Embedding Surgery: query-time localized vector edits fix dense retrieval rankings, +60% nDCG@10
_reachsumit · x · 2026-09-07
A new arXiv paper introduces Embedding Surgery, a lightweight method for adaptive ranking correction in dense retrieval: instead of retraining or rebuilding static indexes, it applies small, localized updates to selected document embeddings at query time, guided by editorial labels, user interactions, or LLM pseudo-labels.
- Formulated as a convex optimization problem that enforces ranking constraints while minimally modifying affected document representations
- Preserves the global structure of the embedding space and runs at low computational cost
- Integrated into standard dense retrieval pipelines and evaluated on TREC Deep Learning, TREC Robust, TREC CAsT, and MS MARCO
- Consistent gains even under noisy or shifting feedback — up to +60.64% relative nDCG@10 on DL-Hard with editorial feedback
A practical path for search, recommendation, and RAG systems to absorb feedback without retraining.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA claims HGX Rubin NVL8 delivers 10x tokens per megawatt vs B200 — thursdai_pod · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27