acceptodds
Under review as a conference paper at ICLR 2027

CoEvoMed: Co-Evolving Supervision Principles and Medical Vision–Language Models

Abstract

Training medical vision–language models requires supervision that adapts as the learner's capabilities change across tasks. Existing pipelines typically rely on repeated manual error analysis and task-specific training data construction, making this adaptation costly. We introduce CoEvoMed, a co-evolutionary framework that couples a medical vision–language learner with a multi-agent supervision process integrating the evolution of construction principles with curriculum allocation. These reusable principles define question intent, guide how responses draw on medical evidence, and clarify conceptual distinctions in training examples. A revision agent updates these principles in response to the learner's evolving task-specific capabilities, guided by comparisons of factual consistency and task alignment between examples generated from the same medical evidence. Under fixed budgets for each task, curriculum allocation adjusts construction effort according to the learner's capability profile and learning progress. Using CoEvoMed, we train an 8B model supporting both 2D and 3D medical tasks. It surpasses Hulu-Med-7B by 1.15–4.80 percentage points on four 2D VQA benchmarks and achieves higher RaTEScore on MIMIC-CXR and IU-Xray, while delivering strong performance on 3D recognition, question answering, and report generation. Results at 32B further support the scalability of the framework.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Medical vision–language models, Co-evolution, Curriculum learning, Multi-agent systems, Supervision construction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2NuwOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CoEvoMed, a framework that couples medical vision-language model training with multi-agent supervision construction. It revises supervision principles and adjusts curriculum allocation using feedback on the learner’s evolving capabilities. The authors evaluate the framework on 2D medical VQA, report generation, and 3D medical tasks, reporting improvements over several baselines and presenting interventions examining feedback alignment and stage-dependent principle utility.

Critical strengths
  1. Stage-dependent supervision is supported by controlled comparisons. Applying original and evolved construction principles at different checkpoints reveals a reversal in their relative effectiveness. This provides concrete evidence that supervision should adapt to the learner’s training stage, supporting a central motivation of the framework.

  2. Broad medical multimodal evaluation. The experiments cover 2D visual question answering, report generation, and 3D tasks, with additional backbone evaluations. This breadth provides useful evidence that the framework can apply across multiple medical task formats and model architectures.

Critical weaknesses
  1. The experimental controls and mechanism analyses do not sufficiently establish the contribution of co-evolution. The method combines teacher-generated supervision, data verification, principle revision, and adaptive curricula, but does not adequately isolate their effects or those of additional resources. Improvements are difficult to attribute to individual principles, and the relationship between principle-selection criteria and subsequent learning gains remains unvalidated. The substantial degradation of the raw-data training baseline also lacks sufficient explanation.

  2. Evaluation coverage and statistical reliability remain limited. The 2D VQA evaluation lacks expert-level, complex medical reasoning tasks, and independent repeated runs or confidence intervals are not clearly reported. On IU-Xray, the 8B model improves over the strongest baseline by only 0.92 RaTEScore points, while the 32B model performs worse than the 8B model. The stability of the gains and benefits of scaling therefore remain unclear.

Questions for authors
  1. Under the same teachers, data sources, and comparable budgets, could you separately establish the contributions of principle evolution and adaptive curricula, including PPO’s advantage over simpler allocation strategies? Could you also explain the raw-data baseline’s degradation and validate the relationship between principle-selection criteria and subsequent learning gains?

  2. How are patient-, study-, and image-level overlaps excluded across support data, internal validation sets, and external tests, particularly between PubMedVision-IU and IU-Xray? Was R7 selected exclusively using a predefined internal validation criterion?

  3. How many independent runs underlie the main results, and could you provide standard deviations and paired confidence intervals? Additionally, why is CURE-4B absent from the IU-Xray comparison, and why does the 32B model underperform the 8B model on this dataset?

  4. Why were complex reasoning benchmarks such as MedXpertQA-MM or the Health & Medicine portion of MMMU not included? How do the existing results support generalization to complex medical reasoning?

3 comments
AuthorsOpenReview ↗

Thank you for the thorough questions: component ablations, overlap screening, paired confidence intervals, and MMMU

Q1 & W1 (Contributions, PPO, raw data, selection criteria). The paper already ablates both components with the same teachers, data, budget, and R5 checkpoint (Table 2, Sec. 5.3; App. D). We add a uniform split across capability cells as the simpler curriculum, to be added to the revision. Change from R5 on each benchmark:

SettingPrinciple evolutionAdaptive curriculumVQA-RADSLAKEPMC-VQAPathVQAMIMIC-CXRIU-Xray
Neither (new)No (original)No (uniform)−6.21−4.52−10.48−2.84−7.65−4.75
Curriculum onlyNo (original)Yes−5.64−3.60−9.60−2.20−6.44−4.11
Principles only (new)YesNo (uniform)+1.55+0.09+0.12+0.81+1.94+0.59
CoEvoMedYesYes+2.62+0.69+0.52+1.13+2.77+1.28

Of the two, principle evolution has the larger effect, and PPO beats uniform allocation on all six with either principle set.

PPO also optimizes cumulative gain (Eq. 5), and its choices follow the learner's state. After localization regressed at R4 (60.27 → 56.16), it switched to repair-heavy for R5 and R6, and localization recovered to 59.36 and 62.10 (Fig. 6).

The raw-data baseline degrades because its sources (App. D.3) are not task-aligned (58.3% of 2D records are figure descriptions; Fig. 10), not verified, not targeted, and not adapted. CoEvoMed's agents handle each, and it reaches 91.60 on SLAKE (Table 9).

Principles chosen by our criteria do produce later learning gains. In our logs, 12 of 15 accepted revisions turned their target cell's net repair (validation cases fixed minus broken) positive in the next window, versus 6 of 20 eligible principles kept unrevised. By design (Sec. 4.3; App. B), a principle is replaced only when the validation cases it fixes no longer outnumber those it breaks, that is, when it stops helping the learner. A replacement is chosen only if a review confirms it targets the errors the learner still makes.

Q2 (Overlaps; R7). We excluded patient-, study-, and image-level overlaps by design: support, validation, and test data are separate partitions (Sec. 1; App. C.6), enforced in our code. We screened all 11 2D and 5 3D sources against every external test at four levels: IDs (including study, case, and PMC article IDs), image paths, pixel hashes (exact and near-duplicate), and normalized text. All checks found zero overlap, including PubMedVision-IU vs. IU-Xray.

R7 was selected on internal validation only: it has the highest equal-weight mean of QA and report scores, whose per-round change is the planner's reward (App. C.4). External benchmark scores never feed back into training or checkpoint selection (App. A, C.6).

Q3 & W2 (Runs and intervals; CURE-4B; 32B on IU-Xray). Each of four learners (Sec. 5.1; App. B) was trained once, and all 16 comparisons beat the base. Table 1 and Fig. 3 use greedy decoding (App. C.1); for the std, we re-evaluated CoEvoMed-8B (R7) by sampling with three seeds. Paired CIs compare the greedy run with Hulu-Med-7B, rerun with the same prompts and its own evaluation code (MIMIC-CXR: top baseline Qwen3.6-27B):

VQA-RADSLAKEPMC-VQAPathVQAMIMIC-CXRIU-Xray
Mean±std82.63±1.2691.74±0.7167.64±0.3269.17±0.3663.31±0.8368.53±0.74
Baseline72.584.861.458.561.9266.06
95% paired CI[5.99, 14.19][4.52, 9.14][6.08, 7.02][9.76, 12.07][0.71, 2.68][1.15, 3.15]

Sampled means stay within one std of the greedy scores, and every CI excludes zero. The rerun scores below Hulu-Med's published numbers, which Table 1 keeps, so its margins are conservative. On IU-Xray, the gain is 2.16 under the same protocol and 0.92 against the published 67.30.

CURE-4B is out of domain (trained on MIMIC-CXR-derived data), scoring 28.65 zero-shot on IU-Xray, below every model in Fig. 3.

IU-Xray references are short (30.5 words; MIMIC-CXR 53.4): 8B matches them (30.7), 32B writes longer ones (40.1).

Q4 & W2 (Complex reasoning). We chose benchmarks that span the data clinical reasoning works on, from 2D images to 3D volumes and from short answers to full reports:

DataBenchmarks
Chest X-rayMIMIC-CXR, IU-Xray (full reports)
CT, MRI, X-rayVQA-RAD (clinician-authored), SLAKE (knowledge-based)
PathologyPathVQA
Literature figuresPMC-VQA (multiple choice)
3D CT volumesM3D-VQA (64 slices), 3D-RAD (open-ended, with measurements), AMOS-MM (multi-organ reports)

On the MMMU Health & Medicine test set (1,752 questions, public answers, App. C.1 protocol), CoEvoMed-8B scores 62.04, versus 49.66 for its base and 61.30 for MediX-R1-8B, a medical model on the same base. We will add it to the revision.

The results support generalization: CoEvoMed ranks first on all nine benchmarks (Table 1; Figs. 3–4).

Thank you for the responses and additional experiments. The new ablations and MMMU results, together with the clarifications on data separation and checkpoint selection, resolve most of my concerns. I am raising my rating to Weak Accept. I still note that sampling multiple times from a single checkpoint does not establish stability across independent training runs.

AuthorsOpenReview ↗

Thank you for the updated assessment

Thank you for reading our responses and for raising your rating. We are glad that the new ablations, the MMMU results, and the clarifications on data separation and checkpoint selection address most of your concerns. All of them are now in the revised PDF (App. C.2, C.4, C.6, D.4–D.7).

We agree on run-to-run stability. The three sampling seeds measure the decoding variability of one checkpoint, not the variability across training runs, and App. C.4 now states that each learner was trained once. The training-level evidence we have is that the pipeline, run independently on four learners from three model families, improves over each base in all 16 comparisons (Sec. 5.1; App. B).

Thank you again for the constructive review.

Reviewer_jcb6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

CoEvoMed is a framework that jointly adapts supervision construction and curriculum allocation for training medical vision–language models. A multi-agent pipeline generates and verifies evidence-grounded training examples, revises construction principles using learner feedback, and allocates data-generation effort according to learning progress. Experiments across 2D and 3D medical tasks show improvements in visual question answering and report generation, with additional analyses supporting the value of matching supervision to the learner’s current capabilities.

Critical strengths
  1. Automatically constructing training data from a model’s learning feedback is an interesting and practically valuable research direction. The paper combines evolving construction principles with curriculum allocation to adapt both supervision content and data allocation to the learner’s current weaknesses.
  2. The experiments cover 2D and 3D medical VQA and report generation across different model scales and backbones. Comparisons of current versus historical feedback and original versus evolved principles provide evidence that supervision should adapt to the learner’s training stage.
Critical weaknesses

The ablation study does not sufficiently isolate the contributions of the individual components. Budget-matched comparisons with fixed principles and a fixed curriculum, principle evolution alone, and curriculum adaptation alone would help determine whether the gains come primarily from high-quality synthetic data or the proposed co-evolution mechanism.

Questions for authors

Starting from the same initialization, training on raw data reduces performance on all four VQA benchmarks, with SLAKE accuracy dropping from 64.84% to 38.15%. What causes this substantial degradation? Would cleaning and reconstructing the data using fixed construction principles also yield substantial gains? A budget-matched comparison would help clarify the additional benefit of adaptive principle evolution.

2 comments
AuthorsOpenReview ↗

Thank you for the precise suggestions: all four component combinations and a same-budget fixed-principle baseline

We thank the reviewer for the positive assessment and the precise suggestions.

W1 (Isolating the contributions of individual components). The paper already contains budget-matched ablations of both components. The main text tests them on VQA-RAD and MIMIC-CXR (Table 2, Sec. 5.3): Panel A for curriculum adaptation, Panel B for principle evolution. For space, App. D extends both to all six benchmarks (Tables 7–8). We have now run the two missing settings and will add them to the revision, so the table covers all four combinations. All settings start from the same R5 checkpoint and train one round with the same pipeline and budget. To remove principle evolution, we keep the original principles. To remove curriculum adaptation, we split the budget uniformly across capability cells, while construction still uses current feedback. Change from the R5 checkpoint on each benchmark:

SettingPrinciple evolutionCurriculum adaptationVQA-RADSLAKEPMC-VQAPathVQAMIMIC-CXRIU-Xray
Both fixed (new)NoNo−6.21−4.52−10.48−2.84−7.65−4.75
Curriculum onlyNoYes−5.64−3.60−9.60−2.20−6.44−4.11
Principles only (new)YesNo+1.55+0.09+0.12+0.81+1.94+0.59
CoEvoMedYesYes+2.62+0.69+0.52+1.13+2.77+1.28

Of the two components, principle evolution has the larger effect: with either curriculum, the original principles lower all six benchmarks, while the evolved principles raise all six. Curriculum adaptation adds a smaller but consistent gain: with either principle set, it beats uniform allocation on all six. Keeping both gives the largest gain on every benchmark.

This also answers whether the gains come from high-quality synthetic data. That data is itself produced by our agents: without them, raw data lower all four VQA scores despite more examples (Table 9; report scores are new). We also ran fixed principles from the same base under the same total budget (new):

From the baseVQA-RADSLAKEPMC-VQAPathVQAMIMIC-CXRIU-Xray
Raw data, no agents43.0138.1545.5135.9446.6549.18
Fixed principles (new, same budget)74.7288.8861.4962.4662.2761.31
CoEvoMed82.4891.6067.9569.4063.6268.22

CoEvoMed is higher on all six benchmarks (+1.35 to +7.76). Fixed principles help early but hurt from R5 (Table 8), so co-evolution is what sustains the gains. We will add this run to the revision.

Q1 (Raw-data degradation; fixed-principle reconstruction). Most of the raw data teach the wrong task. Most 2D records are figure descriptions (58.3%); the rest mix labels, reports, and question–answer pairs in other datasets' formats, such as OmniMedVQA (Fig. 10; App. C.3). Trained on them without reconstruction (App. D.3), the model mainly learns to describe images instead of answering the benchmark questions. The drop is largest on SLAKE (−26.69), whose answers are short and direct. The raw data are also unverified, untargeted, and static. The table lists these four causes and the CoEvoMed component that addresses each:

Property of the raw sourcesWhere CoEvoMed handles it
Not task-aligned: mostly figure descriptions rather than questionsPrinciples define question intent; the constructor turns evidence into supervision for the target capability (Sections 3, 4.1)
Not verified: image–text associations are never checkedThe verifier checks factual consistency and visual dependence (Section 4.1)
Not targeted: the mix follows source availability, not the learner's needsThe planner allocates a fixed budget to weak and regressing capabilities (Section 4.2)
Not adaptive: supervision stays fixed while the learner changesThe revision agent updates principles from residual errors (Section 4.3)

Reconstruction with fixed principles cannot keep up with the learner. Table 8 (App. D.2) compares fixed and evolved principles from the same R5 checkpoint under the same budget:

ConditionVQA-RADSLAKEPMC-VQAPathVQAMIMIC-CXRIU-Xray
R5 reference73.6189.0064.6065.8059.0061.63
C: R5 / original67.97 (−5.64)85.40 (−3.60)55.00 (−9.60)63.60 (−2.20)52.56 (−6.44)57.52 (−4.11)
D: R5 / evolved (CoEvoMed)76.23 (+2.62)89.69 (+0.69)65.12 (+0.52)66.93 (+1.13)61.77 (+2.77)62.91 (+1.28)

The fixed principles lower all six benchmarks (C), while the evolved principles raise all six (D). Fixed principles handle the first three properties above but not the fourth. The gap between D and C is the additional benefit of adaptive principle evolution.

AuthorsOpenReview ↗

Thank you, and a brief follow-up on our response

Dear Reviewer jcb6,

Thank you again for your positive assessment and precise suggestions. As the discussion period is coming to an end, we would like to check whether our response answers your questions. Following your suggestions, the revised PDF (changes in blue) now includes:

  • All four combinations of principle evolution and curriculum adaptation (W1): principle evolution has the larger effect, and adaptive allocation beats uniform allocation on all six benchmarks with either principle set (App. D.4, Table R4).
  • A budget-matched fixed-principle baseline (Q1): from the same base and under the same total budget, fixed principles trail CoEvoMed on all six benchmarks, by 1.35–7.76 points (App. D.3, Table 9).

We would be glad to answer any further questions. We would appreciate it if you could let us know whether these results address your comments and whether they change your assessment.

Reviewer_pUjhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper propose a framework that can co-evolve the curriculum for learning objective and the medcial VLMs. The authors argue that learning gains and learning task usualy evolve unevenly, and this issue limiting the supervision quality. Therefore they are motivated to propose a data and learning objective construction pipeline and model trianing process framework that can evenly improve both. This framework is designed in a way that for every iterative round, the model will first be validated on a validate set and evaluate which capability is insufficient. And the curriculum desginer will adjust the next round learning objective based on this validation report.

Critical strengths
  1. The concept behind the framework design is acceptable. Bascially, this framework is similar to those auto-research project that trying to propose and validate a scientific project by iteratively improve an algorith. But this framework is more about how to leverage medical context to do such job.
  2. The experiments are very abudant and sufficient to prove their method's strenght and also honestly showing their weakness
Critical weaknesses
  1. To me, the multi-agent framework seems unecessary. The experiments do not validate all critical function of each "agent role". Do we reall y need all of these agents or can we easily find another plausible agent process to substitude them.
  2. The organization of this paper is relative poor. The authors move most of the detail about the framework design into the appendix, including most important deisng such as evaluation task composition.
  3. The method demonstrates reactive recovery from selected capability regressions, but does not establish prevention of catastrophic forgetting or stable balanced convergence. Later-round declines, reliance on R7 checkpoint selection, absence of retention ablations, and lack of standard continual-learning metrics leave the central retention claim only partially supported.
Questions for authors
  1. For medVLSynther results, do you use the same base model to replicate their results?
  2. Would you consider to publicate the data generated in your evolving pipeline?
  3. CoEvoMed relies on a hand-designed capability taxonomy to construct validation cells and guide curriculum allocation, but the paper does not justify why these dimensions and categories are necessary. Some datasets, such as SLAKE, already provide question-type annotations that overlap with categories such as organ recognition, abnormality detection, and localization. The authors should report this mapping, validate the generated annotations, and compare the proposed taxonomy against native dataset labels. Otherwise, it remains unclear whether the gains come from meaningful capability diagnosis or an arbitrary relabeling of existing task categories.
2 comments
AuthorsOpenReview ↗

Thank you for the careful review: per-agent ablations, forgetting measurements, and taxonomy validation

W1 (Necessity of the multi-agent framework). We replaced each agent except the constructor, which writes every example. From the same R5 checkpoint, each row trains one round with one change, all else fixed; entries are changes from R5. Rows 1 and 4 are conditions D and C of Table 8.

From R5VQA-RADSLAKEPMC-VQAPathVQAMIMIC-CXRIU-Xray
Full CoEvoMed+2.62+0.69+0.52+1.13+2.77+1.28
Selector → random evidence−2.00−5.28−1.77−3.62−2.73−4.70
Verifier → none−3.33−2.83−3.13−2.91−4.17−4.42
Revision → original principles−5.64−3.60−9.60−2.20−6.44−4.11
Allocation → uniform+1.55+0.09+0.12+0.81+1.94+0.59
Single agent−10.42−11.59−6.74−7.38−7.93−9.50

Random evidence is drawn from the support pool, ignoring the target cell's modality and anatomy (App. C.3). Without the verifier, all constructed examples enter training. The single agent is one LLM that selects evidence, constructs and checks examples, and revises principles, under the same allocation.

Replacing the selector, verifier, or revision agent turns the round's gain into a loss on all six benchmarks. Uniform allocation still gains, but less on all six. The simplest substitute, a single agent, loses most on five of six. From the base, training on about 700,000 raw examples lowers all four VQA scores (Table 9). Each agent is thus needed; no simpler process we tested substitutes them.

W2 (Organization). The layout follows our contributions and scope. The main text presents each contribution: co-evolution formulation (Sec. 3), multi-agent system and curriculum policy (Sec. 4). Sec. 5.1 gives the evaluation task composition (four 2D VQA, two report, three 3D); for space, main analyses use one VQA and one report task. The appendix adds what reproduction needs: algorithm (App. A), settings (App. B), evaluation (App. C), and mechanism analyses (App. D).

W3 (Catastrophic forgetting and retention). CoEvoMed builds on curriculum learning, not continual learning: it trains all tasks every round and replays earlier data (Eq. 7; Sec. 3.3). We still measured forgetting on validation cases (Eq. 6). Through R7, each round lost 2.3–4.7% of learned cases and fixed 10.9–23.8% of unlearned ones. Average cell-level forgetting (Chaudhry et al., 2018) is 1.27 points at R7 and 2.84 at R10.

Retention is the planner's budget share for learned cells (Sec. 4.2). Zeroing it from R5 (as in W1) loses 6.9% of learned cases instead of 3.3%.

Later-round declines come from the data, not our mechanisms. The support pool is fixed (359,355 facts, App. C.3), and 95% of it was used by R7; few later examples contain new evidence (18.0% at R8, 2.2% at R10). Scores peak at R7–R8, as new evidence runs out. At R10, Fig. 7 benchmarks remain within 3 points of their peaks and over 12 above R1.

We chose R7 on internal validation only: it has the highest equal-weight mean of QA and report scores (App. C.4). InternVL3-9B peaks at rounds 4–6 (App. C.5), so our results do not rely on R7.

Q1 (MedVLSynther base model). Table 1 uses MedVLSynther's published Qwen2.5-VL-7B results. We retrained it on our Qwen3-VL-8B base with its released data and GRPO script. CoEvoMed leads both:

ModelPMC-VQAVQA-RADSLAKEPathVQA
MedVLSynther-7B55.4377.5772.3665.56
MedVLSynther (our base)53.1870.9573.3368.69
CoEvoMed-8B67.9582.4891.6069.40

Q2 (Data release). Yes, upon acceptance, with all principle versions and under each source's license. The data are reproducible now: our code is in the supplement, and all sources are public (Table 5).

Q3 (Capability taxonomy). The dimensions are necessary because every mechanism runs on the cells they define (Table 6). The profile is measured per cell (Eq. 1), the planner allocates budget across cells (Eq. 7), each principle serves one cell and is revised when it stops improving (Sec. 4.3), and the selector retrieves evidence by a cell's modality and anatomy (App. C.3). With the cells fixed, a uniform split gains less than current diagnoses (W1) and stale ones lower 11 of 12 scores (Table 7), so gains come from diagnosis, not relabeling.

We built our categories on SLAKE's (Liu et al., 2021): physicians annotated them, and SLAKE spans more modalities and body parts than prior datasets. They describe what a question asks, not a dataset, so they fit all tasks: VQA-RAD and M3D-VQA use similar types (Lau et al., 2018; Fig. 9a).

On 200 SLAKE questions, our labels match native ones: capability 97.5% (Cohen's κ = 0.958, "almost perfect"; Landis & Koch, 1977), modality 100%, body region 96.5%. We will add them and the new results to the revision.

AuthorsOpenReview ↗

Thank you, and a brief follow-up on our response

Dear Reviewer pUjh,

Thank you again for your detailed review. As the discussion period is coming to an end, we would like to check whether our response addresses your concerns. In brief, the revised PDF (changes in blue) now includes:

  • Necessity of each agent (W1): replacing the selector, the verifier, or the revision agent turns the round's gain into a loss on all six benchmarks, and a single agent loses most on five of six (App. D.5, Table R5).
  • Forgetting and retention (W3): per-round forgetting and repair rates, average forgetting, and a zero-retention ablation (App. D.6), together with the explicit checkpoint selection rule (App. C.4).
  • Organization (W2): Sec. 5.3 and App. D now state which component each experiment ablates, with a map from components to ablations (Table R3).
  • MedVLSynther on our base (Q1) and agreement with SLAKE's native labels (Q3) (App. C.2, C.3).

We would be glad to clarify anything further. If these results resolve your concerns, we would appreciate it if you could consider updating your assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank you to all reviewers: summary of revisions in the updated PDF

We thank all reviewers for their feedback. We have uploaded a revised PDF in which all new or changed text is shown in blue. The blue only helps the reviewers and the AC find the changes; if the paper is accepted, the final version will use the standard text color. Section, table, figure, and equation numbers from the submission are unchanged, so the references in our responses still apply; new tables are numbered R1–R5. The main text remains nine pages.

Main text

LocationChangeReviewers
Sec. 5.3, Table 2States that panel A ablates curriculum adaptation and panel B ablates principle evolution, and points to the further ablations in App. DpUjh, jcb6, 2Nuw

Appendix

LocationChangeReviewers
App. C.2, Table R1MedVLSynther retrained on our Qwen3-VL-8B base, next to its published resultspUjh (Q1)
App. C.2MMMU Health & Medicine results (1,752 questions)2Nuw (Q4)
App. C.3Agreement of our capability labels with SLAKE's native labels on 200 questionspUjh (Q3)
App. C.4, Table R2Standard deviations over three decoding seeds, and paired 95% CIs against Hulu-Med-7B re-evaluated under the same protocol (Qwen3.6-27B on MIMIC-CXR)2Nuw (Q3)
App. C.4Checkpoint selection rule: the highest equal-weight mean of QA and report validation scorespUjh (W3), 2Nuw (Q2)
App. C.6Overlap screening between all support sources and the external test sets2Nuw (Q2)
App. D, Table R3Retitled "Ablations of CoEvoMed's components"; a map from each component to its ablations; D.1 and D.2 titles name the ablated componentpUjh, jcb6, 2Nuw
App. D.3, Table 9Report benchmarks added, and a run with fixed principles from the base under the same total budgetjcb6 (W1, Q1)
App. D.4, Table R4 (new)All four combinations of principle evolution and curriculum adaptationjcb6 (W1), 2Nuw (Q1)
App. D.5, Table R5 (new)Replacing individual agents: random evidence for the selector, no verifier, and a single agentpUjh (W1)
App. D.6 (new)Forgetting measurements and the zero-retention ablationpUjh (W3)
App. D.7 (new)Learning gains after accepted principle revisions2Nuw (Q1)

New references: Chaudhry et al. (2018), Landis & Koch (1977), and Yue et al. (2024, MMMU).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.