Learned post-training quantization (PTQ) calibration tunes quantizer parameters on a small calibration set. For text embedders, the natural supervision signal is retrieval relevance, but the conditions under which this calibration improves or degrades performance are not fully understood.
The paper, titled "When PTQ Calibration Helps or Hurts: Dose-Driven Collapse and Validation-Gated Abstention in Quantized Text Embedders," asks two questions that are usually conflated, though the abstract is truncated before stating them explicitly. The title points to a "dose-driven collapse" phenomenon and a proposed "validation-gated abstention" strategy to mitigate it.
Given the partial abstract, the full methodology and results are not available, but the core contribution appears to be a framework for deciding when to trust calibrated quantized embedders and when to abstain from using them, based on validation performance.