acceptodds
Under review as a conference paper at ICLR 2027

MedRoute: Dynamic Specialist Routing for Multi-Agent Medical Diagnosis

Abstract

Large Multimodal Models (LMMs) have shown promising diagnostic ability in medical reasoning, yet they are typically used as monolithic generalists or as agents assigned to static clinical roles. Existing multi-agent medical systems improve specialization through fixed expert panels, role-based discussion, or external knowledge augmentation, but their coordination is often weakly adaptive and not directly optimized for final diagnostic correctness. We propose MedRoute, a reinforcement-learning-based framework that formulates multi-agent medical diagnosis as a learned consultation policy. For each case, MedRoute constructs a compact, case-conditioned specialist pool and instantiates specialists with targeted clinical guidance. A General Practitioner (GP) then dynamically selects specialists conditioned on the input and accumulated diagnostic history, allowing each consultation to inform the next rather than producing individual expert opinions. To address the lack of step-wise supervision signals, we train the router with an outcome-level policy-gradient objective using group-relative advantage normalization, assigning credit to consultation trajectories based on final diagnostic correctness. A Moderator synthesizes the resulting diagnostic trajectory into the final prediction. Across three text-only and four image-text medical QA benchmarks, MedRoute consistently outperforms both single-model and multi-agent baselines, with gains of up to +22.5% on text-only and +30.4% on image-text over the single-model baseline. Our training code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MultiAgent, MedicalDiagnosis, ReinforcementLearning, DynamicRouting, GeneralPractitioner

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_25kuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper treats multi-agent medical QA as a learned consultation policy. A frozen LMM proposes 5–7 case-specific specialist roles; a trained router sequentially selects either an unconsulted specialist or a stop action, conditioned on a pooled embedding of the accumulated diagnostic record; a Moderator synthesizes the record into the final answer. The router is trained with a group-normalized policy gradient on a binary LLM-judge reward (multiplied by a length decay γ^T), using 300 train questions × 16 rollouts per dataset. It is evaluated on three text-only and four image-text benchmarks against single-model and adapted multi-agent baselines sharing a Qwen3-8B / Qwen3.5-9B backbone, with gains up to +14.1 (MedQA) and +34.3 (BT-MRI) percentage points.

Critical strengths

The paper identifies a real, previously unaddressed problem: the order of agent consultation.

Prior medical multi-agent work fixes the roles and the coordination protocol. MedAgents runs a round table, MDAgents adapts a discussion structure, MAM follows a fixed GP → Specialist → Director template. In all of them which specialist is consulted, in what order, and when to stop is a byproduct of prompt design or a hand-written workflow, never a decision that is made and evaluated. This paper proposes MedRoute and makes it explicit.

Critical weaknesses

W1. Substantially Under-Configured Single-Model Baselines

  • Suspect Baseline Scores: Qwen3-8B is reported at 58.4% on MedQA. Independent literature reports 74.87% [1] and 79.4% [2] on MedQA. If the Qwen3-8B's true performance is ~75%, MedRoute’s 72.5% actually underperforms the vanilla base model.
  • Unfair Compute Comparison: No CoT, few-shot, or self-consistency baselines are provided for single models. MedRoute uses up to 3 specialist calls + 1 moderator makes it an unfair comparison of multi-call aggregation vs. zero-shot direct answer, failing to isolate the gain from routing.

W2. Critical Omissions Regarding Reproducibility

  • Undefined TmaxT_{max}: The maximum consultation budget TmaxT_{max} is never given in the settings.
  • Unspecified Reward Judge (RMRM): The LLM acting as RM(⋅)RM(\cdot) is never named, nor is its prompt, batch size, or judge-consistency provided—despite it being the sole supervisory signal for RL training.

W3. Mechanistic Disconnect from Real Clinical Consultation

Specialists cannot be re-consulted once visited. However, clinical consultation is inherently iterative (e.g., returning to radiology with surgical findings). Masking visited roles reduces the framework to a constrained subset selection over static opinions, undermining the claim of an evolving conversational trajectory.

W4. Missing LLM-as-a-Router Baseline

Using an LLM as a router may lead to better performance and acceptable extra cost.

[1] Chen J, Tang Q, Lu Q, et al. Mol for llms: Dual-loss optimization to enhance domain expertise while preserving general capabilities[J]. arXiv preprint arXiv:2505.12043, 2025.

[2] Zhang Z, Sun L, Qian X, et al. FaithMed: Training LLMs For Faithful Evidence-Based Medical Reasoning[J]. arXiv preprint arXiv:2607.01440, 2026.

Questions for authors

Please solve my concerns of Weakness W1-W4.

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

This paper proposes a reinforcement-learning-based framework that formulates multi-agent medical diagnosis as a learned consultation policy. The framework uses frozen specialists to extract image and text embeddings while only train lightweight adaptors. Three QA and four VQA benchmarks are used to evaluate its medical reasoning ability. Experimental results show that the proposed method can achieve the highest performance across several benchmarks.

Critical strengths
  • Clinical applicability. This work develops a routing approach to plan and select specialists for answering both text-only and vision questions, which would be useful in clinical diagnosis.
  • Lightweight training and practical adaptation. The framework uses frozen specialists to extract image and text embeddings while only train some lightweight modules under a small sample size. This adaption is useful in clinical applications where computing resource is limited and privacy data is small.
  • Broad experimental validation. The authors use three QA and four VQA benchmarks to evaluate their method, and compared with diverse LLMs/VLMs. Experimental results show that the proposed method can achieve the highest performance.
Critical weaknesses
  • Limited methodological novelty. A previous work has studied learning history-conditioned agent orchestration from outcome rewards, as shown in “Multi-Agent Collaboration via Evolving Orchestration [1]” (NeurIPS 2025). The proposed method uses a similar centralized policy, and its technical contribution is limited. Besides, the similar idea of GP-specialist workflow for multimodal medical reasoning has been proposed in MMedAgent-RL [2] (ICLR 2026). The conceptual contribution of this work is also marginal.
  • The ablations are insufficient. In Table 6, the first component addition introduces multiple specialist calls and Moderator aggregation in addition to routing, so its improvement cannot be attributed to the router alone. The paper also lacks matched controls for diagnostic history, consultation order, and stopping. The SFT baseline in Table 4 imitates the first three generated specialists in an arbitrary order, making it a weak comparator. The authors need to use matched settings for fair comparisons.
  • The training objective is ambiguous. It seems that Eq. (3) sums only over the consultations before termination (TgT_g) and omits the log-probability of the sampled termination action. It is unclear if this is a mathematical error or reflects the real implementation. If the latter, the stopping decision then would ignore the immediate terminal-reward.
  • The evaluation setup needs clarification. The last rule in fig 11 would favor answers that list multiple or contradictory alternatives. Have the authors done sufficient quality controls? Please justify that rule and provide evidence on its quality.
  • The statistical difference and computational efficiency are not reported. The proposed method has small gains of 0.6 percentage points on DeepLesion and 1.4 percentage points on PathVQA. It is unclear if the proposed method exists significant difference with others. Please report confidence intervals and statistical test results. Besides, this method involves multiple calls of frozen models, which could largely increase computational burden. Without reporting computational efficiency and comparisons with others, it is difficult to evaluate its practical application value.

[1] Dang Y, Qian C, Luo X, et al. Multi-agent collaboration via evolving orchestration[J]. Advances in neural information processing systems, 2025, 38: 165025-165059.

[2] Xia P, Wang J, Peng Y, et al. Mmedagent-rl: Optimizing multi-agent collaboration for multimodal medical reasoning[C]//International Conference on Learning Representations. 2026, 2026: 81513-81546.

Questions for authors
  • Please clarify the implementation details of the generator. How would the performance changes when the same generated guidance is provided from a single agent to a fixed routing strategy?
  • What is the effect of removing history only from the router while leaving it available to specialists?
  • If answers are enforced to semantically correct rather than merely contain the ground-truth concept, how would the performance change?
  • Would the proposed method outperform a router using a fixed and manually defined specialist pool of the same size?
Reviewer_Dc2nOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a multi-agent reinforcement learning framework for medical question answering. Its innovation lies in using reinforcement learning to coordinate the number and sequence of various medical experts, thereby enabling more comprehensive decision-making. A key strength of this work is the introduction of a more flexible model for multi-agent collaboration.

Critical strengths

A key strength of this work is the introduction of a more flexible model for multi-agent collaboration.

Critical weaknesses

My primary concern regarding this paper relates to the accuracy improvements claimed for the proposed method.

  1. Compared to MAM (Zhou et al., 2025), the proposed method yields an improvement of approximately 5%; however, MedRoute utilizes a weaker base model. Could the authors provide a performance comparison across different methods using the same base model?
  2. The authors do not compare the average number of agents employed by MedRoute against other multi-agent approaches; the computational cost of inference should also be included in the comparison.
  3. While the authors compare their method against a single-model baseline, they fail to report results for a single-model multi-turn dialogue setup—specifically, one that involves the same number of conversational turns as the multi-agent approach.
  4. According to the results in Table 3, the improvement offered by MedRoute on the image-text diagnostic task is quite limited compared to existing methods. Furthermore, why is the gain for the image-text diagnostic task lower than that for the text-only diagnostic task?

Moreover, it is worth noting that the core concept—using reinforcement learning to dynamically select the next agent based on prior outcomes and optimizing the execution sequence—is not novel. Existing literature, such as "Multi-Agent Collaboration via Evolving Orchestration" and "Router-R1: Teaching LLMs Multi-Round Routing and Aggregation via Reinforcement Learning," has already explored this approach. Consequently, the central idea of ​​this paper lacks novelty.

Questions for authors

See Weaknesses.

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

In this manuscript, the authors propose MedRoute, a multi-agent framework for medical question answering. For each case, a frozen LMM generates a pool of five to seven specialist roles together with per-specialist briefings; a small learned router (a GPT-2-small transformer over frozen sentence-transformer and SigLIP embeddings) then sequentially selects the next specialist or a stop action, conditioned on the case and the accumulated consultation history. Specialists, which are prompt variants of a single backbone LMM, answer in turn, and a Moderator synthesizes the record into a final answer. The router is trained with a group-normalized outcome-reward policy gradient on 300 training questions per dataset. The authors report gains over single-model baselines (Qwen3-8B, Qwen3.5-9B) and multi-agent frameworks (MAC, MedAgents, MDAgents, MAM) on three text-only and four image-text benchmarks. Overall, learning the consultation order is an interesting idea, but the current experiments do not isolate the contribution of routing from that of the prompt generator.

Critical strengths
  1. Framing multi-agent consultation as a learned sequential routing policy with an explicit stopping action, trained from final-answer outcomes, is a natural and under-explored alternative to fixed discussion workflows.

  2. Only the lightweight router is trained while all agents stay frozen, which keeps training cheap (300 questions per dataset), and the evaluation spans seven benchmarks including several imaging modalities.

  3. The analyses are informative: random versus supervised versus RL routing (Table 4), a component decomposition (Table 6), cross-dataset transfer (Table 5), and the forced-consultation experiment supporting the learned stopping decisions (Table 10).

Critical weaknesses

Major issues

  1. The gains may be driven by the specialist generator G(·) rather than by routing. The dynamic briefing prompt (Figure 9) instructs G to list two to three differential diagnoses and to “tell them which specific finding or mechanism distinguishes the correct answer”, which amounts to largely solving the question. The model used for G is not stated (it runs on a separate GPU); if it is stronger than the agent backbone, the method partly distills a stronger model. Consistent with this, dynamic prompts and pool generation account for 8.4 of the 14.1-point MedQA gain (Table 6), and random routing within the full system already gains 8.5 points (Table 4). The authors should name G, report results with G set to the agent backbone, add a baseline in which the single backbone receives G's briefing directly, and compare with compute-matched baselines (e.g., self-consistency with the same number of LLM calls) and with the backbone's native reasoning mode.

  2. The comparisons are not on equal footing. MedRoute is trained per dataset on in-distribution questions while all baselines are zero-shot, and several single-model numbers are low for these backbones (e.g., Qwen3-8B at 58.4% on MedQA and 50.3% on PubMedQA), which suggests prompting or answer-extraction issues. DeepLesion and BT-MRI are converted into class-label multiple-choice questions, where answer formatting can dominate (BT-MRI rises from 42.9% to 77.2%). There are no seeds or variance estimates, the router is trained for one epoch, the reward-model judge is unspecified, and inference cost (number of calls and tokens per method) is not reported.

Minor issues

  1. There are several inconsistencies: Section 4 and Appendix A.4 refer to five image-text datasets, but four are evaluated; Figure 4 describes an “axial CT plus ECG” whereas the case concerns an echocardiogram with Doppler; the maximum consultation budget Tmax is never given; and the Table 2 caption contains a grammatical error.

  2. Please define the “static prompt” and “fixed pool” variants in Table 6, explain how the router handles pools whose size and role names differ across cases, and make clear that all specialists share one backbone and differ only in prompts. Related work on learned routing and agent-topology optimization for LLM systems (e.g., GPTSwarm, AgentPrune, RouterDC) should be discussed.

Questions for authors

N/A

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.