Story · arXiv

The Recall Trap: A Recall-Maximizing Retriever Configuration Reduces Issue Resolution in Fixed-Budget Code Context (arXiv)

arXiv, 6 pp · Story page

On SWE-bench Verified with a fixed 12-slot context pack, the higher-recall retriever setting served the gold file more often (0.878 vs 0.806) yet resolved fewer issues. Turning off one-chunk-per-file dedup took gpt-5.6-sol from 39.2% to 46.8% single-shot; a pre-registered Qwen3.6-27B replication gained 3.6pp.

In plain words

  • A study found that showing a coding system the correct file more often did not make it fix more software issues.
  • The test gave each system only 12 pieces of project text, so every included piece displaced another.
  • The better setting showed fewer files but included more useful sections from each selected file.
  • Removing a rule that limited each file to one section improved successful fixes in both tested systems.
  • Coding assistant builders should judge search choices by completed repairs, because stronger file-finding scores can hide worse practical results.

Appeared in

Subscribe

Get the brief in your inbox

Pick daily, weekly, or both. Nothing is gated either way: every issue is on the site and in the feeds.

  • Weekdays at 8:45am IST, one lead story and 6 to 9 items.
  • Sundays, an argued synthesis rather than a recap.
  • One click to leave, and quiet days say so in the subject line.
How often

Weekdays 8:45am IST + Sundays. Unsubscribe in one click.

You're asking for The Agentic Brief by email at the cadence you picked. You can unsubscribe in one click from any issue, and your address is never sold or shared.