Simplified on-policy supervised fine-tuning (SFT) with truncation-based length reward improves reasoning efficiency and training speed, outperforming complex RL methods.

Anhao Zhao, Ziyang Chen, Junlong Tong et al.

SkillsBench benchmark shows curated skills improve task pass rate by 16.6 percentage points.

Xiangyi Li, Yimin Liu, Wenbo Chen et al.

Introduces Min-kNN distance for black-box detection of RLVR training data via structural convergence, achieving 70% AUC.

Hongbo Zhang, Yue Yang, Jianhao Yan et al.

Introduces GameDevBench, a benchmark for AI in complex game development, with a success rate of 53.8%.

Wayne Chi, Yixiong Fang, Arnav Yayavaram et al.

Introduced Epistemic Ledger to diagnose and reduce illusory completion in search agents, lowering underverification by up to 26.5%.

Dayoon Ko, Jihyuk Kim, Sohyeon Kim et al.

ANCHOR uses branch-point expansion to generate diverse GUI trajectories, improving model fine-tuning success rates by over 4% on benchmarks.

Jinbiao Wei, Yilun Zhao, Kangqi Ni et al.

AgentCPM-Explore achieves 97.09% GAIA accuracy via parameter merging, reward denoising, and context refinement in a 4B model.

Haotian Chen, Xin Cong, Shengda Fan et al.

This survey organizes graph-based agent memory around extraction, storage, retrieval, and evolution, but reports no unified benchmark or quantitative gain.

Chang Yang, Chuang Zhou, Yilin Xiao et al.

Research-level math questions test AI's autonomous reasoning; framework involves encrypted questions, external resource access, and human validation.

Mohammed Abouzaid, Andrew J. Blumberg, Martin Hairer et al.

The paper proposes a novel uncertainty quantification framework for interactive LLM agents, addressing uncertainty in multi-step reasoning.

Changdae Oh, Seongheon Park, To Eun Kim et al.

LABBench2 enhances AI's real-world biology research capabilities with 1,900 tasks, models' accuracy drops 26%-46%.

Jon M Laurent, Albert Bou, Michael Pieler et al.

Agent-Omit enhances LLM agent efficiency by adaptively omitting context; experiments show comparable performance to frontier methods.

Yansong Ning, Jun Fang, Naiqiang Tan et al.

DeltaEvolve employs structured semantic delta and multi-level database to improve program evolution efficiency, reducing token use by 36.79%.

Jiachen Jiang, Tianyu Ding, Zhihui Zhu

Proposes a multi-step reasoning framework with background stories to improve knowledge internalization and transfer in LLMs, boosting multi-hop reasoning by 10-20%.

Ya Gao, Kalle Kujanpää, Pekka Marttinen et al.

ROMA plus GEPA+ lifts SEAL-0 accuracy by 9.9% over Kimi-Researcher.

Salaheddin Alzu'bi, Baran Nama, Arda Kaz et al.

PCBSchemaGen uses reward-guided structured verification for training-free PCB schematic generation, achieving 81.3% success rate across 227 real IC tasks.

Huanghaohe Zou, Peng Han, Emad Nazerian et al.

Proposes TABROUGE, a training-free, deterministic LCS-based reward, improving table reasoning accuracy by 26.7 points.

Tung Sum Thomas Kwok, Xinyu Wang, Hengzhi He et al.

Using Gemini Deep Think, the study evaluated 700 'Open' Erdős problems, combining AI natural language verification with expert review, revealing problem status and AI limitations.

Tony Feng, Trieu Trinh, Garrett Bingham et al.

Proposes SCMA, a multi-agent RL framework, compresses reasoning chains by 39%, boosts accuracy by 10%.

Yiqun Chen, Jinyuan Feng, Wei Yang et al.

WebArbiter is a principle-guided reasoning reward model using structured text generation, surpassing scalar/template methods with 9.1% improvement on BoN accuracy.

Yao Zhang, Shijie Tang, Zeyu Li et al.