arXiv：cs.AI· Ruoxi Cheng, Haoxuan Ma, Hongyi Zhang, Junming Zhang, Ranjie Duan, Qiaolin Xia, Hao Wang, Yu Lu, Haibo Shi, Xingjun Ma·· 6 小时前
Search-G1：面向选择性接地搜索的共演化表征奖励框架
Search-G1: Co-Evolving Representation-Based Rewards for Selective Grounded Search
AI 导读
Search-G1 是一种将检索必要性与证据依赖耦合的奖励框架，通过两个干预校准的隐藏状态读取实现选择性接地搜索，无需过程标注或 LLM 评判推理即可打分。该框架在四个 QA 基准和两种规模上取得有竞争力的任务准确率，在主要对比中展现出更强的反事实证据依赖与更少的搜索次数。
正文
Abstract:Search-augmented language agents should retrieve external information only when necessary and ground their answers in retrieved evidence. Yet terminal correctness assigns equal credit to direct answers from parametric knowledge, evidence-grounded answers, and answers recalled after redundant search. Process supervision and LLM judges provide richer feedback but can require high-cost annotation or inference; low-cost, non-contrastive policy confidence signals do not directly identify whether retrieved evidence changes an answer. We propose Search-G1, a reward framework that couples retrieval necessity with evidence reliance through two intervention-calibrated hidden-state readouts for selective grounded search. A prompt readout predicts closed-book sufficiency, whose complement proxies policy-relative retrieval necessity; an answer-commit readout predicts evidence reliance as sensitivity to evidence deletion conditional on the realized reasoning prefix. A correctness-first reward gates evidence-reliance credit by retrieval necessity, rewards sufficient direct answers, and penalizes repeated search. After calibration, scoring needs neither process annotations nor LLM-judge inference. As policy-relative targets, trajectory distributions, and representations may drift, Search-G1 periodically regenerates targets and refits its readouts. Across four QA benchmarks and two scales, Search-G1 achieves competitive task accuracy, with stronger counterfactual evidence dependence and fewer searches in the main comparisons, complemented by independent human audits of answer support.
| Subjects: | Computation and Language (cs.CL); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2608.07531 [cs.CL] |
| (or arXiv:2608.07531v4 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2608.07531 arXiv-issued DOI via DataCite |
Submission history
From: Rosy Cheng [view email]
[v1]
Fri, 24 Jul 2026 15:22:57 UTC (4,390 KB)
[v2]
Tue, 18 Aug 2026 05:20:56 UTC (4,390 KB)
[v3]
Fri, 4 Sep 2026 09:19:36 UTC (1 KB) (withdrawn)
[v4]
Sat, 10 Oct 2026 14:57:44 UTC (4,427 KB)
来源：arXiv：cs.AI · arxiv.org