Study: Higher recall in code retrieval drops GPT-5.6 repair success by 7.6pp

Retrieval components for code assistants are usually tuned against retrieval metrics like recall@k, under the assumption that higher recall means better downstream task performance. In a controlled study on SWE-bench Verified, Alexander Adkins and Teimuraz Trapaidze show this assumption can fail. They injected retrieved chunks as a fixed 12-slot context pack with no search tools, toggling a single flag: one-chunk-per-file deduplication. The flag on produced higher recall (gold file in 87.8% of served packs vs 80.6% off), yet the lower-recall configuration — which trades file breadth for within-file depth — significantly improved single-shot resolve rates: gpt-5.6-sol rose +7.6pp (39.2% to 46.8%, n=500, McNemar p=0.0003) and a pre-registered open-weights replication with Qwen3.6-27B rose +3.6pp (n=499, p=0.0133). Both results survived repository-clustered inference, and a random-chunk control ruled out an argmax-selection artifact.

The gain scales with within-file anchor dose, but the effect is not universal. On a lexical BM25 retriever it reverses (-3.2pp, significant cross-paradigm interaction), and it disappears when agents have unrestricted Read access (a powered null). Across four languages in SWE-PolyBench (N=617), the improvement was positive but not significant (+2.6pp, p=0.056), marking a boundary rather than a confirmed extension. Operationally, at a tight fixed budget, the authors recommend not hard-deduplicating by file and A/B testing packing policies against the actual task, not the metric the flag was tuned to. The paper includes a reproducible artifact and is published on arXiv with DOI 10.48550/arXiv.2608.14838.