Belief-Calibrated Optimization: An Explicit World Model for Agentic Optimization
- Source
- arxiv.org
- Author
- Yuhan Chen, Zhihua Tian, Mahavir Dabas, Charith Peris, Rahul Gupta, Ming Jin, Feiyang Kang, Siyuan Zhang, Nan Wang, Ruoxi Jia
- Date
Context
An AI coding that improves another agent's supporting software can preserve its explanations of earlier successes and failures. Belief- Optimization writes those explanations into a persistent document and revises it after each evaluated change. The document gives later optimization attempts an account of how changes affect results. The researchers report higher pass rates than a matched version without that document across five benchmarks, including held-out tasks.
A separate test found that the accumulated document helped a fresh predictor anticipate outcomes better than either no document or a version containing falsified explanations. That supports the value of retaining useful experimental knowledge between attempts.
- AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
- calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
articleAutoWorldModel-Bench: A State-Centric Benchmark for Automated World-Model ResearchMarjan Moodi, Xuankang Zhu, Fernando De Mesentier Silva, Harold Chaput, Mohammad Reza Taesiri
articleThe Double Measurement Confound in Agent Benchmarks: De-Scaffolding, Ground-Truth Scoring, and Reliability Beyond the MeanYonghong Zhang, Shadi Motaali, Vu Phong Dinh, Avin Piroutiniya, Jorge E. L\'opez de Vergara, Luis de Pedro, Ricardo Correia, Isabel M. Parra, Yong Xie
articleCoaching Qwen3 Coder 30B to Think Like a CodeClash Arena AgentIvy Ning Zhang
Checking sign-in…
Loading comments…