Shanghai Jiao Tong's MedSSR Doubles Rare Disease AI Accuracy Without Costly Teachers
MedSSR uses knowledge-guided synthetic questions and self-generated pseudo-labels to lift rare-disease reasoning by up to 5.93 percent without expensive teacher distillation.
ProTakeaways
- MedSSR is an ACL 2026 Findings paper introducing knowledge-enhanced synthesis plus semi-supervised RL for medical LLMs.
- Rare-disease knowledge is injected on demand, giving controllable distribution over synthetic reasoning questions.
- The policy model pseudo-labels its own synthetic data via offline majority voting, avoiding teacher distillation.
- Two-stage RLVR training: self-supervised RL on synthetic data, then supervised RL on human-annotated real data.
- Up to +5.93% gain on rare-disease benchmarks with Qwen and Llama backbones, breaking the prior 3% ceiling.
- Code, environment files, and evaluation datasets available at github.com/tdlhl/MedSSR.
Getting a language model to reason well about medicine usually means paying a large proprietary model to produce chain-of-thought traces, fine-tuning on those traces, and then running reinforcement learning on top. The approach works for common conditions, but it is expensive and stubbornly bad at the long tail. A new ACL 2026 Findings paper from researchers at Shanghai Jiao Tong University argues that both problems share a root cause, and offers a way around them.
The framework is called MedSSR (Medical Knowledge-enhanced data Synthesis and Semi-supervised Reinforcement learning). Rather than buying reasoning traces from a bigger model, it manufactures its own training questions from structured medical knowledge and lets the policy model label them itself before any human-annotated data is used.
Why rare diseases break the standard recipe
Large language models hold promise for complex clinical applications, but their development is hindered by the scarcity of high-quality reasoning data. Existing approaches typically distill chain-of-thought traces from large proprietary models via supervised fine-tuning, then conduct reinforcement learning. These pipelines show limited improvement on underrepresented domains like rare diseases while incurring substantial costs from generating complex reasoning chains.
The authors quantify how stuck the field is: even when employing the strong baseline of fully supervised GRPO, gains on rare disease data fail to surpass a 3% ceiling. Synthetic data is the obvious escape hatch, but generated examples often suffer from low fidelity and factual errors, which are especially unacceptable in clinical scenarios. Any synthesis pipeline in this domain has to incorporate structured medical knowledge to prevent factual errors and ensure clinical rationality.