Token-level matching boosts memory retrieval by 13.3 points with zero training
_reachsumit · x · 2026-10-12
The paper "Beyond the Final Embedding" improves memory retrieval without any training on the encoder:
- Core idea: Most embedding retrievers compare single vectors of queries and memories, even when relevance hinges on a small local piece of content. This method runs sparse matching between each query token and token states from a frozen embedding model, aggregates match scores per memory, and fuses them with whole-memory similarity.
- Results: Gains across 8 benchmarks and multiple model families. With an optional low-rank query adapter (encoder and stored memories stay fixed), average Recall@10 and NDCG@10 improve by 13.3 and 14.3 points on Qwen3-Embedding-4B.
- Downstream: Retrieval gains translate into higher QA accuracy on long-term memory benchmarks.
More from Research
- 1,200 AI answers show Reddit citations are nearly all from Google engines: Gemini 52.4%, ChatGPT 0% — drrickio · 2026-10-12
- Open experiment log pushes integer multiplication bounds down to κ≈0.001467, invites agent collaboration — IgorCarron · 2026-10-12
- LLMs don't guess, they optimize: a statistical mechanics take claims 3,418 gradient steps to coherence — TinfoilTricorn · 2026-10-12
- A critical look at Lean4: why no one can prove the proof assistant is consistent — thomasahle · 2026-10-12
- ThinkDrag: mmlab-ntu adds visual reasoning to drag-based image editing, claims SOTA — mmlab-ntu · 2026-10-12
- Dex-One2Many learns dexterous manipulation from one human video, +71% on unseen setups — Jusuk Lee · 2026-10-12