跳到正文
arXiv cs.CL· Zhexi Feng, Ruiyi Zhang, Yongbo Yang, Pengtao Xie·· 10 小时前AI 评分38

面向编码智能体的状态条件最小充分证据:MSS-Complement 与 SERBench 基准

The Missing Complement: State-Conditioned Minimal Sufficient Evidence for Coding Agents

AI 导读

研究提出“状态条件最小充分证据恢复”任务,并发布 SERBench 基准,用 45 个仓库的 500 个智能体状态衡量证据集合是否满足当前决策的全部标注需求。

正文 · 原文

View PDF HTML (experimental)

Abstract:Retrieval assembles repository context by ranking passages for relevance to the current query. A coding agent halfway through an issue has already read much of what such a ranker returns. Relevance is scored per passage, but sufficiency belongs to the set: independently scored passages can fill the budget with support for one requirement while another goes unmet. We formulate state-conditioned minimal sufficient evidence recovery: given a captured agent state, recover a compact evidence combination supplying what its next decision still lacks. SERBench measures this on 500 held-out states from 45 repositories, recording what the agent has seen, crediting only sets that satisfy every annotated evidence requirement of the current decision, and separating set recovery from candidate discovery. MSS-Complement treats acquisition as set construction, not ranking. Three semantic calls propose a jointly sufficient set, search for what it lacks, and return 4-8 intact source units within 6,144 tokens. One configuration, fixed on calibration data, recovers a complete set for 73.0% of those states at five items and 80.6% at eight, against 61.4% and 72.4% for Qwen3 embedding with reranking. A matched control ranking by similarity alone recovers fewer complete sets, placing the margin over it in the set-level policy, not the computation. The lead persists from frozen repository source with no gold-derived pool. On AMA-Bench it answers from a 76.2% smaller answer prompt, with accuracy 2.08 points above that benchmark's own memory agent. Removing one required group from a complete set costs repair-localization precision under two executors. Retrieval for agents is better posed as recovering what a decision lacks than re-ranking what an issue resembles.
Comments: 32 pages, 3 figures. Benchmark and evaluation resources: this https URL
Subjects: Information Retrieval (cs.IR); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Cite as: arXiv:2609.20050 [cs.IR]
  (or arXiv:2609.20050v3 [cs.IR] for this version)
  https://doi.org/10.48550/arXiv.2609.20050

arXiv-issued DOI via DataCite

Submission history

From: Zhexi Feng [view email]
[v1] Thu, 17 Sep 2026 10:56:47 UTC (8,964 KB)
[v2] Sat, 26 Sep 2026 17:31:40 UTC (6,065 KB)
[v3] Thu, 19 Nov 2026 12:47:39 UTC (5,894 KB)

来源:arXiv cs.CL · arxiv.org