Med Tiv Eval

med-tiv-eval

Scaling Medical Reasoning Verification via Tool-Integrated Reinforcement Learning — Zhang et al. (2026) (arXiv:2601.20221, 2026)

What this evaluates

Evaluates a verifier model's ability to distinguish correct from erroneous reasoning traces in medical question-answering tasks. It measures how well tool-integrated reinforcement learning improves factual justification and reduces hallucination compared to static reward models.

Datasets

• MedQA — total ?; splits: test (-1)
• MedMCQA — total ?; splits: test (-1)
• MMLU-Med — total ?; splits: test (-1)
• MedXpertQA — total ?; splits: test (-1)

Metrics

• accuracy (primary) — range: percent
  • The percentage of correctly judged reasoning traces out of the total evaluated traces. Calculated as (correct predictions / total predictions) * 100.

Input / output format

Input: A medical question and a candidate multi-step reasoning trace generated by a frozen generator model.

Output: Binary judgment (correct/incorrect) or a continuous confidence score $r_\tau \in [0,1]$ derived from softmax probabilities of the 1 and 0 tokens.

Scoring recipe

def calculate_accuracy(predictions, gold):
    correct = sum(1 for p, g in zip(predictions, gold) if p == g)
    return (correct / len(gold)) * 100

Common pitfalls

• Averages accuracy across four benchmarks of vastly different sizes without weighting, which can skew the reported 'Avg.' score.
• Confuses the verifier's judgment accuracy with the generator's final answer accuracy; test-time search strategies (e.g., Best-of-N) further decouple verifier performance from final task performance.

Evidence (verbatim from paper)

We evaluated Med-TIV on four open-source medical question-answering benchmarks: MedQA*(Jin et al., [2020]), MedMCQA(Pal et al., [2022]), MMLU-Med(Hendrycks et al., [2021]), and MedXpertQA(Zuo et al., [2025])*, using accuracy as the evaluation metric.

Citation

@misc{zhang2026medtiv,
  title={Scaling Medical Reasoning Verification via Tool-Integrated Reinforcement Learning},
  author={Zhang et al. (2026)},
  year={2026},
  note={arXiv:2601.20221}
}

• arXiv: 2601.20221