The paper addresses a problem in : when an agent is partway through a task, it has already seen some information, but may still lack specific facts needed for its next decision. Traditional retrieval re-ranks passages by relevance, but this can produce redundant or incomplete evidence. The authors propose MSS-Complement, which treats evidence acquisition as set construction. It uses three semantic calls to propose a sufficient set, find missing pieces, and return 4-8 intact source units within a token budget. On a new benchmark (SERBench) with 500 held-out states, a fixed configuration recovers a complete evidence set for 73.0% of states at five items and 80.6% at eight, outperforming a baseline using Qwen3 embedding with reranking (61.4% and 72.4%) and a similarity-only ranking (66.6%). The approach also shows transferability to AMA-Bench, achieving higher accuracy with a 76.2% smaller answer prompt.
Coding agents retrieve better by targeting missing evidence sets
A state-aware retrieval method assembled complete evidence for coding decisions more often than passage-by-passage ranking baselines.
Academic
Zhexi Feng · Ruiyi Zhang · Yongbo Yang · Pengtao Xie
University of California San Diego
Research Digest··2 min read
In one line
Retrieval for coding agents should recover a minimal set of evidence that completes what the agent still lacks, not re-rank relevant passages.
Before you read this
Three ideas this paper builds oncoding agentAn AI-powered software system that autonomously writes, modifies, or extends source code based on high-level instructions.157 papers →repository-level benchmarkAn evaluation suite that measures agent performance on tasks spanning entire codebases, from construction to repair and governance.45 papers →agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.229 papers →
Read it as
Feng et al. frame retrieval for coding agents as finding the smallest set of evidence that complements what an agent has already seen.
The authors built SERBench from 500 held-out coding-agent states across 45 repositories.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.