Learned post‑training quantization (PTQ) calibration fits quantizer parameters on a small calibration set, and for a text embedder the natural supervision signal is retrieval relevance. We separate two often conflated questions: how far to calibrate and whether to calibrate at all.
Across three embedders (e5‑base‑v2, BGE‑M3, EmbeddingGemma‑300m) with 4‑bit weight quantization we performed 161 calibration runs. A fixed calibration procedure produced outcomes ranging from performance gain to catastrophic degradation, driven primarily by the update dose measured as the functional displacement $D_{\text{func}}$ of the quantized model from its uncalibrated start.
On EmbeddingGemma the same learning‑rate grid yielded an in‑domain NDCG@10 change from +1.47 to -71.53 points; a label‑free cosine objective collapsed at the same dose (‑69.52 vs ‑71.63), indicating that collapse is a property of learnable calibration rather than of retrieval supervision.
Objective sensitivity is heterogeneous: on e5, ranking calibration is substantially more robust than cosine at matched displacement (+14.16 ± 4.34 pp, 95% t‑CI [+3.37, +24.96], three paired slates), whereas on BGE‑M3 the ordering reverses but the difference is unresolved ([-5.18, +2.38]). Dose is the main axis but not a sufficient statistic; equal $D_{\text{func}}$ reached via different paths can differ by up to 1.61 pp.
We therefore treat calibration as an optional intervention. A pre‑registered rule accepts a checkpoint only when a per‑query paired confidence interval on a 25% calibration‑validation split clears the uncalibrated baseline. Applied to six independent slates and five arms, the rule never selected a harmful checkpoint (0/30 decisions), retained 8 of 11 beneficial updates, incurred a mean oracle regret of 1.52 pp, and one abstention forfeited 11.13 pp, illustrating a safety‑opportunity trade‑off rather than a free improvement.
Open issues remain, including label efficiency.
Review