PaperlayerView exact revision on arXiv

Source-linked AI summary

MedEvo: Medical Reasoning Evolution through Knowledge Navigation

Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Quanbin Wang, Jiangwei Lao, Yingao Liu, Jiajia Liu, Jinming Cao, Honglin Lin, Zhihong Lu, Hao Jin, Jian Wang, Lijun Wu

arXiv:2610.17690v1cs.AIcs.CVcs.LGq-bio.QM

TL;DR

Medical multimodal reasoning methods often discard valid reasoning in failed student trajectories, while general-domain correction methods offer limited guidance for medical errors. MedEvo uses medically informed diagnosis to repair trajectories and direct later data selection; with SFT alone, it improves mean accuracy across seven benchmarks.

  • Problem

    Medical multimodal reasoning can fail through image misreading, overlooked clinical clues, or unsupported differential diagnosis, while existing failure-correction methods offer little guidance for these medical errors.

    Source: §1 · p. 1

  • Method

    MedEvo uses medical tags and error types to locate a consequential failure, guide answer-free same-student suffix repair, and prioritize later questions using held-out feedback.

    Source: §1 · p. 2, §3 · p. 4, §3.2.2 · p. 6

  • Results

    Across seven benchmarks, MedEvo raises Qwen3.5-9B mean accuracy from 65.57% to 69.12% with SFT alone.

    Source: §1 · p. 2, §4.2 · p. 10

  • Takeaways & Limitations

    The findings distinguish repair success from learning value and identify independent feedback as key to sustained improvement.

    Source: §1 · p. 2, §7 · p. 13, §7.1 · p. 13

    Longer runs on independent data are needed to assess regression, cost, and saturation of knowledge-structure evolution.

    Source: §7.2 · p. 14

Abstract

from arXiv · show

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.

1 Introduction

MedEvo addresses how to repair medical reasoning failures without discarding valid student reasoning, and how to use failures to guide subsequent training. It combines medically informed local repair with feedback-driven question selection, improving results across seven benchmarks.

1 Introduction

  • Medical reasoning failures can arise from image misreading, overlooked clinical clues, or unsupported differential diagnosis, even when earlier reasoning remains valid.The paper notes that existing failure-correction methods developed for mathematics and software engineering offer little guidance for errors involving medical images, anatomy, or differential diagnosis.Source: §1 · p. 1
  • Across seven benchmarks, MedEvo raises Qwen3.5-9B, Qwen3.5-35B, and Qwen3.5-397B mean accuracy from 65.57%, 67.91%, and 71.83% to 69.12%, 73.04%, and 74.95%, respectively.For Qwen3.5-9B, it also surpasses teacher-distillation SFT, 69.12% vs. 66.93%.Source: Figure 1 · p. 2, §1 · p. 2
  • For each failed rollout, a gold-blind auditor identifies the first consequential error and gives a minimal, answer-free hint so the same student can regenerate the erroneous suffix while retaining its valid prefix.After each round, held-out feedback across medical tags guides question selection for fresh rollouts and repairs.Source: §1 · p. 2

2 Related Work

Related work spans medical multimodal reasoning and supervision quality, localized learning from failures, and self-training or data evolution. MedEvo connects these areas by using medical diagnosis to guide repair and subsequent data selection.

2.1 Medical Multimodal Reasoning and Supervision Quality

  • Medical multimodal systems use biomedical instruction tuning, generalist or bilingual training, and knowledge graphs, while process audits and factuality checks assess reasoning beyond answer accuracy.Prior work also studies process supervision and student compatibility; MedEvo diagnoses failures with a medical codebook and screens repairs for correctness, valid boundaries, integrity, and student-relative learnability.Source: §2.1 · pp. 2–3

2.2 Learning from Failures and Localized Repair

  • Failure-learning methods include mistake–correction pairs, trajectory fusion, and hindsight hints, whereas MedEvo uses the same student to regenerate an answer-blind, medically conditioned suffix while retaining its supported prefix.This distinguishes MedEvo from approaches where a teacher writes the corrected suffix or the auditor accesses the reference answer.Source: §2.2 · p. 3

2.3 Data Evolution and Recursive Self-Improvement

  • Self-training methods filter self-generated reasoning for correctness, but MedEvo instead repairs failures with answer-free hints and uses medical-tag and error-type feedback to select each round’s data.Other medical evolution work may update external knowledge banks or procedures without updating model parameters.Source: §2.3 · p. 3

3 Method

MedEvo uses medical tags and diagnosed error types to guide both recovery within failed trajectories and supervision choices across training rounds. Its process preserves supported reasoning, screens repairs, and uses feedback-set capability gaps to direct later question selection.

3 Method

  • Knowledge Navigation links structured medical knowledge and failure diagnosis to intervention within a trajectory and decisions about what to learn next.Tags describe clinical context, while error types identify the failed reasoning operation; feedback reassesses capability gaps across rounds.Source: §3 · p. 4

3.1.1 Medical Question Tags

  • Questions receive hierarchical tags for modality, anatomy, disease, and difficulty, while a codebook maps four parent error types and twenty subtypes to recovery operations.Only input-supported observable tags are shown to the auditor; answer-derived labels are excluded.Source: §3.1.1 · pp. 4–5, §3.1.2 · p. 5

3.2.1 Student Rollouts and Uncertainty-Based Localization

  • For each selected question, the student samples eight trajectories; token entropy proposes an audit region, and the goal is to preserve steps before the first consequential error.The audit window covers steps k* ± 1, with λ = 0.5 and m = 3; entropy proposes the region, while the auditor locates errors using images, questions, and reasoning.Source: §3.2.1 · pp. 5–6

3.2.2 Gold-Blind Auditing and Medical Error Diagnosis

  • A gold-blind auditor diagnoses the earliest consequential error and supplies a minimal, answer-free hint, or skips the rollout without generating a continuation.The auditor sees the question, images, failed trajectory, entropy proposal, and observable tags, but no reference answer, rationale, or answer-derived metadata.Source: §3.2.2 · p. 6

3.2.3 Prefix Preservation and Same-Student Continuation

  • For repair, the same student keeps the valid prefix verbatim, appends the recovery cue, and regenerates the suffix, sampling up to five continuations.The audit record remains metadata, and evaluation excludes hints and audit outputs.Source: §3.2.3 · p. 6, §3.2.4 · p. 7

3.2.4 Repair Admission and Training Targets

  • Repairs enter training only after correctness, boundary, integrity, and suffix-only learnability checks; admitted examples train on the complete response, including the recovery cue.If no repair is admitted, a qualified correct rollout or valid historical target may be retained; otherwise, nothing is added.Source: §3.2.4 · pp. 6–7

3.3.1 First-Round Coverage

  • The first round generates rollouts for every training-pool question, screens correct trajectories, repairs failures, and fine-tunes from the original student on accepted targets.Before later rounds, feedback-set performance identifies tag-group gaps and error-type distributions; feedback questions guide selection and calibration but never enter training.Source: §3.3.1 · p. 7, §3.3.2 · p. 7, §3.3 · p. 7

3.3.3 Selecting Questions and Updating the Dataset

  • From the second round, feedback priorities allocate generation and repair effort to tag-matched questions, while screened targets replace existing ones only when blinded comparison judges them better.The dataset keeps at most one active target per question, retains valid history when a candidate is not better, and each round retrains from M0 with a fresh optimizer and schedule.Source: Figure 4 · p. 5, §3.3.3 · pp. 7–8

4 Experiments

MedEvo is evaluated on seven medical visual question-answering benchmarks, with a fixed feedback set kept separate from training. It improves mean accuracy over baselines and across rounds while reducing the repair workload.

4.1 Experimental Setup

  • The training pool contains 81,293 questions, while a separate fixed set of 5,000 modality-stratified questions provides feedback; evaluation uses an unweighted mean across seven benchmarks.The benchmarks include MedXpertQA-MM and MMMU-Med.Source: §4.1 · p. 9

4.2 Main Results

  • Qwen3.5-9B mean accuracy rises from 65.57% to 69.12%, exceeding teacher-distillation SFT (66.93%) and improving all seven benchmarks.The largest gain is on MedXpertQA-MM, from 47.40% to 54.50%; paired-bootstrap confidence intervals exclude zero on every benchmark.Source: Table 2 · p. 9, §4.2 · p. 10, Table 3 · p. 10

4.3 Accuracy Gains and Repair Workload

  • Mean accuracy rises from 65.57% to 67.22%, 67.42%, and 69.12% across rounds as questions sent for repair fall from 27,233 to 23,620, with repair success near 71%.Individual benchmark results can fluctuate, as illustrated by PathVQA.Source: Table 1 · p. 9, §4.3 · p. 10

5 Ablation Studies and Analysis

Across model scales, MedEvo improves seven-benchmark mean accuracy, while analyses show that repair design, feedback, and question selection affect downstream learning value. Repair success alone does not predict which supervision is most useful.

5.1 Repair Design in the First Round

  • Type-guided suffix continuation achieves 71.80% repair success and 67.22% mean accuracy, outperforming the reported generic-hint continuation and hint-guided full-rewrite designs.The analysis reports that both preserving the prefix and conditioning hints on error type contribute.Source: Table 4 · p. 10, §5.1 · p. 10

5.2 Contribution of Feedback across Rounds

  • Full MedEvo reaches 69.12% R3 accuracy, versus 66.67%, 67.53%, and 66.21% without feedback, without medical tags, and without error-type feedback, respectively.The branches share the R1 data-generation model and historical dataset; without feedback or error-type feedback, accuracy falls below the 67.22% R1 reference.Source: Table 5 · p. 11, §5.2 · p. 11

5.3 Generalization across Model Scales

  • Qwen3.5-35B mean accuracy rises from 67.91% to 73.04%, and Qwen3.5-397B from 71.83% to 74.95%.At 35B, MedXpertQA-MM peaks in R2; at 397B, R2 improves six of seven benchmarks, with OmniMedVQA essentially unchanged.Source: Table 6 · p. 11, §5.3 · p. 11

5.4 Learning Value of Repair Supervision

  • The longest-trajectory set has lower repair success than random selection (55.70% vs. 67.50%) but higher downstream accuracy (66.78% vs. 64.35%).The comparison uses 8,000 questions in each set for one round.Source: §5.4 · p. 11, Figure 5 · p. 12
  • Three hints yield 67.01% mean accuracy, compared with 66.78% for one hint; accuracy falls to 66.74% and 64.10% with five and eight hints.Each condition uses the longest-8k subset and five continuation attempts.Source: §5.4 · p. 11, Figure 5 · p. 12
  • Training-error feedback raises training accuracy from 63.23% to 78.00% but lowers benchmark accuracy from 65.57% to 64.01%; fixed-set feedback yields 69.12% benchmark accuracy.The fixed-set condition has 71.00% training accuracy; the feedback set contains 5,000 held-out questions.Source: §5.4 · p. 11, Figure 5 · p. 12
  • Entropy localization raises repair acceptance from 19.8–35.8% to 25.2–40.9% across one to eight attempts, while erroneous-step rates rise from 16.2% to 64.1% across entropy bins.On the modality-stratified feedback subset, MedEvo improves accuracy over Base at L1 and L5, but the Heuristic baseline scores higher at L5.Source: Figure 6 · p. 12, §5.4 · p. 12

6 Reliability Analysis

The reliability analysis audits repaired answers and visual grounding, then checks whether accuracy gains persist across independent training runs. Reviewers validate repair quality, while all three runs improve through R3.

6.1 Human Audit

  • 88.3% of audited repairs were judged correct and 78.5% image-grounded; reviewers validated 84.5% of repair boundaries and found 97.2% of hints answer-blind.The audit sample also had 95.5% of submitted trajectories labeled erroneous. Cohen’s κ was 0.89 for answer correctness and 0.70 for grounding; κ values for boundary validation and answer-blindness were 0.85 and 0.94.Source: §6.1 · p. 12, Table 7 · p. 13
  • The audit assesses answer correctness and visual grounding separately, with higher reviewer agreement for correctness than for grounding.The adjudicated pass rates were 88.3% for answer correctness and 78.5% for grounding.Source: §6.1 · p. 12

6.2 Stability across Random Seeds

  • 69.09% ± 0.24 was the three-run R3 mean, compared with 65.57% for Base; every run improved through R3.The weakest run reached 68.84%, and all runs gained more from R2 to R3 than from R1 to R2.Source: §6.2 · p. 13

7 Conclusion and Discussion

MedEvo links local repair of failed reasoning to feedback-guided selection of later training questions. The conclusion emphasizes preserving useful reasoning and evaluating supervision by downstream learning, while identifying boundaries and future directions for the current implementation.

7 Conclusion and Discussion

  • MedEvo improves Qwen3.5-9B, Qwen3.5-35B, and Qwen3.5-397B across seven benchmarks with SFT alone, and the Qwen3.5-9B gains are consistent across three runs.The conclusion distinguishes correcting a response from constructing useful supervision and calls for assessing downstream learning alongside repair success and evidence grounding.Source: §7 · p. 13

7.1 Insights for Medical Self-Evolution

  • Medical tags prioritize questions, while error types guide the local hint and rollback; fixed-set feedback improves benchmark accuracy across rounds, unlike training-error feedback.The taxonomy and error codebook remain fixed while selection priorities and accepted supervision change with the student; the experiments also distinguish answer correctness from visual grounding.Source: §7.1 · pp. 13–14
  • 55.70% repair success for the longest-8k subset was below random selection’s 67.50%, yet its downstream accuracy was higher: 66.78% versus 64.35%.Suffix continuation also outperformed full rewriting in repair success and downstream accuracy; the subset comparison does not isolate trajectory length from question difficulty.Source: §7.1 · pp. 13–14

7.2 Limitations and Future Directions

  • The implementation’s fixed taxonomy and error codebook bound representable weaknesses, and it relies on a strong proprietary auditor for diagnoses.The authors propose studying grounded recovery, expert-validated updates to the knowledge structure, longer independent-data runs, and transfer across modalities, specialties, and institutions.Source: §7.2 · p. 14

Appendix Contents and Reading Guide

  • The appendices map method, taxonomy, dataset, and experimental details, including prompts for feedback planning, auditing, and same-student continuation.The reading guide also points to worked radiograph examples and additional baseline details.Source: Appendix Contents and Reading Guide · p. 19

A.1 Failure Definition and Step Segmentation; A.2 Entropy-Based Localization

  • Failed rollouts are segmented into semantically complete steps, and entropy proposes a local inspection window and provisional rollback boundary for auditor review.The auditor may move the proposed boundary earlier when an entropy surge follows an unsupported premise, or retain a high-entropy step that is valid.Source: §A.1 · p. 20, §A.2 · p. 20

A.3 Gold-Blind Audit and Recovery Operators

  • A gold-blind auditor identifies the first consequential error and chooses a recovery operation; the hint guides correction without exposing the reference answer.Operations include perceptual re-examination, inference reconstruction, differential disambiguation, and a narrowly scoped knowledge cue.Source: §A.3 · p. 20

A.4 Continuation, Admission, and Training Targets; A.5 Feedback Statistics, Quota Allocation, and Dataset Update

  • The same student continues from the retained prefix and hint, with admission checks and learnability gates determining which repairs become training targets.Later rounds allocate question-selection budgets by feedback priorities, retrieve tag-matched candidates, and retain at most one active target per question after blinded comparison.Source: §A.4 · p. 21, §A.5 · p. 21

A.6 Settings of Analysis Experiments

The analysis experiments compare trajectory-repair and feedback-selection variants under specified budgets, then examine entropy and difficulty patterns. The difficulty analysis finds MedEvo ahead at L1–L4, while the HHD-inspired baseline is higher at L5.

Repair designs

  • Repair variants share the first-round rollout and continuation budgets, contrasting full rewrites, no-hint rewrites, and generic-hint suffix continuation.The analysis configuration uses Qwen3.5-9B with the training configuration in Tab. 14 and evaluation in § 4.1.Source: §A.6 · pp. 21–22

Feedback ablations

  • Feedback ablations hold the R1 generation model and historical dataset constant while removing feedback selection, medical tags, or error-type matching.R2 data use the shared R1 model, R3 data use each branch’s own R2 model, and each round’s SFT starts from M0 with a fresh optimizer and schedule.Source: §A.6 · p. 22

Trajectory selection and hint count

  • Trajectory selection compares the 8,000 longest failed trajectories with a uniformly sampled random-8k subset, each trained for one continuation round from M0.The hint-count experiment uses the same longest-8k subset and varies k over {1, 3, 5, 8} with five continuation attempts; k=1 reproduces the 66.78% single-hint baseline.Source: §A.6 · p. 22

Feedback source, retry budget, and entropy

  • Feedback-source experiments replace feedback-set errors with training errors, while retry-budget experiments vary continuation attempts over {1, 3, 5, 8} with and without entropy localization.Final acceptance is measured over 45,000 attempts per setting with 95% Wilson intervals.Source: §A.6 · p. 22

Step-entropy analysis

  • Step-entropy analysis bins min–max-normalized entropy into five equal-width intervals from 0.5 to 1.0 and reports each bin’s erroneous-step rate and step count.The dashed y=100x line is a numerical reference, not an error-probability calibration target.Source: §A.6 · p. 22

Difficulty analysis and Heuristic baseline

  • At L5, the HHD-inspired Heuristic baseline reaches 50.8% accuracy versus 44.8% for MedEvo, while MedEvo is higher at L1–L4.The comparison uses a modality-stratified 2,000-question sample from the 5,000-question feedback set; the sample is only for difficulty analysis, while capability-gap assessment uses the full feedback set.Source: §A.6 · p. 22

B.1 Dataset Roles and Evaluation Scope

The study separates training, feedback, and benchmark roles, then evaluates on seven medical benchmarks using an unweighted mean of benchmark accuracies.

Dataset roles

  • Training questions supply supervision, the fixed feedback set guides diagnosis and selection, and benchmarks measure performance.Test coverage is counted in benchmarks.Source: §B.1 · p. 22, Table 9 · p. 22

Evaluation scope

  • Evaluation uses seven benchmarks: PMC-VQA, VQA-RAD, SLAKE, PathVQA, OmniMedVQA, MedXpertQA-MM, and the medical subset of MMMU.The benchmarks use MedUniEval from Hulu-Med, built on Lingshu’s MedEvalKit.Source: §B.1 · p. 22
  • Benchmark accuracies are averaged without weighting by benchmark size, while modality and error-family results describe the feedback set separately.Source: §B.1 · p. 22

B.2 Round-wise Data Construction

Across rounds, repair workload declines while the training set remains fixed in size. Repair success rates stay stable, and mean attempts per repair decrease.

B.2 Round-wise Data Construction

  • Repair submissions fall from 27,233 to 23,620, a reduction of 3,613 questions (13.27%) relative to R1, while their share of the fixed question bank drops from 33.50% to 29.06%.The workload unit is a submitted question.Source: §B.2 · pp. 22–23
  • The final training set remains fixed at 81,293 questions in each round, combining qualified original correct trajectories, accepted repairs, and retained valid historical targets.Source: §B.2 · p. 23
  • Repair success rates are 71.85%, 70.68%, and 71.11% across rounds, while mean attempts per repair fall from 3.37 to 2.12.The number of successfully repaired questions decreases from 19,567 to 16,797 as fewer questions are submitted; each column identifies the generator and training round.Source: §B.2 · p. 23

B.3.1 Evolution across Twenty Error Subtypes

The section tracks error rates and repair outcomes across rounds, alongside the taxonomy, codebook, and information controls used to characterize and manage the data. Error-subtype rates fall below R0 at R3, but relative reductions and repair success differ by family.

B.3.1 Evolution across Twenty Error Subtypes

  • At R3, all twenty error subtypes are below R0; the knowledge-gap family’s R3/R0 ratio is approximately 0.89, versus 0.72–0.75 for the other families.Figure 7 normalizes subtype bars by R0, retains original percentages as labels, and compares R3 with R0; the relative reduction interpretation assumes a consistent counting denominator across rounds.Source: §B.3.1 · p. 23, Figure 7 · p. 23

B.3.2 Round-wise Feedback Statistics

  • Feedback statistics report modality error rates separately from error-family percentages, which sum the five subtypes in each family and are not directly comparable to modality rates.The tables track consecutive-round changes and highlight low rates and R3 values.Source: §B.3.2 · p. 23, Table 11 · p. 24, Table 10 · p. 24

B.4 Detailed Error Statistics for the Current Rounds

  • Reasoning-error repair success rises from 56.43% at R1 to 83.18% at R3, while knowledge-gap repair success falls from 90.28% to 48.82%.These rates describe submitted repair candidates; Tab. 13 separates repair volume from success rate.Source: §B.4 · p. 24, Table 13 · p. 24

C.1.1 Four-Dimensional Medical Tag Dictionary

  • The medical taxonomy organizes questions along modality, anatomy, disease, and difficulty axes, using hierarchical labels and multi-label assignments.It includes modality, anatomy, and disease branches, and directs annotators to use the deepest supported node rather than guess a leaf.Source: §C.1.1 · p. 25, §1 · pp. 25–26

C.1.2 Error Codebook: Four Parents and Twenty Subtypes

  • The error codebook defines five observable subtypes under each of four parent categories, with boundary rules for primary labels and unresolved audits.It distinguishes visual misreading from omitted inference evidence, cross-image correspondence errors from conflicting extracted evidence, and candidate comparison from missing prerequisite knowledge.Source: §C.1.2 · p. 26, K5 Missing investigation or management… · p. 27

K1 Missing finding interpretation

C.1.3 Information Access Matrix

  • The information-access matrix distinguishes the auditor’s observable inputs from hidden gold answers, reference rationales, and labels, and constrains what reaches continuation.An authorized evaluator may identify a wrong training answer without transmitting the answer; generation receives only a boolean routing decision.Source: §C.1.3 · p. 27

C.1.4 Structured Fields and Plan Validation

  • Structured records capture question metadata, feedback, plans, and repair details, while validation rejects cross-split IDs, exceeded budgets, unsupported hints, and changed hyperparameters.Plans are saved before generation and counts logged.Source: §C.1.4 · pp. 27–28

C.1.5 Training Configuration and Evaluation Scope

This material specifies training and evaluation conventions, then details how MedEvo audits failed trajectories and continues repairs. Case studies illustrate these operations, while related work situates the approach among medical reasoning and localized-repair methods.

C.1.5 Training Configuration and Evaluation Scope

  • Training uses BF16 on GPUs, and the reported evaluation compares each backbone’s best complete checkpoint with Base across seven benchmarks and their equally weighted mean.The MoE configuration table records training settings; the supplied passages do not provide its cell values.Source: §C.1.5 · p. 28, Table 14 · p. 28

D.1 Shared Question-Answering Prompt / D.2 Main-Agent Feedback and Next-Round Planning

D.3 Stage 1 Audit Prompt

  • The auditor uses uncertainty to locate a candidate region, then checks medical evidence and reasoning dependencies to identify the first consequential failure and preserve the valid prefix.It receives no reference answer or answer-derived labels; the entropy proposal is a localization prior, not an error label.Source: §D.3 · pp. 29–31

D.4 Stage 2 Continuation Prompt

  • The same student continues from the retained prefix with a minimal, answer-blind cue, while audit labels remain metadata and are not supplied to the student.The SFT export retains the cue and uses full-response loss on repaired and original correct examples; final evaluation supplies no audit or external hint.Source: §D.4 · p. 31

E.1 Case 08: Chest radiograph

  • In the chest-radiograph case, a hint prompts rechecking basal-opacity morphology, after which the continuation selects pleural effusion.The case is presented as an initial failed-rollout excerpt followed by a hint and continuation; the initial excerpt is not the retained safe prefix.Source: Student model output · p. 32, §E.1 · p. 32

E.2 Case 31: Elbow radiograph / Student continuation after the hint

F.1 Medical Reasoning, Distillation, and Data Selection

  • Prior medical-reasoning work spans instruction tuning, process-level audits, factuality checks, and capacity-aware trajectory selection; MedEvo connects medical error diagnosis with student-centered supervision.The related-work discussion notes that rubric scores can leave some clinical hallucinations undetected [17].Source: §F.1 · p. 35

F.2 Learning from Failed Trajectories and Localized Repair

  • Prior localized-repair methods study hints, backtracking, prefix reuse, and self-correction; MedEvo instead combines medically grounded same-student repair with gold-blind auditing and tag-conditioned hints.The discussion also distinguishes approaches that use preference learning, verified traces, or reflective self-distillation.Source: §F.2 · p. 35

F.3 Data Evolution and Recursive Self-Improvement

The related-work discussion distinguishes MedEvo’s weight-updating, feedback-driven training from evolving external knowledge banks and surveys broader failure-feedback and data-evolution settings. It also describes the external models used to contextualize benchmark comparisons.

F.3 Data Evolution and Recursive Self-Improvement

  • MedEvo diagnoses medical groups and error patterns on a fixed feedback set, then generates and repairs training-pool trajectories for subsequent self-distillation through SFT.This updates model weights, unlike MediSkill-Evo’s evolving knowledge banks without backbone fine-tuning.Source: §F.3 · p. 35

F.4 Failure Feedback beyond Medical Reasoning

  • Structured failure feedback also appears in agent-safety and robotics approaches.Source: §F.4 · p. 35

F.5 Analysis of Frontier-Model Reasoning

  • Frontier-model reasoning work analyzes efficiency, step types, and tree structure alongside outcomes.Source: §F.5 · p. 35

F.6 Broader Data and Environment Evolution

  • Broader recursive-improvement settings include difficulty-guided chart synthesis, multimodal generation, data-strategy optimization, evolving environments, and autonomous medical curricula.MedEvo is described as evolving data for medical reasoning repair and supervision.Source: §F.6 · p. 35

G Additional Baseline Details / G.1 Backbones and External Comparison Models

  • The external comparison set includes general-purpose and medicine-specialized models, grouped by proprietary access and open-source parameter scale.These comparisons contextualize performance and differ from MedEvo’s controlled ablations.Source: §G.1 · p. 36, G Additional Baseline Details · p. 36

G.2 Trajectory-Supervision Baselines

The trajectory-supervision baselines compare hint-based recovery, combining incorrect and correct trajectories, and teacher-generated SFT with MedEvo’s repaired student trajectories. The comparison notes clarify how benchmark averages and component ablations should be read.

G.2 Trajectory-Supervision Baselines

  • HHD provides a hint-based supervision comparison alongside MedEvo’s medical diagnosis, prefix preservation, and feedback-driven data updates.The cited method uses concise hints to scaffold recovery [1].Source: §G.2 · p. 36
  • TrajFusion combines incorrect and correct trajectories, whereas MedEvo constructs a corrected suffix from the same student’s retained prefix and uses medical feedback to prioritize repairs.TrajFusion is presented as a comparison for learning from failed reasoning [10].Source: §G.2 · p. 36
  • Teacher SFT uses Qwen-3.8-Max-generated supervision as a comparison with MedEvo’s repaired student trajectories.The passage describes this baseline as isolating supervision.Source: §G.2 · p. 36

G.3 Reading the Comparisons

  • The seven-benchmark average weights benchmarks equally, while repair success measures the fraction of submitted questions corrected; component ablations isolate specified design choices.Bold and underlining mark the highest and second-highest distinct scores within each model group and column, with ties sharing a rank.Source: §G.3 · p. 36

Related papers

Loading 2610.17690v1…