLM-ready inputs, achieving 35% accuracy improvement.

Hongjin Qian, Zheng Liu

Leveraging Gemma 3, MetricX-25 and GemSpanEval enhance translation evaluation and error detection with multi-task learning and generative span output.

Juraj Juraska, Tobias Domhan, Mara Finkelstein et al.

WebLeaper introduces a tree-structured reasoning framework with multi-source data synthesis, significantly improving WebAgent's search efficiency and effectiveness.

Zhengwei Tao, Haiyang Shen, Baixuan Li et al.

R3 framework uses reinforcement learning to optimize retrieval in RAG, improving performance by 5.2%.

Jiawei Zhou, Lei Chen

Pie system enables flexible, efficient LLM serving via APIs and inferlets, boosting throughput by 1.3x-3.4x.

In Gim, Zhiyao Ma, Seung-seob Lee et al.

Constructed TicToc dataset to evaluate LLMs' temporal awareness; models show less than 65% alignment with human preferences under timestamp info.

Yize Cheng, Arshia Soltani Moakhar, Chenrui Fan et al.

Introduces TIR-Judge, an RL framework integrating tools to improve LLM judging, surpassing multiple benchmarks with only 8B parameters.

Ran Xu, Jingjing Chen, Jiayu Ye et al.

EchoMind introduces a multi-level benchmark to evaluate SLMs in speech understanding, vocal cue perception, reasoning, and empathetic dialogue generation.

Li Zhou, Lutong Yu, You Lyu et al.

E2Rank unifies retrieval and listwise reranking using continued training on a single embedding model, achieving state-of-the-art results with high efficiency.

Qi Liu, Yanzhao Zhang, Mingxin Li et al.

Proposes GEMs with diverse architectures and high-quality corpora, achieving up to 3.6% accuracy improvement in Greek NLP tasks.

Alexandra Apostolopoulou, Konstantinos Kanaris, Athanasios Koursaris et al.