PACT: Learning Diverse Diagnostic Strategies via Privileged Synthesis and Branch Consensus

Researchers introduce PACT, a training framework that enables large language models to master multiple diagnostic reasoning strategies simultaneously for clinical decision-making. The method uses supervised dialogue synthesis with complete medical records and a consensus-based training approach, achieving state-of-the-art performance on a new Chinese medical diagnosis benchmark.

PACT addresses a fundamental challenge in medical AI: clinical diagnosis inherently requires flexible switching between multiple reasoning paradigms, yet existing LLM-based medical agents struggle to learn these diverse approaches without mutual interference. The framework represents a meaningful advancement in how AI systems can be trained to handle complex, multi-faceted professional tasks that demand contextual reasoning shifts.

The technical innovation centers on two key components. The Doctor-Patient-Supervisor synthesis approach maintains training integrity by using complete electronic medical records for quality control while strategically hiding information from the model during dialogue—mimicking real clinical constraints. This prevents data leakage while generating validated training examples across four distinct diagnostic paradigms. The Periodic Anchor Consensus Training mechanism trains specialized LoRA branches for each reasoning paradigm, then periodically merges them through sign-based consensus, allowing the model to develop specialized expertise without catastrophic forgetting.

For the AI research community, PACT demonstrates that structured multi-paradigm training with periodic consolidation outperforms naive mixing strategies. The construction of a dynamic multi-turn Chinese medical diagnosis benchmark fills an important gap, as most medical AI benchmarks focus on static question-answering rather than interactive consultation processes.

The broader significance lies in moving medical AI toward more sophisticated, human-like reasoning that mirrors how experienced clinicians actually work. Success metrics spanning both diagnostic accuracy and consultation quality suggest the framework produces more comprehensive improvements than single-objective optimization. This approach may inform training methodologies for other professional domains requiring multi-strategy reasoning—legal analysis, financial advisory, or engineering design.

• →PACT framework enables LLMs to learn multiple diagnostic reasoning paradigms without interference through specialized branch training and periodic consensus consolidation
• →Doctor-Patient-Supervisor synthesis maintains data integrity while preventing information leakage during dialogue-based training
• →New Chinese medical diagnosis benchmark evaluates both diagnostic outcomes and consultation process quality, addressing gaps in existing medical AI datasets
• →Periodic anchor aggregation mechanism provides a scalable approach to multi-strategy learning applicable beyond medical domains
• →State-of-the-art performance demonstrates structured multi-paradigm training outperforms proprietary and medical-specialized baseline models