Search-G1: Grounded Search Agents via Representation-Based Intrinsic Rewards
- Source
- arxiv.org
- Author
- Cheng Ruoxi, Ma Haoxuan, Zhang Hongyi, Zhang Junming, Duan Ranjie, Xia Qiaolin, Wang Hao, Lu Yu, Shi Haibo, Ma Xingjun
- Date
Context
Search-G1 trains an AI to distinguish useful research from unnecessary searching. Its reward scheme estimates whether the model could answer without outside information and whether its answer actually depends on retrieved evidence. The second estimate is by checking what changes when evidence is removed. Correct answers receive additional credit for necessary, evidence-dependent searches, while repeated searches are penalized. The system periodically recalibrates these estimates as training changes the model. The authors report shorter answering trajectories with competitive accuracy across search-based question-answering tests. This targets a specific inefficiency: an can reach the right answer while still making searches that contribute little.
- calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
- 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.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
articleSearchAuditor: Auditing and Attributing Failures in Long-Horizon Search AgentsZhixiang Liang, Yifei Liu, Yidan Huang, Haozhe Zhao, Beichen Huang, Jiaqi Wang, Nan Duan, Qiong Cao
videoWhere RL Will Take Search — Maximilian-David Rumpf, SID.aiAI Engineer
articleCodeGrep: An RL-Trained Retrieval Agent for LLM Coding AgentsWuya Chen, Yihao yang, Yang Cao, Yue Lin
Checking sign-in…
Loading comments…