MERGE: three small open LLMs plus auto-tuned prompts boost BM25 retrieval
_reachsumit · x · 2026-09-30
The MERGE framework uses three heterogeneous 7-8B open-source LLMs to independently produce query expansions, then a larger LLM synthesizes them into a single BM25 query. Prompt engineering is automated via an Automatic Prompt Optimization loop scored by downstream retrieval quality, running a tournament between champion prompts and optimizer drafts. It is retriever-agnostic—one BM25 pass, no rank fusion, no re-indexing—and improves BM25 across five BEIR benchmarks (NQ, SciFact, FiQA, Touche-2020, DBPedia).
More from Research
- CycleDistill: synthetic-data self-distillation lifts Indic MT by 20+ chrF in one cycle — cneuralnetwork · 2026-10-25
- ICLR 2027 still issuing desk rejects as community battles AI slop submissions — Temporary_Switch_339 · 2026-10-25
- TEMPO benchmark: best of 16 retrievers hits only 32.0 NDCG@10; RAG with no retrieval scores higher — mageed · 2026-10-25
- Random speckle spectrometers make some of the best-looking photonic chips — jwt0625 · 2026-10-25
- Studies show the same artwork scores lower when labeled AI-made — bias, not quality — AIandDesign · 2026-10-25
- Anthropic uses Claude to formally verify 100+ real-world open-source libraries in weeks — tianyin_xu · 2026-10-25