RAG collapse: LLM answers converge when retrieving self-authored content, 79.6% simulations collapse
_reachsumit · x · 2026-08-25
A new paper shows that when RAG systems retrieve content they authored themselves, responses collapse toward near-identical answers, and even a single self-authored reference can trigger it. Across 1,528 simulations with three model families and 1,019 prompts, 79.6% ended in collapse. The self-bias persists even after controlling for reference quality.
More from Research
- AAAI 2027 review debate: Should empirical papers without code be auto-rejected? — SimpleObvious4048 · 2026-08-25
- Study asks: Are LLM agents time-aware and budget-conscious? — maksym_andr · 2026-08-25
- SA-RSQ: Sparse Representation Framework for Multi-modal Recommender Systems — _reachsumit · 2026-08-25
- Revisiting N2DCG: Empirical Reformulation for Carousel Recommendation — _reachsumit · 2026-08-25
- Semantic subword tokenization improves generative recommenders by reducing intra-item attention overload — _reachsumit · 2026-08-25
- Spotify study: better reasoning traces can hurt recommender effectiveness — _reachsumit · 2026-08-25