aiqbench analysis
This research investigates the efficacy of hindsight-guided distillation for fine-tuning smaller models on rare disease diagnosis, revealing that accuracy gains stem primarily from filtering out "ground truth" hallucinations rather than the distillation process itself. The study highlights a critical calibration gap and frequency-dependent knowledge transfer limitations in supervised fine-tuning compared to reinforcement learning approaches. These findings offer essential insights into optimizing reasoning chains for high-stakes medical AI applications where data scarcity is common.
Hindsight-Guided Rationale Distillation for Rare Disease Diagnosis
arXiv:2610.03176v1 Announce Type: new Abstract: We study hindsight-guided distillation for rare disease diagnosis on ZebraMap: a 1.5B student is fine-tuned on chain-of-thought traces from a 8B teacher that observes the ground-truth diagnosis during generation. Absolute accuracy remains low for all models - the task is hard at this scale - but within this ceiling a filtered variant (StudentF) achieves a small, statistically significant accuracy advantage over the teacher (p < 0.001), concentrated in better-represented diseases. The unfiltered student does not significantly outperform the teacher (p = 0.129), establishing that contamination filtering - not hindsight distillation alone - drives the gain. The gap traces to an artifact we term GT hallucination. Label-visible generation causes the teacher to embed "ground truth is X" phrases in its reasoning chain; SFT copies the pattern. At inference, the unfiltered student reproduces the phrase in 33.9% of cases, with severe accuracy degradation when the hallucinated label is wrong. A regex filter removing these slots reduces contamination to near-zero, producing the observed gain - though the effect remains small. We precisely quantify this gain-cost tradeoff, document frequency-dependent knowledge transfer absent from the RL-trained teacher, and characterize a calibration gap that SFT does not close - identifying both as directions for future work.
Abstract:We study hindsight-guided distillation for rare disease diagnosis on ZebraMap: a 1.5B student is fine-tuned on chain-of-thought traces from a 8B teacher that observes the ground-truth diagnosis during generation. Absolute accuracy remains low for all models - the task is hard at this scale - but within this ceiling a filtered variant (StudentF) achieves a small, statistically significant accuracy advantage over the teacher (p < 0.001), concentrated in better-represented diseases. The unfiltered student does not significantly outperform the teacher (p = 0.129), establishing that contamination filtering - not hindsight distillation alone - drives the gain. The gap traces to an artifact we term GT hallucination. Label-visible generation causes the teacher to embed "ground truth is X" phrases in its reasoning chain; SFT copies the pattern. At inference, the unfiltered student reproduces the phrase in 33.9% of cases, with severe accuracy degradation when the hallucinated label is wrong. A regex filter removing these slots reduces contamination to near-zero, producing the observed gain - though the effect remains small. We precisely quantify this gain-cost tradeoff, document frequency-dependent knowledge transfer absent from the RL-trained teacher, and characterize a calibration gap that SFT does not close - identifying both as directions for future work.
Computation and Language (cs.CL)
arXiv:2610.03176 [cs.CL]
(or arXiv:2610.03176v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2610.03176 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 295 words
Related coverage
LiteOdyssey: A Lightweight Reasoning AI Agent for Interpretable Rare-Disease Diagnosis
arXiv · Jun 16, 2026
AI by Default
LessWrong · Oct 17, 2026
MedLatentDx: Latent Multi-Agent Communication for Cross-Hospital Rare-Disease Diagnosis
arXiv · Jun 15, 2026
Learning from Repaired Reasoning: Root-Cause-Guided On-Policy Distillation
arXiv · Oct 5, 2026