Rewarding Reasoning, Not Answers: Fixing and Bounding Test-Time Reinforcement Learning on Medical QA

• URL: http://arxiv.org/abs/2609.16660v1
• Date: Tue, 15 Sep 2026 05:31:55 GMT
• Title: Rewarding Reasoning, Not Answers: Fixing and Bounding Test-Time Reinforcement Learning on Medical QA
• Abstract summary: Test-time reinforcement learning adapts a model on its own unlabeled test set using majority-vote pseudo-labels.<n>We show that this recipe collapses on medical multiple-choice QA: accuracy stagnates while output diversity rapidly declines.<n>This diagnosis motivates PROSE, Process Reward Guided Self-Training, which rewards reasoning quality instead of answer agreement.
• Score: 40.18486858532692
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Test-time reinforcement learning adapts a model on its own unlabeled test set using majority-vote pseudo-labels and has shown strong results in mathematics. We show that this recipe collapses on medical multiple-choice QA: accuracy stagnates while output diversity rapidly declines. Through a controlled experiment that keeps the questions, model, and optimizer fixed while changing only the answer space, we trace this failure to answer-space structure rather than domain difficulty. In small answer spaces, incorrect rollouts often collide on the same wrong pseudo-label and reinforce it; in large answer spaces, they disperse and receive little reward. This diagnosis motivates PROSE, Process Reward Guided Self-Training, which rewards reasoning quality instead of answer agreement. PROSE scores each reasoning step with a medical process reward model, assigns the trajectory reward as the minimum score across steps, and enforces answer-format constraints. Without labels, PROSE substantially improves a general Llama model, surpassing purpose-built medical models and matching much larger systems. Because the process signal is internalized into the policy, the adapted model requires no reward model at inference and transfers its gains to unseen datasets. We further show that the minimum aggregation is essential: mean aggregation can be exploited, saturating the proxy reward while degrading accuracy.

Related papers

• Measuring Reward Hacking and Reasoning-Answer Decoupling Under Position-Confounded Optimization [1.1316247605466565]
  We train language models with GRPO on math problems where the correct answer is always option A, then evaluate on an unseen test set with unbiased answer positions.<n>Across Qwen2.5, Llama 3.x and Gemma-3 models, biased training often drives option-A rates above 0.90 in smaller models.<n>We track reasoning-answer decoupling: capable models generate reasoning that reaches the correct numeric answer while still selecting A.
  arXiv  Detail & Related papers  (2026-08-15T23:13:57Z)
• Correct Answers from Sound Reasoning: Verifiable Process Supervision for Language Models [94.68358825189738]
  We propose a post-training framework for verifiable domains that jointly optimize prediction accuracy and reasoning quality.<n>We evaluate VPS on chess, a controlled testbed where reasoning steps can be deterministically verified against engine signals.<n>VPS preserves accuracy while significantly improving reasoning quality, reducing win-rate error by up to 30% and restoring consistency to near saturation.
  arXiv  Detail & Related papers  (2026-04-03T15:19:46Z)
• Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models [78.68818219506313]
  This paper describes a multi-answer reinforcement learning approach for training LMs to perform distributional reasoning over multiple answers during inference.<n>Across question-answering, medical diagnostic, and coding benchmarks, we observe improved diversity, coverage, and set-level calibration scores compared to single answer trained baselines.
  arXiv  Detail & Related papers  (2026-03-25T22:20:25Z)
• Answer-Consistent Chain-of-thought Reinforcement Learning For Multi-modal Large Langauge Models [33.398631680508814]
  We propose Answer-Consistent Reinforcement Learning that modifies the GRPO algorithm with an auxiliary consistency check.<n>We design a consistency-verification reward that grants a high reward only if both the original and the post-shuffle answers agree and are correct.<n>We evaluate ACRE on challenging Video Reasoning benchmarks and multimodal math reasoning benchmarks, achieving an average 2.2% and 1.5% improvement.
  arXiv  Detail & Related papers  (2025-10-11T08:32:52Z)
• Hybrid Reinforcement: When Reward Is Sparse, It's Better to Be Dense [36.71358559780692]
  HERO is a reinforcement learning framework that integrates verifier signals with reward-model scores in a structured way.<n> HERO consistently outperforms RM-only and verifier-only baselines, with strong gains on both verifiable and hard-to-verify tasks.
  arXiv  Detail & Related papers  (2025-10-08T17:09:41Z)
• The Majority is not always right: RL training for solution aggregation [53.1050856072799]
  We train an aggregator model to review, reconcile, and synthesize a final, correct answer.<n>A key ingredient is careful balancing of easy and hard training examples.<n>We find our method, AggLM, outperforms both strong rule-based and reward-model baselines.
  arXiv  Detail & Related papers  (2025-09-08T16:39:38Z)
• Mitigating Spurious Correlations Between Question and Answer via Chain-of-Thought Correctness Perception Distillation [25.195244084313114]
  Chain-of-Thought Correctness Perception Distillation (CoPeD) aims to improve the reasoning quality of the student model.<n>CoPeD encourages the student model to predict answers based on correct rationales and revise them when they are incorrect.
  arXiv  Detail & Related papers  (2025-09-06T05:33:17Z)
• The Perils of Optimizing Learned Reward Functions: Low Training Error Does Not Guarantee Low Regret [64.04721528586747]
  We show that a sufficiently low expected test error of the reward model guarantees low worst-case regret.<n>We then show that similar problems persist even when using policy regularization techniques.
  arXiv  Detail & Related papers  (2024-06-22T06:43:51Z)
• Test-Time Adaptation with CLIP Reward for Zero-Shot Generalization in Vision-Language Models [76.410400238974]
  We propose TTA with feedback to rectify the model output and prevent the model from becoming blindly confident. A CLIP model is adopted as the reward model during TTA and provides feedback for the VLM. The proposed textitreinforcement learning with CLIP feedback(RLCF) framework is highly flexible and universal.
  arXiv  Detail & Related papers  (2023-05-29T11:03:59Z)
• Robust Question Answering Through Sub-part Alignment [53.94003466761305]
  We model question answering as an alignment problem. We train our model on SQuAD v1.1 and test it on several adversarial and out-of-domain datasets.
  arXiv  Detail & Related papers  (2020-04-30T09:10:57Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.