MDTeamGPT: Mitigating Context Collapse and Enabling Self-Evolution in Medical Multi-Agent Reasoning

Long, multi-round, multi-role interaction trajectories inevitably lead to severe information dilution and context window overload, triggering context collapse which destabilizes reasoning.

“Long, multi-round, multi-role interaction trajectories inevitably lead to severe information dilution and context window overload, triggering context collapse which destabilizes reasoning.”

• Also stated in: § Tools Usage Long-Term Memory · hedged

  “As trajectory tokens rapidly accumulate due to multi-expert interactions, the system is often forced to compress historical context for subsequent turns.This process frequently leads to coarse summarization, causing the sudden loss of fine-grained details and critical patient information, thereby destabilizing reasoning efficiency and accuracy (Maharana et al., 2024;Zhang et al., 2025).”

• cites Maharana et al., 2024 sourcePartly supported

  Section 6.1 (p. 7) reports that truncated context windows hinder understanding of very long conversations and that long-context models struggle with complex reasoning over recalled information. It does not establish that long, multi-role interactions inevitably cause severe information dilution or a distinct ‘context collapse’ that destabilizes reasoning; the section also reports improvements on some questions with longer context. (LLM, conf=0.87)

  Experimental — accuracy not yet measured
• cites Zhang et al., 2025 sourcePartly supported

  Section 2.2 (p. 3) describes context collapse as information loss and an accuracy drop when an LLM repeatedly rewrites accumulated context. It does not say that long, multi-round, multi-role interactions inevitably cause collapse through information dilution or context-window overload. The Introduction (p. 2) instead argues that long, detailed contexts can be effective. (LLM, conf=0.87)

  Experimental — accuracy not yet measured

Even when image inputs are removed, retained textual descriptions of each case are typically sufficient to support clinically meaningful judgments.

“As a result, even under experimental settings where image inputs are removed, the retained textual descriptions of each case are typically sufficient to support clinically meaningful judgments.”

• cites Schmidgall et al., 2024 sourcePartly supported

  Section 3.7 says earlier experiments supplied measurement results through text, and the Discussion reports diagnostic accuracy in the benchmark. But the paper does not report removing images from the same cases or establish that their retained text was typically sufficient for clinically meaningful judgments. (LLM, conf=0.78)

  Experimental — accuracy not yet measured

The AgentClinic multimodal dataset construction protocol has the key advantage of completeness of clinical textual information.

“The key advantage of this paradigm lies in the completeness of its clinical textual information.”

• cites Schmidgall et al., 2024 sourceCould not verify

  The dataset table (Additional information, page unavailable) describes AgentClinic’s text and multimodal cases, and Appendix H.1 says the MIMIC-IV cases include corresponding lab events and medical records. Neither passage identifies completeness of clinical textual information as a key advantage of the multimodal dataset construction protocol. Searches for that phrase and for complete clinical text in the construction protocol found no supporting passage. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

MDTeamGPT reduces the average runtime per consultation by 15% (from 236 s to 201 s) compared with Simple Voting.

“MDTeamGPT reduces the average runtime per consultation by 15% (dropping from 236 s to 201 s)”

Models with weaker baseline capabilities (e.g., llama3-8B) demonstrate a markedly steeper improvement slope than stronger counterparts such as gpt-4o.

“models with weaker baseline capabilities (e.g., llama3-8B) demonstrate a markedly steeper improvement slope than stronger counterparts such as gpt-4o.”

In the baseline, by round 4, the token count plummeted from 4,746 to 1,584, exhibiting early signs of collapse.

“by round 4, the token count plummeted from 4,746 to 1,584, exhibiting early signs of collapse.”

MDTeamGPT exhibits the lowest variance across all datasets.

“In addition, MDTeamGPT exhibits the lowest variance across all datasets, confirming its strong stability in complex medical consultation scenarios.”

With continuous injection of historical consultation cases, all models exhibit a consistent upward trend in diagnostic accuracy on both MedQA and PubMedQA, eventually achieving stable convergence after accumulating approximately 600 cases.

“with the continuous injection of historical consultation cases, all models exhibit a consistent upward trend in diagnostic accuracy on both MedQA and PubMedQA, eventually achieving stable convergence after accumulating approximately 600 cases.”

MDTeamGPT substantially outperforms existing baselines across both text-based and multimodal tasks.

“Experimental results show that MDTeamGPT substantially outperforms existing baselines across both text-based and multimodal tasks”

Leveraging a knowledge base constructed from MedQA to assist reasoning on PubMedQA yields a 2.2% accuracy gain over the corresponding baseline.

“leveraging a knowledge base constructed from MedQA to assist reasoning on PubMedQA yields a 2.2% accuracy gain over the corresponding baseline.”

MDTeamGPT sharply decreases token consumption by 23% (from 15,203 to 11,701) compared with Simple Voting.

“and sharply decreases token consumption by 23% (from 15,203 to 11,701).”

When a knowledge base derived from PubMedQA is applied to MedQA testing, accuracy improves by 3.8% compared to the baseline without a knowledge base.

“when a knowledge base derived from PubMedQA is applied to MedQA testing, the accuracy improves by 3.8% compared to the baseline without a knowledge base.”

Overall accuracy remains highly stable between 86.8% and 88.1% regardless of team size (k between 2 and 5).

“the results demonstrate that the overall accuracy remains highly stable between 86.8 percent and 88.1 percent regardless of the team size.”

Relative to SeM-Agents, MDTeamGPT achieves gains of 1.3% and 1.8% on PubMedQA in accuracy and F1.

“and 1.3% and 1.8% on PubMedQA in terms of accuracy and F1 score, respectively”

Combining both knowledge bases achieves the best overall performance, with an average accuracy of 87.7%.

“combining both knowledge bases in Experiment 7 achieves the best overall performance, with an average accuracy of 87.7%”

Relative to SeM-Agents, MDTeamGPT achieves gains of 1.7% and 2.2% on MedQA in accuracy and F1.

“Relative to SeM-Agents, which is limited to voting over correct experiences and lacks an aggregation mechanism, MDTeamGPT further achieves gains of 1.7% and 2.2% on MedQA”

On SLAKE, overall evaluation score s reaches optimum at K=5 (6.7).

“the accuracy for open-ended questions (57.3%) and the overall evaluation score (s = 6.7) still reach their optimum at K = 5”

Extreme configurations involving only 2 or as many as 7 agents are rare.

“Extreme configurations involving only 2 or as many as 7 agents are rare”

Under the pressure of the context window, a typical context collapse occurs at the tenth round, where the token count abruptly drops to 2,326.

“Under the pressure of the context window, a typical context collapse occurs at the tenth round, where the token count abruptly drops to 2,326.”

The baseline method exhibits pronounced instability when handling multimodal context.

“As shown in Figure 5, the baseline method exhibits pronounced instability when handling multimodal context.”

On MedQA, specialist selection distribution is primarily concentrated around 3 agents.

“Specifically, the selection distribution on MedQA is primarily concentrated around 3 agents”

Top-K sensitivity analysis shows a pronounced inverted U-shaped performance curve across all four datasets.

“From the experimental results of the Top-K sensitivity analysis, we observe a pronounced inverted U-shaped performance curve across all four datasets (see Figure 9).”

Using ChainKB in isolation is generally less effective than CorrectKB.

“we observe that using ChainKB in isolation is generally less effective than CorrectKB.”

Most cases reach a stable consensus within 2 to 4 rounds.

“Experimental results confirm that most cases reach a stable consensus within 2 to 4 rounds.”

Model performance remains relatively stable within K=4 to K=6.

“and that model performance remains relatively stable within the range of K = 4 to K = 6”

K=5 achieves the best overall performance trade-off across most metrics.

“Specifically, K = 5 achieves the best overall performance trade-off across most metrics”

On SLAKE, overall evaluation score s decreases to 6.5 at K=6.

“and both decrease to 57.1% and 6.5, respectively, when K = 6.”

On PubMedQA and multimodal tasks (VQA-RAD, SLAKE), specialist selection distribution shifts slightly toward 4 to 5 agents.

“whereas for PubMedQA and multimodal tasks (VQA-RAD, SLAKE), it shifts slightly toward 4 to 5 agents.”

As the number of training cases increases from 0 to 2,500, all base models exhibit a significant upward trend across both VQA-RAD and SLAKE datasets.

“As the number of training cases increases from 0 to 2,500, all base models exhibit a significant upward trend across both VQA-RAD and SLAKE datasets, validating the framework's ability to continuously optimize diagnostic performance by leveraging historical experience in multimodal settings.”

PubMedQA and multimodal tasks requiring complex visual parsing need nearly 4 rounds of exploration for deep reasoning.

“whereas PubMedQA and multimodal tasks requiring complex visual parsing need nearly 4 rounds of exploration for deep reasoning.”

The image-only result remains significantly lower than that of the full system.

“However, this result remains significantly lower than that of the full system (Figure 11 (a))”

On VQA-RAD, closed-ended question accuracy peaks at K=5, reaching 82.7%.

“On the multimodal dataset VQA-RAD, the accuracies for open-ended and closed-ended questions also peak at K = 5, reaching 55.1% and 82.7%, respectively, with the composite score s attaining 7.2.”

The imaging-report-only setting trails the full multimodal input by only a small margin.

“We observe that the imaging-report-only setting exhibits surprisingly strong performance, trailing the full multimodal input by only a small margin”

MDTeamGPT achieved favorable scores in core metrics such as accuracy, explainability, and safety, with average scores reaching 3.8.

“MDTeamGPT achieved favorable scores in core metrics such as accuracy, explainability, and safety, with average scores reaching 3.8.”

On VQA-RAD, composite score s attains 7.2 at K=5.

“On the multimodal dataset VQA-RAD, the accuracies for open-ended and closed-ended questions also peak at K = 5, reaching 55.1% and 82.7%, respectively, with the composite score s attaining 7.2.”

On SLAKE, open-ended question accuracy reaches optimum at K=5 (57.3%).

“the accuracy for open-ended questions (57.3%) and the overall evaluation score (s = 6.7) still reach their optimum at K = 5”

On SLAKE, closed-ended questions achieve marginally higher accuracy at K=6 (81.7%) compared to K=5 (81.3%).

“Although its closed-ended questions achieve a marginally higher accuracy at K = 6 (81.7%, compared to 81.3% at K = 5)”

Introducing Residual Context alone without the Lead Physician yields only limited performance gains.

“Specifically, introducing Residual Context alone without the Lead Physician yields only limited performance gains”

The imaging-report-only setting substantially outperforms the image-only configuration.

“and substantially outperforming the image-only configuration.”

The score for Completeness is 3.0.

“Meanwhile, the scores for”

As the trajectory progresses, token consumption accumulates and reaches a peak at the ninth round (approximately 15,736 tokens).

“As the trajectory progresses, token consumption accumulates and reaches a peak at the ninth round (approximately 15,736 tokens).”

On VQA-RAD, open-ended question accuracy peaks at K=5, reaching 55.1%.

“On the multimodal dataset VQA-RAD, the accuracies for open-ended and closed-ended questions also peak at K = 5, reaching 55.1% and 82.7%, respectively, with the composite score s attaining 7.2.”

On SLAKE, open-ended accuracy decreases to 57.1% at K=6.

“and both decrease to 57.1% and 6.5, respectively, when K = 6.”

The Lead Physician's specialist selection distribution shows a consistent tendency to assign 3 to 5 specialists per consultation across MedQA, PubMedQA, VQA-RAD, and SLAKE.

“Observations indicate that, despite the tasks encompassing both unimodal and multimodal contexts, as well as open-ended and closed-ended formats, the agent exhibits a consistent tendency to assign 3 to 5 specialists per consultation.”

Incorporation of external tools leads to comprehensive improvements across fundamental metrics, particularly Accuracy.

“As illustrated in Figure 12, the incorporation of external tools leads to comprehensive improvements across fundamental metrics, particularly Accuracy.”

The score for Diversity is 2.8.

“Meanwhile, the scores for”

Joint input of imaging reports and raw images achieves the best performance across all backbone models.

“As shown in Figure 11, the joint input of imaging reports and raw images achieves the best performance across all backbone models.”

In the absence of an external knowledge base, the Image only setting performs slightly better than Imaging text only.

“Notably, in the absence of an external knowledge base (Figure 11 (b)), the”

The score for Efficiency is relatively low at 2.6.

“However, we observe that the score for”

MDTeamGPT demonstrates strong robustness in multimodal context.

“In contrast, MDTeamGPT demonstrates strong robustness.”

MedQA requires only 2.3 rounds on average.

“MedQA requires only 2.3 rounds on average”

A slight decline in Recall is observed with external tools.

“Notably, however, a slight decline in Recall is observed.”

The proposed framework effectively compensates for inherent disadvantages in parameter scale, enabling smaller models to obtain substantial marginal gains through iterative experience accumulation.

“This finding indicates that the proposed framework effectively compensates for inherent disadvantages in parameter scale, enabling smaller models to obtain substantial marginal gains through iterative experience accumulation.”

The dual knowledge bases successfully bridge potential critical informational gaps when the number of specialists is minimized.

“In contrast when the number of specialists is minimized the dual knowledge bases successfully bridge potential critical informational gaps.”

MDTeamGPT offers a highly cost-effective solution for scalable real-world deployment.

“Consequently, by substantially reducing resource overhead while maintaining high performance, MDTeamGPT offers a highly cost-effective solution for scalable real-world deployment.”

The Lead Physician mechanism and the Residual Context structure effectively filter out irrelevant noise to prevent hallucination cascades when redundant perspectives are introduced.

“When the number of experts increases to introduce redundant perspectives the Lead Physician mechanism and the Residual Context structure effectively filter out irrelevant noise to prevent hallucination cascades.”

CorrectKB and ChainKB function not merely as repositories of consultation instances, but as encapsulations of transferable reasoning capabilities, thereby enabling robust generalization across diverse scenarios.

“By preserving these higher-order reasoning paradigms, CorrectKB and ChainKB function not merely as repositories of consultation instances, but as encapsulations of transferable reasoning capabilities, thereby enabling robust generalization across diverse scenarios.”

Simple linear summarization is insufficient to support long-term medical reasoning.

“Thus, simple linear summarization is insufficient to support long-term medical reasoning.”

MDTeamGPT does not strictly depend on a flawless initial triage stage.

“This confirms that MDTeamGPT does not strictly depend on a flawless initial triage stage.”

The improvements primarily arise from the synergy of lead physician aggregation, residual context, and the dual knowledge base design based on CorrectKB and ChainKB.

“These improvements primarily arise from the synergy of lead physician aggregation, residual context, and the dual knowledge base design based on CorrectKB and ChainKB.”

The full value of historical data can only be realized by jointly leveraging correct-experience guidance and chain-of-thought based reflective error correction.

“confirming that the full value of historical data can only be realized by jointly leveraging correct-experience guidance and chain-of-thought based reflective error correction.”

The system's downstream reasoning process exhibits strong resilience and reliability against initial triage perturbations.

“Ultimately the system's downstream reasoning process exhibits strong resilience and reliability against initial triage perturbations.”

Knowledge retrieval is deferred to the second round to ensure independent reasoning.

“To ensure independent reasoning, knowledge retrieval is deferred to the second round, primarily to resolve conflicts.”

The knowledge base is indispensable in high-precision multimodal diagnosis because it preserves experience in medical image interpretation.

“reaffirming the indispensable role of the knowledge base in high-precision multimodal diagnosis, as it preserves experience in medical image interpretation.”

The significant increase in Semantic Score (s) indicates that multimodal evidence actively retrieved by tools effectively enhances clarity and completeness of reasoning chains, providing a more robust logical foundation for final answers.

“In particular, the significant increase in the Semantic Score (s) indicates that multimodal evidence actively retrieved by tools effectively enhances the clarity and completeness of reasoning chains, thereby providing a more robust logical foundation for the final answers.”

The separation of imaging evidence summaries from original professional reports enables precise experimental control over whether explicit textual descriptions of visual findings are provided to the model.

“This separation enables precise experimental control over whether explicit textual descriptions of visual findings are provided to the model.”

The slight decline in Recall is primarily attributed to enriched evidence chains introducing greater specific constraints into the reasoning process.

“This is primarily attributed to the enriched evidence chains, which introduce greater specific constraints into the reasoning process.”

The model demonstrates strong robustness.

“demonstrating strong robustness”

The system has adaptive capability to dynamically adjudicate optimal team size based on specific case requirements.

“confirming the system's adaptive capability to dynamically adjudicate the optimal team size based on specific case requirements.”

When K is too large (e.g., K=7), disruptive effects of noisy information become apparent.

“whereas when K is too large (e.g., K = 7), the disruptive effects of noisy information become apparent.”

The low Efficiency score is primarily attributed to time cost associated with multi-round multi-agent interactions.

“This is primarily attributed to the time cost associated with multi-round multi-agent interactions”

The case itself possessed strong confounding characteristics, as the patient's clinical presentations (syncope, pulse abnormalities, elevated ESR) simultaneously pointed to multiple possible complications.

“First, the case itself possessed strong confounding characteristics, as the patient's clinical presentations (syncope, pulse abnormalities, elevated ESR) simultaneously pointed to multiple possible complications.”

K=5 achieves the best balance between coverage and signal-to-noise ratio.

“Given that K = 5 achieves the best balance between coverage and signal-to-noise ratio”

MDTeamGPT retains high practical value in real-world clinical settings even when certain modalities are missing.

“This property allows MDTeamGPT to retain high practical value in real-world clinical settings, even when certain modalities are missing.”

Agentic context engineering can effectively ensure coherence and accuracy in long-range reasoning even in scenarios involving complex visual information.

“These findings provide strong evidence that, even in scenarios involving complex visual information, agentic context engineering can effectively ensure coherence and accuracy in long-range reasoning.”

The framework can continuously optimize diagnostic performance by leveraging historical experience in multimodal settings.

“As the number of training cases increases from 0 to 2,500, all base models exhibit a significant upward trend across both VQA-RAD and SLAKE datasets, validating the framework's ability to continuously optimize diagnostic performance by leveraging historical experience in multimodal settings.”

Complementary multi-source knowledge is necessary.

“This finding not only highlights the necessity of complementary multi-source knowledge”

There are clear differences in information density across modalities in medical diagnosis.

“These results reveal clear differences in information density across modalities in medical diagnosis.”

Without knowledge assistance, the model tends to rely on intuitive visual perception rather than obscure clinical terminology.

“This suggests that without knowledge assistance, the model tends to rely on intuitive visual percep-tion rather than obscure clinical terminology.”

Additional redundant information at K=6 introduces reasoning interference, degrading quality of answers to complex open-ended questions.

“the additional redundant information introduces reasoning interference, thereby degrading the quality of answers to complex open-ended questions.”

MDTeamGPT retains excellent performance when extended to multimodal settings.

“Overall, these results provide strong evidence that MDTeamGPT retains excellent performance when extended to multimodal settings.”

MDTeamGPT is reliable in real-world medical evaluation scenarios.

“This validates its reliability in real-world medical evaluation scenarios.”

Within this framework, the visual modality primarily serves a role of evidence anchoring and complementary validation.

“In contrast, within this framework, the visual modality primarily serves a role of evidence anchoring and complementary validation.”

Within multimodal medical contexts, the model can reach its capability boundary through a specific amount of accumulated experience, thereby achieving robust performance convergence with a manageable data scale.

“This suggests that within multimodal medical contexts, the model can reach its capability boundary through a specific amount of accumulated experience, thereby achieving robust performance convergence with a manageable data scale.”

The discussion mechanism requires a strong guiding role to fully realize its effectiveness.

“indicating that the discussion mechanism requires a strong guiding role to fully realize its effectiveness.”

This forced lossy compression of multimodal historical information triggers a cascading effect, leading to a sharp decline in both open-ended recall and closedended accuracy.

“This forced lossy compression of multimodal historical information triggers a cascading effect, leading to a sharp decline in both open-ended recall and closedended accuracy.”

MedQA's lower average rounds are due to its relatively clear diagnostic features.

“due to its relatively clear diagnostic features”

The system effectively achieves a balance between comprehensive cross-disciplinary coverage and avoiding information redundancy.

“This distribution characteristic suggests that the system effectively achieves a balance between pursuing comprehensive cross-disciplinary coverage and avoiding information redundancy.”

K=5 has strong cross-dataset generalization capability.

“we therefore select K = 5 as the default configuration with strong cross-dataset generalization capability.”

Imaging reports distilled by expert clinicians encapsulate highly concentrated diagnostic semantics and are often sufficient to support a complete reasoning chain.

“Imaging reports distilled by expert clinicians encapsulate highly concentrated diagnostic semantics and are often sufficient to support a complete reasoning chain.”

This property enables construction of strict controlled experiments to quantitatively assess benefits or potential noise introduced by visual information.

“This property enables the construction of strict controlled experiments, allowing us to quantitatively assess the benefits or potential noise introduced by visual information.”

The system effectively and dynamically adjusts discussion depth based on case complexity.

“As shown in Table 9 the system effectively and dynamically adjusts the discussion depth based on case complexity.”

When K is too small (e.g., K=3), retrieved evidence is insufficient to provide adequate references.

“Overall, when K is too small (e.g., K = 3), the retrieved evidence is insufficient to provide adequate references”

The MDTeam consultation failure primarily stemmed from limitations of two objective factors.

“This MDTeam consultation failure primarily stemmed from limitations of two objective factors.”

• AssumptionRests on an assumption: The correct answer for the failed case is B: Myocardial infarction.

Simple linear summarization is insufficient to effectively preserve complex interaction histories that include high-density visual descriptions.

“These results indicate that simple linear summarization is insufficient to effectively preserve complex interaction histories that include high-density visual descriptions.”

Increasing retrieval size to K=6 may slightly improve matching of simple closed-ended questions by covering more factual details.

“This phenomenon suggests that while increasing the retrieval size (K = 6) may slightly improve the matching of simple closed-ended questions by covering more factual details”