acceptodds
Under review as a conference paper at ICLR 2027

M³DT-RL: Multimodal Multidisciplinary Diagnostic Teaming with Reinforcement Learning

Abstract

Multidisciplinary consultation integrates complementary specialty perspectives, but training a shared policy for its distinct diagnostic roles remains challenging. We study reinforcement learning over the full multidisciplinary diagnostic trajectory, termed *MDT-RL*, for multimodal differential diagnosis (DDx) reranking. MDT-RL faces two challenges: trajectory-level rewards provide coarse credit across stages, while uneven learning progress makes fixed rollout allocation inefficient. We introduce **M³DT-RL**, combining *Modular Stage Training* (MST) and *Mastery-Driven Sampling* (MDS). MST jointly optimizes full MDT trajectories and isolated Initial, Expert, and Final tasks with revision-aware credit, while MDS tracks task mastery and adaptively reallocates rollouts toward insufficiently learned tasks. We further construct **M²DDx**, containing 4,225 acute and emergency-related cases and 22,208 images, with 537 cases published from March 2026 onward reserved for temporal holdout evaluation. Across six multimodal DDx benchmarks, **M³DT-RL-4B** achieves macro-average Hit@1/MRR of **60.89%/75.62%**, and **M³DT-RL-9B** reaches **62.93%/77.12%**. Both outperform the compared commercial MLLMs, larger open general and medical MLLMs, and prompted multi-agent workflows. These results show that multidisciplinary diagnosis depends not only on model scale or agent count, but also on how diagnostic roles and their coordination are trained. Our code is available at: https://anonymous.4open.science/r/M3DT-RL-2460/

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Clinical Reasoning, Multi-images Reasoning, Agentic RL

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MJuBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces M3DT-RL\text{M}^3\text{DT-RL}, a reinforcement learning framework for multimodal differential diagnosis (DDx) reranking. A single shared policy simulates an entire multidisciplinary consultation across four stages: initial triage, specialist selection, parallel specialist consultation, and final evidence-based revision. To resolve coarse credit assignment and inefficient rollout allocation, the authors propose: (1) Modular Stage Training (MST), which co-trains full trajectories with isolated stage tasks using revision-aware credit (rewarding successful rescue and penalizing corruption); and (2) Mastery-Driven Sampling (MDS), which prioritizes tasks at intermediate mastery via an EMA-tracked Gaussian kernel with a positive floor. The authors also construct AcuteM$^2$DDx (4,225 cases / 22,208 images), featuring an independently collected, leakage-free temporal holdout test set. Across six benchmarks, fine-tuned Qwen3.5-4B and 9B variants achieve state-of-the-art results (up to 62.93% Hit@1 / 77.12% MRR), outperforming frontier commercial MLLMs, larger open models (27B/32B), and prompted multi-agent baselines.

Critical strengths
  1. AcuteM2DDx is a substantial and thoughtfully constructed resource. The benchmark contains 4,225 cases and 22,208 images across 14 modalities, with 537 post-March-2026 temporal-holdout cases, explicit diagnostic-timepoint control, leakage-aware SFT/RL partitioning, MONDO-based concept normalization and alias handling, and dual-model verification with selective physician review.

  2. The empirical study is broad and internally consistent. Two backbones × six benchmarks, consistent component ablations (Outcome-only 57.10 → +MST 61.23 → +MST+MDS 63.00 → +SFT 65.51), and a good 4B result (60.89 macro Hit@1) that exceeds 27B/32B general and medical models.

Critical weaknesses

W1. Unrealistic Closed-Set Setup and Risk of Synthetic Artifacts

  • Clinical gap: Real-world emergency differential diagnosis (DDx) is inherently open-ended; assuming a predefined five-candidate set severely oversimplifies clinical decision-making.
  • Distractor artifacts: Distractors were generated by LLMs (Gemini/GPT) and mixed with true labels from clinical reports. Models may achieve high accuracy merely by exploiting stylistic, linguistic, or granularity differences between AI-generated distractors and human-written gold-standard diagnoses, rather than by performing genuine clinical reasoning.

W2. Lack of Statistical Significance and Robustness Evaluation

  • All experimental results (Tables 1 and 2) report only point estimates. Given the inherent randomness of RL exploration and LLM generation, the paper lacks error bars, standard deviations across multiple random seeds, or 95% confidence intervals (CIs) to confirm statistical significance and rule out random variance.

W3. Inconsistencies and Editorial Oversights

  • Inconsistent number of specialty roles: The main text specifies a pool of “12 specialty roles” (line 369), but the consultation prompt in Appendix F.2 explicitly lists 13 distinct roles.
  • Typographical errors:
    • Misspelling in the Figure 1 caption: “commmercial models” (line 046).
    • Section 4.1 contains a duplicated heading: “Implementation Details. Implementation Details.” (line 295).
Questions for authors

Q1: Can the authors respond to the concerns raised in W1 and W2? Specifically, can they provide empirical evidence to mitigate these concerns?

Q2: Is Team Collaboration Really Necessary?

To rigorously distinguish whether the performance gains stem from the multidisciplinary team structure or merely from increased inference compute or data curation, can the authors provide the following two single-agent baselines:

  1. Token-Matched Self-Consistency: Measure the average test-time token budget consumed by a full MDT consultation, and evaluate a single-agent baseline using self-consistency methods (e.g., majority voting), scaled to roughly match the above inference token cost.

  2. Symmetrically Trained Single Agent (SFT + RL): Train a single-agent model under the same data recipe:

    • SFT: Fine-tune on high-quality single-physician diagnostic reasoning trajectories distilled from the 27B model series.
    • RL: Use outcome-based reinforcement learning with a binary (0/1) reward based on Top-1 diagnostic correctness.
1 comment
AuthorsOpenReview ↗

Response to MJuB — candidate artifacts, uncertainty, and single-physician controls

W1 and Q1. Clinical scope and candidate artifacts. Our task is to rerank a supplied differential diagnosis list. Open-ended generation of that list is a separate clinical task. To test your concern about the candidates themselves, we removed the history, images, and image identifiers from 1,461 cases in AcuteM²DDx, MedReaMM, and MRareBench. Each case had one source-derived reference diagnosis and four model-generated distractors. The ranking prompts and scoring rules were retained.

Candidate-only macro Hit@1 was 22.81% and 19.66% for the direct 4B and 9B backbones, and 57.69% and 56.69% for the corresponding follow-up MDT models. Random choice gives 20%. On MedReaMM, simply choosing the longest candidate achieved 60.59% Hit@1. Restoring the clinical inputs to the same trained checkpoints raised their three-set macro Hit@1 to 64.16% and 65.02%, with improvements for both models on every dataset. The candidate lists therefore contain predictive signals. Restoring the history and images together adds diagnostic information. This experiment does not distinguish AI authorship, diagnostic specificity, ordering, or other candidate features as sources of the candidate-only signal (Conclusion and Appendix I.7, Table 15).

W2 and Q1. Uncertainty. We used 20,000 within-dataset case-bootstrap replicates, weighting datasets equally and resampling matched cases together for comparisons. Appendix I.4 reports 95% Hit@1 intervals for every row of Tables 1 and 2, and MRR@5 intervals wherever the required ranks were retained. Six-set Hit@1 was 62.93% [61.17, 64.71] for 9B and 60.89% [59.09, 62.66] for 4B. These intervals measure test-case uncertainty. The paired Hit@1 differences for MDS and the roster reward were +1.77 [−0.05, 3.56] and +0.49 [−1.32, 2.28] points. An independent rerun of the 9B backbone gave six-set MRR@5 of 62.91% [61.59, 64.23], compared with the original 63.01%.

We also repeated the revision-credit ablation. Its five-set Hit@1 effect was +2.68 points in the original pair and −1.75 in the repeat. The two pairs used different GPU configurations. MST and MDS still have one training run per configuration (Appendix I.5).

Q2. Token-matched single-physician self-consistency. We retrained Table 2's Single-Agent RL baseline for 100 updates from base Qwen3.5-9B without SFT. Using its original single-physician prompt with full histories and images, we drew four to six samples per case at temperature 0.7 against each dataset's measured MDT token budget. Valid draws voted for their first choice, with ties broken by summed ranks. The comparator is the main 9B step-140 MDT. Scores below are percentages.

DatasetVote Hit@1MDT Hit@1Vote MRR@5MDT MRR@5
AcuteM²DDx61.4573.3775.8884.01
DDx-TRACE69.6371.2081.2582.75
MRareBench46.8251.5165.0470.30
MedReaMM73.4477.4483.1586.87
MedThinkVQA40.8354.0362.6371.07
Five-set macro58.4465.5173.5979.00

Voting consumed 23.19M input plus output tokens and MDT consumed 23.73M. The paired MDT advantages were 7.07 Hit@1 points [4.64, 9.52] and 5.41 MRR@5 points [3.86, 6.97] (Appendix I.6, Table 14). Training also differs between these systems. For your second control, the 27B teachers produced eligible single-physician demonstrations for 2,731 of the original 3,019 SFT cases. The single-physician SFT plus binary top-1 RL pipeline has no completed evaluation to report.

W3. Editorial corrections. The main text and the consultation prompt in Appendix F.2 now list the same 12 specialist roles. The ED attending is a separate role. The Figure 1 caption now reads “commercial models”, and Section 4.1 has a single “Implementation Details” heading.

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

In this work, the authors proposed M3DT-RL, a reinforcement learning framework for training a single shared policy across the full multidisciplinary diagnostic trajectory (MDT-RL) in multimodal differential diagnosis reranking. The approach addresses two core difficulties: trajectory-level rewards that assign only coarse credit across diagnostic stages, and uneven learning progress that renders fixed rollout allocation wasteful. M3DT-RL combines Modular Stage Training (MST), which jointly optimizes complete MDT trajectories alongside isolated Initial, Expert, and Final tasks using revision-aware credit assignment, with Mastery-Driven Sampling (MDS), which monitors per-task mastery and adaptively redirects rollouts toward under-learned tasks. The authors also introduce M2DDx, a dataset of 4,225 acute and emergency cases with 22,208 images, reserving 537 cases published from March 2026 onward as a temporal holdout.

Critical strengths
  • The author introduces M2DDx dataset alongside the method, which makes it a valuable addition on top of the training pipeline the author proposed.

  • The paper provides complete ablations to isolate the effect of each of the components. The improvements in Figure 4 and table 1 seems pretty promising, especially considering a 4B model can match or provide better performance than frontier models. The ablation isolates Single-Agent RL, Outcome-Only MDT-RL, Prompted MDT, MST, MDS, and SFT init, and Figure 4 also reports oracle results as a ceiling. These results helps us understand the effectiveness of each component better, while pointing out spaces for improvements.

Critical weaknesses
  • The work seems to have a lot of overlaps over prior works. For example, MDS is essentially a difficulty based curriculum learning conditioned on the EMA of pass rate. This has been explored in quite a few other workds, like Reinforce-Ada (https://arxiv.org/abs/2510.04996) and AdaRFT (https://arxiv.org/abs/2504.05520). Reinforce-Ada in particular is cited but not compared in experiments. It would be nice if the authors can elaborate on the novelty of the method, and compare with more related methods instead of simple GRPO.

  • MDT seems not improving the performance much. For example, Fig. 4(a) shows the final-minus-initial Hit@1 gap is at most 0.62 points, and Fig. 5(d) shows final Hit@1 stays within 65.2 to 66.5 from 1 to 11 experts. Oracle results do show pretty good improvements though, so that's good, but multiple experts, at least under this paper's settings, seems not to be very useful.

  • It would be useful to include a few compute analysis of the method. Multi agents naturally increase turns of conversation for each training/eval step, and it's useful to do compute aligned comparison instead of step aligned (as each step under multi agents setting may cost much more)

Questions for authors

Please see weaknesses

1 comment
AuthorsOpenReview ↗

Response to RkkP — adaptive sampling, expert contributions, and measured compute

W1. Comparison with Reinforce-Ada and AdaRFT. Using recent success to guide training is shared with prior work. The distinction is what the estimate controls. MDS selects which case and stage to revisit. Initial ranking and Final revision for the same patient have separate success histories, and every selected task receives four rollouts. AdaRFT selects problems near a target on a precomputed difficulty scale and updates that target using batch reward. In our MDT adaptation of Reinforce-Ada-Est (Ada-EMA), tasks are drawn uniformly, estimated pass rates allocate two to eight rollouts per task, and the observed group pass rate rescales the advantages.

We ran Ada-EMA from the same 9B SFT checkpoint with the same task pool, revision reward, 140 updates, and 8,960 optimized trajectories as MDS. Five-set Hit@1/MRR@5 was 65.51%/79.00% for MDS and 64.42%/78.27% for Ada-EMA. The paired differences (Ada-EMA minus MDS) were −1.09 Hit@1 points [−2.79, 0.62] and −0.73 MRR@5 points [−1.77, 0.31]. MDS had higher Hit@1 on four datasets and tied on two. It had higher MRR@5 on all six and used 37.62M optimization sequence tokens versus 42.72M for Ada-EMA. At this trajectory budget MDS used fewer tokens and achieved competitive scores. Both paired intervals include zero (Appendix I.5, Table 13).

W2. What do additional experts contribute? From one to eleven specialists, Oracle Hit@1 rose from 70.94% to 78.08%, while Final moved from 65.23% to 65.66% and peaked at 66.52% (Fig. 5d). We examined the main 9B predictions across all six datasets. Among initially incorrect cases, Final recovered 178/295 (60.34%) when at least one specialist ranked the reference first, and 11/1,868 (0.59%) when none did. Correct specialist suggestions were associated with successful corrections. Replacing the specialists with GPT-5.4-mini raised Final Hit@1 by 0.56 points on AcuteM²DDx and 2.67 on MRareBench (main text, RQ4). The paired outcomes also show the cost of revision. On the five ablation datasets, the main 9B model corrected 92 initial errors and lost 127 initial hits (Appendix I.3, Table 9).

W3. Actual compute. Appendix I.1 reports training tokens, allocated GPU-hours, and per-case inference usage. We compared the main 9B MDT with a separately trained single physician voting on the same 2,372 cases. MDT used 23.73M input plus output tokens and scored 65.51% Hit@1. Voting used 23.19M and scored 58.44%. Voting used 2.26% fewer tokens overall, with per-dataset deviations from −8.21% to +18.08%. The paired difference favored MDT by 7.07 Hit@1 points [4.64, 9.52] (Appendix I.6, Table 14). The inference comparison therefore uses approximately matched tokens. The training comparison above holds the trajectory count fixed and reports the actual tokens processed by each method.

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

The paper trains one shared multimodal policy to perform every role of a multidisciplinary-team workflow (initial diagnosis, specialist selection, specialist assessment, final adjudication) for five-candidate differential-diagnosis reranking. After SFT initialization, RL combines Modular Stage Training (full trajectories plus isolated stage tasks, with Final-stage revision credit) and Mastery-Driven Sampling (rollout allocation by recent task success). The authors also build AcuteM²DDx (4,225 cases, including 537 temporally held-out test cases). Across six datasets the 9B model achieves the best macro-averaged Hit@1/MRR among evaluated models (62.93%/77.12%), +18.02/+14.11 points over its backbone.

Critical strengths
  1. Large, broad gains and a carefully built benchmark. Improvements over the backbone hold on all six datasets (Table 1), and AcuteM²DDx uses source/image-aware splits and a temporally held-out test set (Appendix B). This establishes that the complete training pipeline works on the constructed task, although about 80% of the Hit@1 margin over GPT-5.4-mini comes from MedThinkVQA and AcuteM²DDx.
  2. Ablations support the two training mechanisms. Without SFT initialization, MST adds 4.13 Hit@1 points and MDS another 1.77; removing revision credit from the complete (SFT-initialized) model costs 2.68 points (Table 2, Fig. 4c). No seed variance is reported for these differences.
Critical weaknesses
  1. The gains cannot yet be attributed to multi-agent coordination, which the paper's conclusions emphasize. Single-Agent RL alone reaches 57.87% five-dataset Hit@1 (backbone 45.97%, full method 65.51%; Table 2), yet its SFT data, initialization and budget are not described, and MST's isolated stage tasks change supervision and generation cost. Consultation's own benefit is unclear: in the cumulative-stage experiment (SFT/MDS settings unspecified, so not attributable to the final model), Final−Initial Hit@1 ranges from −0.47 to +0.62 points (Fig. 4a), and final Hit@1 stays at 65.23–66.52% from 1 to 11 experts (Fig. 5d). Without a matched single-agent or self-revision control, the results support the complete pipeline but do not isolate the team design's contribution.
  2. The increment over the closest trained-agent methods is under-specified. MMedAgent-RL (ICLR 2026) already RL-trains triage/attending roles with a curriculum based on specialist accuracy, and MAPoRL (ACL 2025) already rewards corrections and penalizes incorrect revisions; both are cited. The identifiable differences (a shared policy for specialists, joint full/isolated tasks, instance-level MDS, Final-only credit after normalization) should be delineated and, where feasible, compared against these designs (e.g., with a plain extra Final top-1 credit), not only through the removal ablations in Table 2 and Fig. 4c.
Questions for authors
  1. (Clarification) For Single-Agent RL and each MDT ablation, what SFT data, initialization, and generation/optimization token and compute budgets were used? Please also give the MDS/reward coefficients, isolated-context construction, checkpoint-selection rule (no validation split is described) and seed-level uncertainty.
  2. (Clarification) What are the final model's paired Initial→Final transitions, training team size and inference calls/tokens, and each model's protocol-failure rate? Which settings and data split produced Fig. 4?
2 comments
AuthorsOpenReview ↗

Response to UGYE (1/2) — training gains, consultation, and revision credit

Thank you for asking us to separate the effect of training the policy from the benefit of consulting specialists at test time.

W1. What accounts for the gain? On the five Table 2 benchmarks, Single-Agent RL scored 57.87% Hit@1 and Outcome-Only MDT-RL scored 57.23%. Auditing the latter's MedThinkVQA results corrected its Hit@1/MRR@5 to 42.78%/62.41%. The resulting Hit@1 gain from adding MST is 4.00 points (57.23% to 61.23%). Adding MDS gives another 1.77 points (63.00%, paired 95% CI for the gain [−0.05, 3.56]). The SFT checkpoint scored 59.69% before RL and 65.51% after full training, a paired gain of 5.82 points [3.66, 8.04].

We retrained the Table 2 single physician and tested self-consistency on the same 2,372 cases. Across five datasets, voting scored 58.44% Hit@1 and 73.59% MRR@5 using 23.19M input plus output tokens. The main 9B MDT scored 65.51% and 79.00% using 23.73M. The paired MDT advantages were 7.07 Hit@1 points [4.64, 9.52] and 5.41 MRR@5 points [3.86, 6.97]. This comparison favors the complete MDT system at a similar inference budget. Its training also differs from that of the single physician.

Within the main 9B model, consultation changed five-set macro Hit@1 from 66.61% at Initial to 65.51% at Final. It corrected 92 initial errors and lost 127 initial hits. For 4B, macro Hit@1 changed from 63.61% to 63.96%. In a separate experiment that held the MDT-trained 9B weights fixed, Initial voting reached 65.59% versus 63.54% for full consultation (vote minus MDT +2.04 points [0.59, 3.51], from unrounded scores). The complete system therefore outperformed the trained single physician at a comparable inference budget (Appendices I.3 and I.6).

W2. What differs from MMedAgent-RL and MAPoRL? MMedAgent-RL trains triage and attending physicians using externally supplied specialist opinions and adjusts the attending's entropy bonus with specialist accuracy. Our shared policy also learns to generate the specialist reports. It trains on complete consultations and on isolated Initial, Expert, and Final decisions, with mastery tracked separately for each task instance. MAPoRL conditions correction rewards on the peers' previous majority and also rewards influence on subsequent answers. Our credit uses the attending's Initial-to-Final change and the fraction of specialists whose top choice was wrong. We center it within the group and add it only to the Final advantage after normalization.

We tested your suggested plain Final top-1 bonus of 0.30. Five-set Hit@1 was 65.51% with revision-aware credit, 63.96% with the plain bonus, and 62.83% without extra Final credit. Plain minus revision-aware was −1.55 points [−3.19, 0.14], and plain minus no extra credit was +1.13 [−0.51, 2.77]. Both paired intervals include zero. For revision-aware minus no extra credit, the original pair gave +2.68 points [0.96, 4.42]. A second pair, trained from the same SFT initialization with a new seed, gave 63.16% versus 64.91%, a difference of −1.75 [−3.50, −0.02]. The direction varied across the two paired runs.(Appendices I.4–I.5).

AuthorsOpenReview ↗

Response to UGYE (2/2) — training budgets, checkpoint selection, and Figure 4

Q1. Data, initialization, and budgets. Both main policies and their reference models started from the corresponding epoch-2 SFT checkpoint (step 644). We distilled Initial, Expert, and Final targets from Qwen3.5-27B, Qwen3.6-27B, and Qwen3.8-27B. Retained trajectories had valid first attempts and a correct Final top choice. Initial and Expert targets additionally required a correct top choice at their own stage. Single-Agent RL, Outcome-Only MDT-RL, MST, and MST+MDS in Table 2 started from base Qwen3.5-9B without SFT. The first two were evaluated at update 100 and the latter two at update 160. At 64 trajectories per update, the corresponding nominal budgets are 6,400 and 10,240 trajectories, compared with 8,960 for the main models.

An isolated Expert receives a specialty and, when available, a previously generated ranking with probability 0.75. Otherwise the reference diagnosis is placed uniformly among the five ranks. An isolated Final task receives an initial ranking and two reports, with confirm/correct/resist contexts sampled in proportions 0.3/0.4/0.3. Appendix H.1 gives the reward and MDS coefficients. The correction, adoption, and corruption magnitudes are 0.30, 0.08, and 0.30. The MDS EMA rate, kernel center, width, and floor are 0.25, 0.5, 0.2, and 0.05.

We selected RL step 140 using the highest recorded Hit@1 on the 537-case AcuteM²DDx temporal test split. At step 140, 4B scored 72.07% and 9B scored 73.37%. These cases therefore contributed to checkpoint selection.

The main 4B/9B runs used 11.79/17.08 allocated GPU-hours and 38.40M/37.62M optimization sequence tokens. Recorded generation totaled 3.15M/3.24M output tokens, including unused trajectories. For the 100-update single-agent rerun, 6,400 trajectories used 10.10M optimization sequence tokens, including 2.38M loss-bearing response tokens. Generation logs record 2.51M output tokens across 6,694 of 6,695 calls.(Appendix I.1).

Q2. Calls, failures, and Figure 4. The main models trained and ran full consultations with two specialists. A complete consultation requires five calls before repair. Across 5,223 cases, 9B averaged 4.946 calls, 8,035.50 input tokens, and 661.57 output tokens per case. The corresponding 4B means were 4.931 calls, 7,943.38 input tokens, and 621.18 output tokens. Table 8 reports the protocol-failure count for every model (column P, denominator 5,223), separately from candidate mismatches, truncation, service failures, and workflow errors. The 9B and 4B protocol-invalid rates were 136/5,223 (2.60%) and 192/5,223 (3.68%). Table 9 gives the paired Initial-to-Final counts by dataset.

Figure 4(a–c) uses step-140 models on the five Table 2 test sets, with two specialists and temperature 0. The first three cumulative-stage runs in (a,b) used base 9B and difficulty sampling. The fourth run's saved predictions reproduce +0.46 points for Final minus Initial and 69.11% Oracle Hit@1, but its initialization and sampling configuration could not be recovered. The reward ablations in (c) used the same SFT initialization and RL task pool. Panel (d) uses the 537 AcuteM²DDx test cases, evaluated every 20 updates from 20 to 160 in the MDS comparison runs.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers for recognizing the value of AcuteM²DDx and for asking what drives performance in five-candidate differential diagnosis reranking. Their questions led us to test the contribution of consultation, compare alternative training objectives, and measure how much information the candidate list itself provides.

Training and consultation (Tables 9 and 14). On 2,372 cases across five datasets, the main 9B MDT achieved 65.51% Hit@1, compared with 58.44% for a retrained single-physician RL model using self-consistency voting. The two systems used 23.73M and 23.19M inference tokens, respectively. The paired MDT advantage was 7.07 percentage points for Hit@1 (95% CI [4.64, 9.52]) and 5.41 points for MRR@5 [3.86, 6.97]. This comparison evaluates the complete systems at similar inference cost. The MDT received SFT followed by RL, and the single physician started RL from the base model. Within the main MDT models, consultation changed five-set Hit@1 from 66.61% to 65.51% for 9B and from 63.61% to 63.96% for 4B. The contribution of consultation at inference therefore varied across the two models.

Adaptive sampling and revision rewards (Appendix I.5, Table 13). With the same 9B initialization and 8,960 training trajectories, MDS achieved 65.51% Hit@1 using 37.62M optimization sequence tokens. Our adaptation of Reinforce-Ada-Est (Ada-EMA) achieved 64.42% using 42.72M. MDS processed fewer tokens, and the paired 95% interval for the Hit@1 difference included zero. The suggested plain Final top-1 bonus yielded 63.96% Hit@1, compared with 65.51% for revision-aware credit and 62.83% without extra Final credit. The effect of revision-aware credit relative to no extra credit was +2.68 points in the original paired runs and −1.75 points in the replication.

Candidate cues and clinical context (Appendix I.7, Table 15). On 1,461 cases with one source-derived diagnosis and four model-generated distractors, the follow-up 4B and 9B models achieved 57.69% and 56.69% Hit@1 from the candidates alone. Random choice gives 20%. On MedReaMM, choosing the longest candidate alone yielded 60.59%. Restoring the history and images to the same checkpoints raised macro Hit@1 to 64.16% and 65.02%, with gains on all three datasets. These results confirm that candidate wording can help identify the reference diagnosis and that patient context provides additional information for ranking.

The revision also identifies the temporal test split used to select step 140 (Appendix H.1) and reports measured computational costs, per-model failures, paired Initial-to-Final outcomes, and case-bootstrap confidence intervals (Appendices I.1–I.4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.