Scaled GenSelect and LLM-as-a-Judge to millions of tokens, enhancing math proof verification.

Sadegh Mahdavi, Branislav Kisacanin, Shubham Toshniwal et al.

GEM uses entropy-guided generative preference modeling for few-shot LLM alignment, improving preference prediction by 5-10% and task performance significantly.

Yiyang Zhao, Huiyu Bai, Xuejiao Zhao

PCRS-TKA employs tree-structured knowledge and PLMs, achieving 19.70% MRR@10 improvement over baselines.

Yongwen Ren, Chao Wang, Peng Du et al.

SparseAutoencoder-based lightweight reward model achieves high accuracy with less than 1% trainable parameters.

Dengcan Liu, Jiahao Li, Zheren Fu et al.

AIA Forecaster combines agentic search, reconciliation, and calibration to achieve expert-level forecasting performance.

Rohan Alur, Bradly C. Stadie, Daniel Kang et al.

Introduces Klear-Qwen3-AgentForge, combining supervised fine-tuning and multi-turn RL on Qwen3-8B, achieving state-of-the-art multi-task agentic performance.

Qi Wang, Hongzhi Zhang, Jia Fu et al.

PublicAgent employs a multi-agent architecture based on LLMs for open data analysis, validating five core design principles.

Sina Montazeri, Yunhe Feng, Kewei Sha

ExplicitLM introduces a million-scale external memory bank with a two-stage differentiable retrieval, achieving 43.67% improvement on knowledge tasks.

Chengzhang Yu, Zening Lu, Chenyang Zheng et al.

Proposes the Oversight Game framework using Markov Potential Games to ensure AI autonomy aligns with human safety.

William Overman, Mohsen Bayati

Confidence-based hybrid approach improves AI-human fact verification accuracy to 89.3%, outperforming individual ratings.

Rishub Jain, Sophie Bridgers, Lili Janzer et al.

FM Agent combines LLM reasoning with large-scale evolutionary search, achieving state-of-the-art results across multiple domains autonomously.

Annan Li, Chufan Wu, Zengle Ge et al.

APTBench evaluates base LLMs' agentic potential during pretraining via real-world trajectories, focusing on planning, action, and atomic skills.

Jiarui Qin, Yunjia Xi, Junjie Huang et al.

Proposes Huxley-Gödel Machine using CMP estimation to achieve human-level coding agent self-improvement.

Wenyi Wang, Piotr Piękos, Li Nanbo et al.

SSoT prompts LLMs with random strings and operations, achieving near-ideal distribution fidelity and enhanced diversity.

Kou Misaki, Takuya Akiba

Proposes WorldTest protocol to evaluate AI models' generality via environment-level queries; humans outperform AI in experiments.

Archana Warrier, Dat Nguyen, Michelangelo Naim et al.

Reinforcement learning enhances adaptive search strategies, improving accuracy by 15% on HotpotQA with multi-step decision models.

Minhua Lin, Zongyu Wu, Zhichao Xu et al.

Proposed NP-ENGINE framework integrates instance generation, rule verification, and heuristics, boosting LLMs' NP-hard optimization reasoning.

Xiaozhe Li, Xinyu Fang, Shengyuan Ding et al.

ToolPRM uses fine-grained step scoring to improve structured function calling inference, outperforming coarse reward models.

Jianghao Lin, Yuanyuan Shi, Xin Peng et al.

ProofFlow constructs dependency graphs for faithful proof autoformalization, achieving PROOFSCORE 0.545, surpassing baselines.

Rafael Cabral, Tuan Manh Do, Xuejun Yu et al.

PaperArena creates a challenging benchmark for tool-augmented cross-paper reasoning, with an average accuracy of 38.78%, dropping to 18.47% on hard subsets.

Daoyu Wang, Mingyue Cheng, Shuo Yu et al.