Quantization Preserves the Answer but Loses the Evidence: The "Proxy Objective Trap" in Medical LLMs
Translation and analysis of the paper "When Quantization Preserves Accuracy but Not Evidence: Explanation-Aware Post-Training Quantization for Medical LLMs" (University of Alberta).
A clinical scenario
A doctor uses an LLM to assist diagnosis. A 60-year-old male patient presents with chest pain and sweating; the ECG shows ST elevation. The model answers "acute myocardial infarction," citing "ST elevation suggesting transmural ischemia." The doctor nods — correct answer, correct reasoning, trustworthy.
Now the model is quantized for edge deployment — weights compressed from 16-bit to 4-bit. The quantized model still answers "acute myocardial infarction." Accuracy is unchanged. But the rationale changes: "Patient age over 50, chest pain is a common symptom, consider acute MI."
The answer is still right, but the reasoning shifted from the key clinical evidence (ST elevation) to generic factors (age + chest pain). If the doctor reads the rationale to understand the model's decision basis, they may wrongly conclude that "age + chest pain" suffices to diagnose MI — clinically problematic.
Core finding: accuracy and rationale support decouple
| Token type | Share in rationale | Share in quantization loss (KL divergence) | |---|---|---| | Generic/discourse | 75.3% | 83.1% | | Clinical cues | 22.2% | 15.6% | | Numeric cues | 0.67% | 0.38% | | Answer/option words | 1.80% | 0.99% |
The distribution of quantization loss across token types does not match the token distribution itself. Generic discourse tokens are over-protected; sparse but critical clinical evidence tokens are relatively neglected. Standard PTQ objectives (minimizing KL divergence) primarily protect tokens that keep sentences fluent, not tokens that make rationales credible.
Why this is a "proxy objective trap"
Standard PTQ optimizes reconstruction error between the quantized and full-precision models, with token-level KL divergence as the proxy. This proxy is reasonable in most cases — but its blind spot is that all tokens are treated equally, while their contributions to rationale trustworthiness differ enormously:
- A wrong "therefore" does not affect credibility.
- A wrong "ST elevation" destroys the whole rationale.
- OSTQuant baseline: preserves task accuracy, but answer-rationale support degrades significantly.
- Explanation-aware method: retains accuracy while better preserving answer choice and rationale-answer consistency.
- RAG retrieval relevance ≠ trustworthiness
- V-JEPA pixel reconstruction ≠ rotation understanding
- Moral label aggregation ≠ moral truth
- PTQ accuracy ≠ rationale support
- The Faithfulness Cache depends on full-precision model quality; correct answers can still come with flawed rationales.
- Evidence-token identification (KL-based clustering) involves heuristics and hyperparameters.
- The method is instantiated on OSTQuant only; applicability to GPTQ, AWQ, SmoothQuant, etc. is untested.
- Only medical QA is evaluated; legal/scientific/code-explanation settings are untested, though the idea should generalize.
- For deployment engineers: don't evaluate quantized models by accuracy alone in explanation-critical settings — check rationale-answer consistency.
- For AI evaluation researchers: another clean instance of the proxy objective trap.
- For medical AI: rationale-answer consistency should become part of standard evaluation; if quantization quietly changes the "why" even when the "what" is unchanged, clinical trustworthiness must be reassessed.
The proxy (token-level KL) aligns with the true objective (rationale-answer consistency) on the training distribution, but fails under quantization-induced distribution shift.
Method: Faithfulness Cache and explanation-aware losses
Step 1: Build a Faithfulness Cache
Generate rationales from the full-precision (teacher) model on calibration samples, identify the evidence tokens that actually support the answer, and record them with the answer behavior — only including samples the full-precision model answered correctly.
Step 2: Explanation-aware PTQ objectives
Building on OSTQuant (a transformation-based PTQ method), add two loss terms:
1. Evidence token protection loss: upweight KL divergence on cached evidence tokens. 2. Rationale-conditioned answer behavior loss: protect the answer prediction distribution conditioned on the evidence tokens, preserving the evidence-to-answer mapping.
Experiments: W4A4KV4 medical QA
Setup: W4A4KV4 quantization, four 7B–8B medical and instruction-tuned models, benchmarks MedExQA, MedExpQA, ChallengeClinicalQA.
The structure of this result matters: accuracy-only evaluation hides the problem; only rationale-answer consistency metrics expose it.
Why this matters for medical AI evaluation
Most medical LLM evaluation uses answer accuracy, implicitly assuming that a correct answer implies trustworthy reasoning. This paper shows that assumption is fragile: the quantized model may keep accuracy and fluency while silently swapping the evidence ("ST elevation" → "age over 50"). The recommendation: in explanation-critical settings, PTQ evaluation should include rationale-answer consistency metrics, not just accuracy.
Broader implications: proxy failure is a recurring pattern
Same structure each time: a measurable proxy aligns with a hard-to-measure true objective on the training distribution, then fails under distribution shift.
Honest assessment: limitations
A deeper question: what counts as "evidence"?
The paper implicitly assumes rationale tokens split into "evidence" vs. "discourse." This clustering-based distinction is a practical approximation; a more principled definition might come from counterfactual reasoning — which tokens, if replaced, significantly change the answer distribution? An open question for follow-up work.
Conclusion
This is a diagnostic paper: it identifies an overlooked problem (quantization preserves accuracy but not evidence), proposes a practical fix (explanation-aware PTQ), and exposes a broader evaluation blind spot.
Paper: When Quantization Preserves Accuracy but Not Evidence: Explanation-Aware Post-Training Quantization for Medical LLMs arXiv: https://arxiv.org/abs/2609.24799 Code: https://github.com/dut0817/EAQuant Affiliation: University of Alberta