Aligning Medical Vision Language Models at the Capability Boundary with Entropy Adaptive Rubric Grounding

What it claims

MedCBR is a two-stage alignment recipe for medical VLMs. It first runs off-policy, difficulty-weighted DPO on samples the SFT model solves only some of the time. It then runs on-policy GRPO with clinical rubric rewards switched on by prediction entropy. The authors (Shanghai Jiao Tong University and Hunan University, arXiv October 2026) report 61.2% average accuracy across six medical VQA benchmarks covering radiology, pathology and general medical knowledge. The fetched text ends partway through the method section, so no per-benchmark table, baseline numbers or ablation values are available here.

How it works

Stage 1 (CBP, counterfactual boundary preference) has four steps:

1. For each training sample, draw N trajectories from the SFT policy and compute the empirical pass rate p against the verified answer.
2. Keep only samples with 0.2 ≤ p ≤ 0.8. The authors treat p near 1 as already mastered and p near 0 as noisy. With N=16, the selected set has Jaccard overlap above 0.92 with an N=128 reference set, so 16 samples per item is enough.
3. For each kept sample, pair the correct and incorrect trajectories that share the longest token-level common prefix. The pair then differs only after the diagnostic branch point, which is meant to stop the model learning stylistic preferences.
4. Train with DPO against π_sft, weighting each pair by w(p) = 1 − 2|p − 0.5|. The weight peaks at p = 0.5. Sampling happens once, offline.

Stage 2 (MRG, instantiated as EGPO on GRPO) distills a reusable rubric library from the Stage 1 trajectories. The method figure names four rubrics:

• Anatomy: structure identification and localization.
• Signs: findings and imaging signs.
• Differential: differential diagnosis reasoning.
• Uncertainty: limitations, alternatives and next steps.

Entropy is computed per rubric, and a gate selects a subset of rubrics for each sample. The selected rubrics combine with the verifiable answer reward into a composite reward, weighted by λ in the paper's objective. Gating is meant to keep rubric noise off samples the model already handles reliably. The exact gating rule falls past the truncation point.

The key coupling is the KL anchor. Stage 2's KL term references π_cbp, the Stage 1 policy, rather than π_sft.

flowchart LR
  SFT[SFT policy] --> S[N rollouts, pass rate p]
  S --> B{0.2 ≤ p ≤ 0.8?}
  B -- no --> X[discarded]
  B -- yes --> P[max-LCP pairs]
  P --> D[DPO, weight w p]
  D --> CBP[π_cbp]
  P --> R[rubric library]
  CBP --> G[GRPO + entropy-gated rubrics]
  R --> G
  CBP -. KL anchor .-> G

Evidence

The available text gives two results. The first is the 61.2% six-benchmark average. The second is the claim that the full method beats the sum of the two stages' individual gains (super-additivity), with the largest gains on fine-grained differential diagnosis. Neither result comes with benchmark names, baseline scores or a base model in the fetched text. The N=16 versus N=128 Jaccard figure is the only quantified design check shown.

Limits

• The fetched page stops at the Stage 2 figure, so results tables, ablations, the base VLM, model size, compute and λ/β values are not visible.
• No code, weights or licence is mentioned.
• The method assumes verified ground-truth answers for every training item.
• The authors state that clinician-facing integration, quantized inference and other deployment conditions are outside the evaluation, which covers medical QA only.