Dense retrievers favor passages naming answers their backbone already knows, study finds
_reachsumit · x · 2026-11-11
A new arXiv paper by Avigad, Ravfogel and Goldberg shows dense retrievers suffer from "search-by-answer bias": they prefer passages mentioning answer objects their backbone model already knows, even over passages supporting valid alternative answers — and even when the favored passage expresses a wrong relation.
- Controlled factual contrasts built from LAMA/TREx, measuring token-prediction-head scores for canonical vs alternative objects
- Across eight retrievers (DPR, Spider variants, Contriever, E5, Qwen3-Embedding), higher token scores consistently predict higher retrieval scores
- Implication: retrieval evidence selection is skewed by model priors, with real consequences for RAG evidence diversity
More from Research
- PufferLib: open-source RL library trains 60M samples/sec on one GPU — jsuarez · 2026-11-11
- Artificial Analysis Intelligence Index weights shift sharply across versions, chart reveals — HamelHusain · 2026-11-11
- Benchling's BenchBench-Refusals: GPT-6.1 Sol tops bio-safety refusals at 88.4%, unverified Opus 5.5 just 40.4% — nlarusstone · 2026-11-11
- Anthropic's top-cited AI scientists rack up 2x OpenAI's citations — NegotiationOk7535 · 2026-11-11
- MAEP: training-free modality-aware pruning cuts MoE-MLLM expert memory over 90% — nota-ai · 2026-11-11
- Fable 5.2 tops SemiAnalysis' GDPx chip design leaderboard at 16.4%, but some tasks saw no gains — ycombinator · 2026-11-11