Quantize by Drift: label-free mixed-precision quantization for text embedders hits 0.911 Spearman
_reachsumit · x · 2026-10-08
A new paper introduces Quantize by Drift, a label-free mixed-precision post-training quantization method for text embedders. Instead of relevance labels, it measures quantization-induced representation drift: quantize one module, re-encode the corpus, and record how far output embeddings move from full-precision positions. Across five embedders, configuration-level drift orders mixed-precision plans against held-out retrieval quality at macro Spearman 0.911, transports across corpora and domains, and requires one additive allocation under a packed-byte budget with no labels or search. Pre-registered checks hold 3/3, though drift is numerically below same-budget uniform precision at the main budget.
More from Research
- MarODE, a Markovian ODE framework for scoring LLM reasoning traces, accepted at TMLR — Tanmoy_Chak · 2026-10-08
- Empirical eval shows AI-written unit and integration tests don't improve agent success rates — GabGarrett · 2026-10-08
- OpenAI researcher speculates math AI could soon prove P != NP for good — markjeffrey · 2026-10-08
- Cryptographer: AI doing amazing math puts elliptic curve hardness in no danger — markjeffrey · 2026-10-08
- Temperature 0 doesn't make LLMs deterministic: 1,000 runs of Qwen3-235B yielded 80 outputs — lmoroney · 2026-10-08
- macro2mind Trains LLMs on Prediction Markets to Simulate Individuals, +15.5 Points Zero-Shot — youjiaxuan · 2026-10-08