Proposes Ling-2.6 and Ring-2.6 models using hybrid linear attention and transfer pretraining, boosting long-context efficiency and reasoning.

Ang Li, Ben Liu, Bin Han et al.

Under fixed inference budgets, vanilla actors often outperform or match augmented methods in web agent tasks, with multi-run variance highlighting evaluation challenges.

Sina Hajimiri, Masih Aminbeidokhti, Jose Dolz et al.

AC-GPT introduces a simple modification enabling arbitrary conditional evaluation within a single pass, preserving left-to-right prediction.

Yinhan Lu, Eric Elmoznino, Léo Gagnon et al.

Proposes a method combining lexical token analysis and multi-word similarity to quantify cultural localization in AI-generated stories, revealing only 9-17% of vocabulary accounts for cultural differences.

Shaily Bhatt, Supriti Vijay, Jeremiah Milbauer et al.

Introduces operads as a formal framework for question decomposition, with operadic consistency correlating strongly with model accuracy across multiple datasets.

Nathaniel Bottman, Kyle Richardson

NTS-CoT reduces hallucinations in news timeline summarization using Chain-of-Thought, improving AR-1 by 23.4%.

Feng Lyu, Huiqin Yan, Sijing Duan et al.

SKIM framework compresses procedural knowledge in LLMs, maintaining task performance with 30%-60% compression rate.

Changyue Wang, Weihang Su, Qingyao Ai et al.

This study evaluates decoding-time truthfulness methods (layer-contrast, intervention, logit adapters) on instruction-tuned LLMs, finding limited effects under strict controls, with reasoning prompting more robust.

Ao Sun

H-SAL uses self-description text for implicit debiasing, outperforming explicit-label debiasing.

Shun Shao, Zheng Zhao, Anna Korhonen et al.

Proposes Arbor framework with Hypothesis-Tree Refinement for autonomous research, achieving over 2.5× improvement across tasks.

Jiajie Jin, Yuyang Hu, Kai Qiu et al.

VISTA introduces a hybrid user simulator combining UI and API actions, with six metrics for realism and coverage, outperforming existing methods in diverse scenarios.

Yunan Lu, Ryan Shea, Yusen Zhang et al.

KATE framework enhances LLM tool use by integrating experiential knowledge and expanding reasoning width, achieving a 15% improvement on BFCL-V3.

Yupu Hao, Zhuoran Jin, Huanxuan Liao et al.

Predictor-gated bank-wise sparsity converts dense LLMs into 4× sparse models, improving long-text efficiency.

Ruixuan Huang, Jinyuan Shi, Hantao Huang et al.

REAL constructs a temporal, confidence-aware graph, boosting long-term memory by 22.72% in LLMs.

Keer Lu, Liwei Chen, Guoqing Jiang et al.

UniSVQ combines scalar and vector quantization via affine transforms, achieving near VQ performance with minimal storage overhead.

Haoyu Wang, Haiyan Zhao, Xingyu Yu et al.

This paper reveals that RLHF achieves shallow alignment by compressing partisan signals without removing the underlying partisan structure, as shown through internal representation analysis of Llama 3.1 8B.

Wendy K. Tam

AdvGRPO framework combines dense multi-channel rewards and advantage decoupling for joint attacker-defender training, achieving over 90% attack success rate and superior defense robustness.

Blake Bullwinkel, Eugenia Kim, Amanda Minnich et al.

SG-OPD introduces sign-consistency gating and phased teacher sampling, improving mathematical reasoning accuracy by 1.98 and 7.50 points over standard OPD.

Haoran Xu, Hongyu Wang, Yifei Gao et al.

RadOT-Eval employs structured evidence optimal transport for radiology report evaluation, achieving a Spearman correlation of 0.715 with total errors.

Weixin Liu, Juming Xiong, Yang Li et al.

SAPO framework optimizes policy by validating skill marginal utility, improving skill quality and performance.

Zhiwei Zhang, Yudi Lin, Nikki Lijing Kuang et al.