Track manipulation, locomotion and embodied systems.
Search papers, labs, and topics across Lattice.
MedEvo addresses the underutilization of valuable reasoning within failed trajectories of medical multimodal reasoning models by employing a novel knowledge-guided navigation approach. This method leverages a medical tag taxonomy and error-type diagnosis to identify and repair consequential errors, retaining valid prefixes and regenerating erroneous suffixes. MedEvo significantly improves model accuracy across multiple benchmarks, outperforming standard supervised fine-tuning and teacher-distillation methods, and revealing that failed trajectories contain preserved reasoning and that independent feedback is crucial for sustained learning.
Failed medical reasoning trajectories hold significant untapped learning potential, with MedEvo unlocking it to boost model accuracy beyond teacher-distillation.
Medical multimodal reasoning models are typically trained on teacher-generated rationales, discarding the valid reasoning in students'own failed trajectories, while general-domain correction methods offer little guidance for errors in image reading, anatomical context, or differential diagnosis. We propose MedEvo, which uses medical knowledge, namely a tag taxonomy and error-type diagnosis, to navigate both trajectory repair and data selection. For each failed rollout, a gold-blind auditor locates the first consequential error and provides a minimal, answer-free hint; the same student retains the valid prefix and regenerates the erroneous suffix. After each round, MedEvo identifies capability gaps across medical tags on a held-out feedback set and prioritizes related questions for new rollouts and repairs. With supervised fine-tuning (SFT) alone, MedEvo raises Qwen3.5-9B mean accuracy across seven benchmarks from 65.57% to 69.12% (47.40% to 54.50% on the multimodal track of MedXpertQA (MedXpertQA-MM)), surpassing teacher-distillation SFT (66.93%). It also raises Qwen3.5-35B and Qwen3.5-397B mean accuracy from 67.91% to 73.04% and from 71.83% to 74.95%, respectively. Our analyses show that failed trajectories contain reasoning worth preserving, that repair success does not imply learning value, and that independent feedback is key to sustained improvement.