ResMem distills retrieved evidence into parametric memory, cutting RAG storage by 94.7%
_reachsumit · x · 2026-11-04
A new arXiv paper (2611.00450) introduces ResMem, which distills previously retrieved evidence into a parametric memory for frozen LLMs so online retrieval can be dropped at inference.
- The LLM reads retrieved text offline to generate teacher targets; a trained memory learns an additive logit residual aligning student predictions with the teacher distribution, with self-distillation confined to the memory module
- Unlike kNN-based parametric memory (e.g. MLP Memory), it avoids storing massive token representations, cutting teacher-construction storage by 94.7% on a Wikipedia subset
- Across five QA datasets, mean token F1 rises from 32.83% to 34.48% over the frozen baseline
More from Research
- François Chollet: modeling brain wiring as an engineering problem constrained by genome size and time — CatAstro_Piyush · 2026-11-04
- OffsecCyber dataset combines HackerOne reports and CVEs to train pentesting LLMs — AccBalanced · 2026-11-04
- Monofact rate in training data governs LLM scientific discovery, preprint finds — CatAstro_Piyush · 2026-11-04
- qwenTPU compiles a TPU into Qwen3.5 weights, achieving exact integer matmul — CatAstro_Piyush · 2026-11-04
- Jev model improves ICL demonstration retrieval via distillation into a lightweight MLP — kalyan_kpl · 2026-11-04
- Indie researcher open-sources 10 years of math work: 146k files, AI-accelerated over 3 years — HamOnWhy · 2026-11-04