Technical Summary: BODHI Framework for Curiosity-Driven and Humble AI in Clinical Decision Support

1. Problem Statement

Large Language Models (LLMs) exhibit significant limitations when applied to clinical decision support, primarily characterized by overconfidence and inflexible reasoning. Despite high statistical pattern recognition capabilities, LLMs often conflate confidence with medical understanding, producing declarative outputs even when evidence is insufficient. This behavior contributes to automation bias, where clinicians may defer to incorrect AI suggestions, and sycophantic behavior, where models comply with illogical medical requests.

Existing approaches to mitigate these issues fall short:

• Uncertainty Quantification (UQ): Methods like conformal prediction estimate uncertainty but fail to modify the model's behavior or communication style when uncertainty is high.
• Chain-of-Thought (CoT): While improving reasoning accuracy, standard CoT optimizes for correctness rather than appropriate epistemic conduct (e.g., admitting uncertainty).
• Fine-tuning: Approaches like ConfiDx require model modification, limiting practical deployment and generalizability.
• Conceptual Frameworks: Previous work has called for "epistemic humility" as a design principle but has not provided concrete engineering implementations to operationalize these virtues.

The core problem is the gap between philosophical proposals for humble AI and the engineering reality of deploying LLMs that can reliably constrain their behavior to align with epistemic virtues without architectural changes.

2. Methodology: The BODHI Framework

The authors propose BODHI (Balanced, Open-minded, Diagnostic, Humble, and Inquisitive), an engineering framework that operationalizes curiosity and humility through a two-pass chain-of-thought protocol. The framework does not require model fine-tuning; it functions via structured prompting.

2.1 Core Architecture

The framework operates through six integrated steps:

1. Clinical Complexity Assessment: Evaluates the query for diagnostic ambiguity, urgency, and data completeness.
2. Prior Confidence Evaluation: Estimates the model's epistemic state based on training distribution coverage.
3. Parallel Processing:
   • Curiosity Module: Identifies information gaps and generates clarifying questions.
   • Humility Module: Assesses confidence boundaries and triggers for deferral.
4. Virtue Activation Matrix (VAM): Maps clinical complexity against model confidence to select one of four epistemic stances:
   • Q1 (Proceed & Monitor): High confidence, low complexity.
   • Q2 (Watchful & Alternatives): High confidence, high complexity.
   • Q3 (Clarify & Review): Low confidence, low complexity (triggers high curiosity/humility).
   • Q4 (Escalate & Reframe): Low confidence, high complexity (triggers deferential humility and escalation).
5. Adaptive Response Generation: Generates responses based on the selected stance.
6. Continuous Learning: Incorporates clinician feedback to refine thresholds.

2.2 The Two-Pass Protocol

• Pass 1 (Structured Analysis): The model generates an internal, auditable analysis containing seven mandatory fields: Task Type, Audience, Primary Hypothesis, Key Uncertainties, Clarifying Questions (1–2 required for non-emergencies), Red Flags, and Safe Recommendations. This makes the epistemic state explicit.
• Pass 2 (Constrained Output): The model generates the clinician-facing response conditioned on Pass 1. It applies Cross-Cutting Epistemic Constraints:
  • Specificity: Use concrete numbers/timeframes where available.
  • Active Inquiry: Convert conditional statements into direct questions.
  • Alternatives: Present differential diagnoses when confidence is below threshold.
  • Task Routing: Adjusts behavior based on context (Emergency, Technical, Hybrid, or Conversation modes).

2.3 Evaluation Design

• Dataset: 200 challenging clinical vignettes from HealthBench Hard, selected for cases requiring nuanced judgment where baseline models underperform.
• Models: GPT-4.1-mini and GPT-4o-mini.
• Experimental Design: Evaluated across 5 random seeds (42–46) with 200 cases each, yielding 1,000 observations per model.
• Conditions: Baseline (standard prompting) vs. BODHI (two-pass protocol).
• Metrics: Overall HealthBench score, Context-Seeking Rate (proportion of responses with clarifying questions), Hedging Frequency, Helpful & Safe score, Emergency Referral Rate, and Communication Quality.
• Statistics: Nonparametric bootstrap resampling (10,000 iterations), paired t-tests, and Cohen's d effect size computation.

3. Key Results

The study demonstrates that BODHI significantly alters LLM behavior toward epistemic virtues without model modification.

3.1 Primary Outcomes

• Overall Clinical Quality:
  • GPT-4.1-mini: Improved from 2.5% to 19.1% (+16.6 percentage points, $p < 0.0001$).
  • GPT-4o-mini: Improved from 0.0% to 2.2% (+2.2 percentage points, $p < 0.0001$).
• Curiosity (Context-Seeking):
  • GPT-4.1-mini: Increased from 7.8% to 97.3% (+89.6pp). Effect size (Cohen's $d$) = 16.38 (Very Large).
  • GPT-4o-mini: Increased from 0.0% to 73.5% (+73.5pp). Effect size (Cohen's $d$) = 19.54 (Very Large).
• Humility (Hedging):
  • GPT-4.1-mini: Increased from 1.7% to 21.9% (+20.3pp, $d = 5.80$).
  • GPT-4o-mini: Increased from 0.8% to 5.2% ($d = 1.16$).
• Safety: Emergency referral rates increased significantly for both models (e.g., GPT-4.1-mini from 12.3% to 28.6%).

3.2 Trade-offs

• Communication Quality: Scores decreased for both models (GPT-4.1-mini: -12.5pp; GPT-4o-mini: -11.3pp). The authors attribute this to the shift from confident, declarative statements to nuanced, question-containing outputs. They argue this is an appropriate trade-off for high-stakes clinical domains where overconfidence poses greater risks than verbose, hedged responses.
• Model Capacity: GPT-4.1-mini showed larger gains in overall score and hedging behavior compared to GPT-4o-mini, suggesting model capacity influences the effectiveness of epistemic constraints.

3.3 Reproducibility

Performance was consistent across all 5 random seeds. For GPT-4.1-mini, context-seeking rates under BODHI ranged from 92.9% to 100.0%, with three seeds achieving 100%. This indicates the framework's effects are robust and not artifacts of specific sampling.

4. Key Contributions

1. Operationalization of Epistemic Virtues: The paper provides the first prompting-based framework that empirically demonstrates the ability to constrain LLMs to exhibit curiosity (information gathering) and humility (hedging/deferral) in clinical contexts.
2. Engineering Implementation: Unlike previous work that discusses epistemic humility conceptually or requires fine-tuning, BODHI offers a two-pass chain-of-thought protocol that is immediately deployable via prompting.
3. Virtue Activation Matrix: A novel decision logic that maps clinical complexity and model confidence to specific behavioral stances (Proceed, Watchful, Clarify, Escalate).
4. Empirical Validation: Demonstrates "very large" effect sizes (Cohen's $d > 1.2$) on curiosity and humility metrics, proving that behavioral modification is possible at the prompting level.

5. Significance and Claims

The authors claim that BODHI offers a pathway toward safer clinical AI deployment by ensuring systems operate as collaborative partners that know when to ask questions and when to defer, rather than as "overconfident oracles."

• Safety over Confidence: The framework prioritizes epistemic safety (acknowledging uncertainty) over the appearance of confidence, addressing the risk of automation bias where clinicians follow incorrect AI advice due to high confidence scores.
• Scalability: Because it relies on prompting rather than model retraining, the framework can be applied across different model variants and clinical domains without the computational cost of fine-tuning.
• Limitations Acknowledged: The authors modestly note limitations, including reliance on a single benchmark (HealthBench Hard), evaluation of only two models from one provider, and the absence of clinician-in-the-loop validation. They acknowledge that the two-pass architecture increases latency and that the framework's effectiveness may vary across different institutional contexts.

The paper concludes that while a quality-safety tradeoff exists (reduced communication scores), it is a necessary and appropriate adjustment for high-stakes medical decision support, aligning AI behavior with the epistemic virtues required for safe human-AI collaboration.