HypothesisMed: Inference-Time Answer Fusion and Structured Hypothesis-Space Reporting for Biomedical Question Answering

Technical Summary of HypothesisMed

Problem Statement
Current evaluations of Large Language Models (LLMs) for biomedical question answering (QA) rely heavily on final answer accuracy. However, the paper argues that accuracy alone is insufficient for reliability-sensitive biomedical workflows. A model may select the correct option yet fail to produce machine-parseable outputs, ignore structured instructions, or confidently commit to incorrect answers when the available answer space is incomplete, duplicated, or contradictory. Existing benchmarks like MedQA, MedMCQA, and PubMedQA typically do not assess whether a model can recognize these "weak" answer spaces or provide auditable reliability metadata alongside its predictions.

Methodology: The HypothesisMed Pipeline
The paper introduces HypothesisMed, an inference-time reliability pipeline designed to separate answer selection from structured reliability reporting. The system operates through the following components:

• Prompting Strategies: The pipeline evaluates four inference-time methods:
  1. Direct Prompting: Minimal reasoning, direct answer selection.
  2. Chain-of-Thought (CoT): Explicit reasoning steps before selection.
  3. HypothesisMed-v3 Prompting: A structured prompt designed to output specific metadata.
  4. Proposed Fusion: A combination of the above.
• Structured SPACE Labels: A core innovation is the SPACE field (a label name, not a model), which HypothesisMed-v3 uses to classify the answer space status into three categories:
  • VALID: One medically supported best answer exists.
  • INCOMPLETE: The correct answer appears missing or insufficiently represented.
  • CONTRADICTED: Options are internally inconsistent, duplicated, or medically contradictory.
• Answer Fusion: The final answer is selected via majority voting over parseable candidates generated by the direct, CoT, and HypothesisMed-v3 prompts. If no strict majority exists, a deterministic fallback order (Direct $\to$ CoT $\to$ HypothesisMed-v3) is used.
• Evaluation Metrics: Beyond standard Answer Accuracy, the study measures:
  • Parse Coverage: The fraction of outputs successfully parsed into a valid option.
  • SPACE Coverage: The fraction of outputs containing a valid SPACE label.
  • SPACE Accuracy: The correctness of the SPACE label (e.g., correctly identifying a VALID space vs. falsely claiming it is INCOMPLETE).
  • False Commitment: The rate of high-confidence (confidence $\ge$ 0.5) incorrect answers.

Experimental Setup
The study evaluates four open-weight models: Qwen2.5-7B-Instruct, Phi-4-mini-instruct, DeepSeek-R1-Distill-Qwen-32B, and BioMistral-7B.

• Primary Evaluation: 1,000 examples per dataset (MedQA, MedMCQA, PubMedQA) for all four models.
• Scaled Evaluation: 10,183 examples per model for Qwen2.5-7B and Phi-4-mini.
• SPACE Stress Test: An expanded 12,000-example test per model where answer spaces are artificially perturbed (removing correct answers to create INCOMPLETE sets, duplicating answers to create CONTRADICTED sets) to test the model's ability to detect defects.

Key Results

1. Accuracy and Reliability Trade-offs: In the primary four-model evaluation, the proposed Answer Fusion + HypothesisMed-v3 pipeline improved weighted answer accuracy over the best direct or CoT baselines for all models while simultaneously increasing parse coverage and SPACE coverage.
   • Qwen2.5-7B: Achieved 0.6367 accuracy (vs. 0.6003 CoT baseline) with 100% parse coverage and 99.97% SPACE coverage.
   • Phi-4-mini: Showed substantial gains, improving from 0.4817 (CoT) to 0.5767 (Proposed), with a massive reduction in false commitments (from 0.9838 to 0.4054).
2. Model-Specific Behaviors:
   • DeepSeek-R1-32B: Showed improved accuracy with fusion (0.1690 to 0.3877), but the authors note this primarily reflects improved formatting/usability rather than a pure reasoning gain, as its baselines had very low parse coverage.
   • BioMistral-7B: Demonstrated poor structured-output compliance (near-zero SPACE coverage), indicating that domain specialization does not guarantee adherence to structured output constraints.
3. Scaled Evaluation: In the 10,183-example test, fusion substantially improved Phi-4-mini's accuracy (0.4296 to 0.5192). For Qwen2.5-7B, CoT remained slightly higher in raw accuracy than fusion, but the fusion method provided complete parse/SPACE coverage and significantly lower false commitment.
4. SPACE Stress Test: The ability to diagnose answer-space defects remained difficult. In the stress test, Qwen2.5-7B achieved 0.3074 SPACE accuracy, and Phi-4-mini achieved 0.4168, suggesting that prompt-only structured inference struggles to reliably detect subtle answer-space anomalies.

Significance and Claims
The paper explicitly states that its main contribution is not a universal state-of-the-art claim for biomedical QA accuracy. Instead, the significance lies in:

• Separability of Capabilities: Demonstrating that answer accuracy, parseability, structured reliability reporting, calibration, and false-commitment behavior are distinct, separable capabilities that must be evaluated independently.
• Auditable Workflow Components: Providing a reproducible inference-time framework for evaluating biomedical LLMs as auditable workflow components under structured reliability constraints, rather than merely as answer generators.
• Structured Reliability Reporting: Introducing a method to generate machine-readable metadata (SPACE labels) that allows downstream systems to detect when a model is operating on an incomplete or contradictory answer space, thereby mitigating the risk of confident incorrect commitments.

The authors conclude that while answer-space diagnosis remains a challenge for current models, the HypothesisMed pipeline offers a necessary step toward evaluating biomedical LLMs with the rigor required for safe, automated deployment.