A new training method lets AI models get sharper at spotting rare diseases without touching their underlying weights.
The technique, called Policy Iteration with Human Feedback (PIHF), keeps a pretrained language model frozen and instead edits a versioned, natural-language policy and toolset sitting on top of it. A language-model critic and a clinical expert review the model's reasoning and tool-use traces, flag recurring failure patterns, and draft candidate fixes. The clinical expert keeps final say over whether a revision gets adopted or rolled back, and results are checked with Recall@1 and Recall@5 after each change. Researchers tested the approach on ultra-rare-disease diagnosis benchmarks across four models spanning 3 to 49 billion parameters, including proprietary GPT-5.4 and open-weight Qwen3.6-35B.
Recall@1 improved for all four models, with GPT-5.4 and Qwen3.6-35B posting the largest jumps: 32.7 and 31.1 percentage points, respectively. That matters because rare-disease diagnosis is exactly the kind of data-starved problem where retraining a model on more examples usually isn't an option. Editing a shared, human-reviewed policy instead of fine-tuning weights is cheaper, and it leaves a paper trail an expert can audit or reverse.
The catch: the paper only publishes hard numbers for two of the four models. The other two also improved, but by how much is left unstated, which makes it hard to judge whether these gains hold up outside the best-case examples the authors chose to highlight.