The Missing Complement: State-Conditioned Minimal Sufficient Evidence for Coding Agents
Ranking
Overall
78
Content
95
Popularity
39
Observed public metrics from 1 member.
Merged summary
TL;DR - This paper reframes retrieval for coding agents as finding a compact, state-conditioned set of evidence that supplies every fact missing from the agent’s current context. Its MSS-Complement method improves evidence completeness and downstream repair localization while using substantially smaller prompts.
- SERBench evaluates retrieval on 500 held-out agent states from 45 repositories, requiring retrieved sets to cover every annotated fact needed for the next decision.
- MSS-Complement constructs jointly sufficient evidence sets through three semantic calls and returns 4–8 intact source units within 6,144 tokens.
- It achieves 73.0% complete-set recovery with five items and 80.6% with eight, versus 61.4% and 72.4% for Qwen3 embeddings with reranking.
- On AMA-Bench, it reduces answer-prompt size by 76.2% while improving accuracy by 2.08 points over the benchmark’s memory agent.
Sources (1)
The Missing Complement: State-Conditioned Minimal Sufficient Evidence for Coding Agents
Public signals
Semantic Scholar citations 0 · Semantic Scholar influential citations 0
TL;DR - This paper reframes retrieval for coding agents as finding a compact, state-conditioned set of evidence that supplies every fact missing from the agent’s current context. Its MSS-Complement method improves evidence completeness and downstream repair localization while using substantially smaller prompts.
- SERBench evaluates retrieval on 500 held-out agent states from 45 repositories, requiring retrieved sets to cover every annotated fact needed for the next decision.
- MSS-Complement constructs jointly sufficient evidence sets through three semantic calls and returns 4–8 intact source units within 6,144 tokens.
- It achieves 73.0% complete-set recovery with five items and 80.6% with eight, versus 61.4% and 72.4% for Qwen3 embeddings with reranking.
- On AMA-Bench, it reduces answer-prompt size by 76.2% while improving accuracy by 2.08 points over the benchmark’s memory agent.