Andrii Bidochko

• Updated: August 24, 2026
• 6 min read

Teaching agentic AI to learn expert reasoning for rare disease diagnosis

Direct Answer

The paper introduces liteOdyssey, an agentic diagnostic system that converts expert reasoning for rare disease diagnosis into a reusable policy using Policy Iteration with Human Feedback (PIHF). By grafting this clinician‑gated policy onto off‑the‑shelf large language models (LLMs), the authors boost first‑rank diagnostic accuracy from roughly one‑third to nearly sixty percent while keeping the model under direct human oversight.

Background: Why This Problem Is Hard

Rare disease diagnosis is a classic “needle‑in‑a‑haystack” problem. Over 7,000 rare conditions affect an estimated 400 million people worldwide, yet most clinicians encounter only a handful in their careers. The diagnostic pathway therefore relies heavily on a small pool of domain experts who can synthesize sparse phenotypic data, genetic test results, and nuanced clinical narratives into a coherent differential diagnosis.

Current AI approaches struggle for three intertwined reasons:

• Data scarcity: Public datasets contain limited labeled cases for each rare disease, preventing conventional supervised learning from converging on reliable patterns.
• Reasoning opacity: Large language models excel at pattern matching but lack the step‑by‑step logical scaffolding that clinicians use to rule out mimics and prioritize hypotheses.
• Governance gaps: Deploying a black‑box model in a high‑stakes clinical setting raises liability and trust concerns; clinicians need a mechanism to inspect, correct, and ultimately control the AI’s suggestions.

These bottlenecks explain why off‑the‑shelf LLMs rank the correct disease first in only about 35 % of benchmark cases, far below the performance of specialized diagnostic engines that rely on curated knowledge bases and hand‑tuned heuristics.

What the Researchers Propose

The authors present a novel learning loop called Policy Iteration with Human Feedback (PIHF). PIHF adapts the classic reinforcement‑learning concept of generalized policy iteration to the realm of in‑context learning:

• Policy extraction: When the base LLM makes a diagnostic suggestion, clinicians review the output, flag errors, and provide corrected reasoning steps.
• Policy consolidation: The collected corrections are distilled into an explicit, rule‑like policy that maps a given patient context to a ranked list of disease hypotheses.
• Policy deployment: This policy is injected back into the LLM’s prompting chain, effectively turning the model into an agentic system that follows the expert‑derived decision logic.

Crucially, the policy remains clinician‑gated: any new suggestion must pass through the same human‑in‑the‑loop verification before being added to the policy repository. This creates a transparent, auditable knowledge layer that can be inspected, revised, or rolled back at any time.

How It Works in Practice

The operational workflow of liteOdyssey can be broken down into four sequential modules:

1. Patient Encoding: Structured phenotypic data (e.g., HPO terms), laboratory values, and narrative notes are transformed into a concise prompt that the LLM can ingest.
2. Base LLM Inference: An off‑the‑shelf model (e.g., GPT‑4, Claude, or an open‑weight alternative) generates an initial differential diagnosis and a natural‑language rationale.
3. Clinician Review & Feedback: A domain expert examines the output, annotates mis‑ranked diseases, and supplies a corrected reasoning chain. This feedback is recorded in a policy database.
4. Policy‑Guided Re‑ranking: The policy engine intercepts future prompts, applies the stored decision rules, and re‑orders the LLM’s raw suggestions to align with expert reasoning before presenting the final list to the clinician.

The following diagram (placeholder) illustrates the iterative PIHF loop, showing how each failure becomes a learning opportunity that enriches the policy:

What sets this approach apart from simple prompt engineering or retrieval‑augmented generation is the explicit separation between the language model’s generative capacity and the deterministic, human‑validated policy layer. The model continues to provide rich contextual knowledge, while the policy enforces a disciplined, expert‑driven ranking strategy.

Evaluation & Results

To validate liteOdyssey, the researchers conducted three complementary experiments:

Benchmark Diagnostic Accuracy

Using a public rare‑disease benchmark comprising 1,243 cases across 722 conditions, they measured the proportion of cases where the correct disease appeared at the top of the differential. Without any policy, the base LLM achieved 26.5 % top‑rank accuracy. After applying the PIHF‑derived policy, accuracy rose to 59.3 %—a more than two‑fold improvement that matches the best published specialist systems while requiring far less computational overhead.

Generalization to Unseen Diseases

Importantly, the policy was trained on a subset of 1,193 cases (679 diseases) and then evaluated on the remaining 50 cases (43 diseases) that were excluded from policy development. The uplift remained virtually identical (≈58 % vs. 26 % without policy), demonstrating that the learned reasoning generalizes beyond the specific diseases it was trained on.

Cross‑Model Transferability

The same policy was deployed unchanged on both closed‑weight (e.g., proprietary) and open‑weight (e.g., Ollama) LLMs. In each scenario, the policy produced comparable gains, confirming that the expert reasoning encapsulated in the policy is model‑agnostic and can be ported across different inference back‑ends.

Additional ablation studies ruled out alternative explanations such as mere prompt lengthening or direct access to external knowledge bases; the policy itself accounted for the majority of the performance boost.

Why This Matters for AI Systems and Agents

liteOdyssey showcases a pragmatic pathway for turning scarce expert knowledge into a scalable AI capability without retraining massive models. For AI practitioners building diagnostic agents, the key takeaways are:

• Modular governance: By decoupling reasoning (policy) from knowledge generation (LLM), teams can enforce compliance, audit decisions, and satisfy regulatory requirements.
• Rapid adaptation: New disease insights or updated clinical guidelines can be incorporated by simply adding fresh clinician feedback to the policy store, avoiding costly model fine‑tuning cycles.
• Resource efficiency: The approach delivers specialist‑level performance with a fraction of the compute and data footprint traditionally needed for domain‑specific LLMs.
• Interoperability: Because the policy layer is model‑agnostic, it can be integrated into existing AI pipelines, orchestration frameworks, or enterprise AI platforms such as the Enterprise AI platform by UBOS.

These advantages align with emerging best practices for building trustworthy, agentic AI that remains under human control—a prerequisite for any high‑risk application like medical diagnosis.

What Comes Next

While the results are compelling, several open challenges remain:

• Policy complexity management: As the number of diseases and feedback instances grows, maintaining an efficient, searchable policy database will require sophisticated indexing and versioning strategies.
• Multi‑modal integration: Extending the workflow to incorporate imaging, genomics, and wearable sensor data could further improve diagnostic breadth but will demand richer representation formats.
• Long‑term clinical impact studies: Prospective trials measuring patient outcomes, time‑to‑diagnosis, and cost savings are needed to move from benchmark performance to real‑world adoption.
• Regulatory alignment: Formalizing the policy‑as‑software concept within existing medical device frameworks will be essential for FDA or EMA clearance.

Future research may explore automated policy synthesis, where the system suggests candidate rules based on aggregated clinician feedback, thereby reducing the manual curation burden. Additionally, integrating liteOdyssey with broader health‑AI ecosystems—such as electronic health record (EHR) platforms, clinical decision support tools, and patient‑facing chatbots—could unlock end‑to‑end diagnostic pipelines.

Developers interested in prototyping similar agentic systems can start by reviewing the original paper and experimenting with the UBOS platform overview, which offers modular components for policy management, LLM orchestration, and secure data handling.

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.