GUIDER
A reinforcement learning method for clinical guideline reasoning. It jointly trains a decision-flow controller and a reasoning policy, and employs multi-flow sampling together with a process reward based on answer gating to address both the path-answer gap and the problem of path diversity.
Timeline (last 90 days)
The GUIDER method jointly trains two components: a decision flow controller and a reasoning policy.
The decision flow controller learns from guideline-derived supervision to generate multiple reasonable decision flows.
GUIDER evaluates the policy's process reward on the multiple flows generated by the controller and gates that reward by answer correctness.
GUIDER's joint training belongs to the on-policy reinforcement learning paradigm, with process rewards recomputed as the policy updates.
On the MedGUIDE benchmark, GUIDER with a Qwen3-14B backbone achieves 82.5 weighted accuracy and 73.5 guideline compliance.
On MedGUIDE, GUIDER exceeds the strongest clinical reinforcement learning baseline by 3.3 and 3.1 points in weighted accuracy and guideline compliance, respectively.
On the Qwen3-14B and Gemma4-12B backbones, GUIDER outperforms CRPO, MRPO and FaithMed on the out-of-domain benchmarks AMEGA-LLM, RealMedQA and CancerGUIDE.
GUIDER's "multi-flow sampling + answer gating" idea can be borrowed for domains requiring process supervision, such as multi-step tool calling and complex planning.
All Facts (8)
The GUIDER method jointly trains two components: a decision flow controller and a reasoning policy.
50%UnverifiedThe decision flow controller learns from guideline-derived supervision to generate multiple reasonable decision flows.
50%UnverifiedGUIDER evaluates the policy's process reward on the multiple flows generated by the controller and gates that reward by answer correctness.
50%UnverifiedGUIDER's joint training belongs to the on-policy reinforcement learning paradigm, with process rewards recomputed as the policy updates.
50%UnverifiedOn the MedGUIDE benchmark, GUIDER with a Qwen3-14B backbone achieves 82.5 weighted accuracy and 73.5 guideline compliance.
50%UnverifiedOn MedGUIDE, GUIDER exceeds the strongest clinical reinforcement learning baseline by 3.3 and 3.1 points in weighted accuracy and guideline compliance, respectively.
50%UnverifiedOn the Qwen3-14B and Gemma4-12B backbones, GUIDER outperforms CRPO, MRPO and FaithMed on the out-of-domain benchmarks AMEGA-LLM, RealMedQA and CancerGUIDE.
50%UnverifiedGUIDER's "multi-flow sampling + answer gating" idea can be borrowed for domains requiring process supervision, such as multi-step tool calling and complex planning.
50%