The Missing Complement: State-Conditioned Minimal Sufficient Evidence for Coding Agents
The paper argues coding-agent retrieval should assemble the missing evidence set, not just rank similar passages.
The authors define a state-conditioned task: given what an agent has already seen, recover the compact source evidence still needed for its next decision. Their SERBench benchmark tests 500 held-out states across 45 repositories, scoring only evidence sets that cover every required fact. MSS-Complement uses three semantic calls to propose, check, and return 4-8 intact source units within 6,144 tokens, beating embedding-plus-reranking baselines at five and eight items. The paper also reports gains on AMA-Bench with a much smaller answer prompt. ArXiv · AI/CL/LG's note
The authors define a state-conditioned task: given what an agent has already seen, recover the compact source evidence still needed for its next decision. Their SERBench benchmark tests 500 held-out states across 45 repositories, scoring only evidence sets that cover every required fact. MSS-Complement uses three semantic calls to propose, check, and return 4-8 intact source units within 6,144 tokens, beating embedding-plus-reranking baselines at five and eight items. The paper also reports gains on AMA-Bench with a much smaller answer prompt. ArXiv · AI/CL/LG's note
score 5