// AI news · 2026-11-16 · Dev.to LLM
Decomposition retrieval worsened; simple approach better
Decomposing job requirements for retrieval degraded performance, with MRR dropping from 0.47 to 0.63.
Published 2026-11-16 00:29 (Beijing time)Source Dev.to LLMAuto-summarised, fact-checked against source
01Summary
Removing decomposition restored MRR to 0.92, showing that splitting queries lost semantic context. A simple budget-check rule improved bid decision F1 from 0.89 to 0.98. Scam detection was misapplied to injected instructions; fixing by scanning only non-instructional text eliminated false rejections.
Summarised from the original article by an open-source model, then checked automatically: every number and English name must appear in the source. The original article is authoritative.
02Key points
- Decomposition reduced MRR from 0.47 to 0.63, losing conjunctive query semantics.
- MRR recovered to 0.92 when decomposition was removed, indicating poor retrieval quality.
- A simple budget check improved bid F1 from 0.89 to 0.98.
- Scam detection misflagged injected instructions; fixing by excluding instruction contexts resolved the issue.
03Source
The clever retrieval trick lost. A boring if statement won.
Dev.to LLM · published 2026-11-16 00:29 Beijing time
05Related news
- Same source · Miktam Bible Part 2: On-Device AI Study Assistant2026-11-16 · Dev.to LLM
- Same source · ModelFold Updates with 703 Multimodal Model Entries2026-11-16 · Dev.to LLM
- Same topic · 2026 GOTY Race Could Harm Indies2026-11-16 · Polygon
- Same topic · Nvidia CEO Resignation Sparks China's AV Backup Planning2026-11-16 · TechCrunch
- Same day · AI Tools Compare in Detecting Amazon Scam Emails2026-11-16 · Android Police
From AIMap AI news: collected from 1,100+ official and media sources. Summaries only restate the original article; follow the link for the full story. All stories of 2026-11-16 · Archive by date · All AI news · Today