Proposed MobileIPL with iterative preference learning significantly improves mobile GUI reasoning, outperforming SOTA models by over 10% on key benchmarks.

Kun Huang, Weikai Xu, Yuxuan Liu et al.

CoLM enables progressive scaling and elastic inference; CoLM-Air reaches about 3× faster 1M-token prefilling.

Kaitao Song, Xiaohua Wang, Xu Tan et al.

Introduced MedCaseReasoning dataset with 14,489 clinical cases; fine-tuning improves diagnostic accuracy by 29% and reasoning recall by 41%.

Kevin Wu, Eric Wu, Rahul Thapa et al.

Proposes Critique-Guided Distillation (CGD), leveraging teacher critiques to improve reasoning, achieving +7% average gains on benchmarks.

Berkcan Kapusuzoglu, Supriyo Chakraborty, Zain Sarwar et al.

Time-R1 employs a three-stage RL fine-tuning framework to endow a 3B-parameter LLM with comprehensive temporal reasoning, outperforming models over 200 times larger in future event prediction and creative scenario generation.

Zijia Liu, Peixuan Han, Haofei Yu et al.

HAPO employs history-aware policy optimization to reduce response length by 33-59% with only 2-5% accuracy loss in LLMs.

Chengyu Huang, Zhengxin Zhang, Claire Cardie

MetaSPO optimizes system prompts via meta-learning, enhancing performance across 14 datasets.

Yumin Choi, Jinheon Baek, Sung Ju Hwang

Qwen3 integrates thinking and non-thinking modes, with 235B parameters, enhancing multilingual and multi-task performance.

An Yang, Anfeng Li, Baosong Yang et al.

InForage, a reinforcement learning framework based on Information Foraging Theory, optimizes dynamic retrieval and reasoning, boosting complex task performance.

Hongjin Qian, Zheng Liu

This study introduces ChemRAG-Bench and ChemRAG-Toolkit, boosting chemistry task performance by 17.4% through multi-source retrieval-augmented generation.

Xianrui Zhong, Bowen Jin, Siru Ouyang et al.

MiMo-7B employs multi-stage data mixing and multi-token prediction to enhance reasoning, trained on 25 trillion tokens, with post-training on 130K problems, outperforming larger models.

LLM-Core Xiaomi, :, Bingquan Xia et al.

DynamicRAG uses reinforcement learning to adaptively rerank documents, boosting QA performance with state-of-the-art results.

Jiashuo Sun, Xianrui Zhong, Sizhe Zhou et al.

Proposes a quantitative measure (DEq(p)) for distracting effect, using synthetic and retrieval-based hard distractors to improve RAG accuracy by 7.5%.

Chen Amiraz, Florin Cuconasu, Simone Filice et al.

The clem:todd framework employs self-play with large language models (LLMs) to systematically benchmark task-oriented dialogue systems across architectures, datasets, and metrics, revealing performance trade-offs.

Chalamalasetti Kranti, Sherzod Hakimov, David Schlangen

Evaluated nine large language models' zero-shot performance on Bengali health query summarization; Mixtral outperformed others in ROUGE-1 and ROUGE-L.

Ajwad Abrar, Farzana Tabassum, Sabbir Ahmed

ZeroSearch enhances LLM search capabilities via simulated searches, with 14B model surpassing real search engines.

Hao Sun, Zile Qiao, Jiayan Guo et al.

Guiding LLM behavior using reward signals, employing techniques like RLHF and RLAIF for dynamic feedback learning.

Xiaobao Wu

RM-R1 formulates reward modeling as a reasoning task using chain-of-thought and Rubrics, outperforming larger models by up to 4.9%.

Xiusi Chen, Gaotang Li, Ziqi Wang et al.

This study analyzes length bias in LLM-based MT evaluation, proposing FSP and fine-tuning to improve consistency across input granularities.

Tobias Domhan, Dawei Zhu

DeepCritic introduces a two-stage framework leveraging Qwen2.5-72B-Instruct to enhance mathematical critique, outperforming GPT-4o and DeepSeek with significant accuracy gains.

Wenkai Yang, Jingwen Chen, Yankai Lin et al.