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

1. Problem Statement

Clinical diagnosis is inherently a multi-turn, uncertain conversation requiring doctors to flexibly employ multiple reasoning paradigms (e.g., step-by-step hypothesis testing, parallel differential diagnosis, urgency-based prioritization, or evidence-driven revision) under incomplete information. While Large Language Models (LLMs) demonstrate strong medical knowledge, existing medical agents struggle to learn these diverse reasoning paradigms simultaneously without interference.

Two primary challenges hinder the development of robust diagnostic agents:

1. Data Synthesis: Generating high-quality, controlled diagnostic dialogues is difficult due to the "information asymmetry" between doctors and patients. A synthesis system needs complete Electronic Medical Records (EMRs) for quality control, but the simulated doctor agent must not see hidden clinical answers (e.g., final diagnosis, lab results) to avoid leaking information or asking leading questions. Existing datasets often lack controlled trajectories or suffer from hallucinations when generated via unconstrained role-play.
2. Model Training: Integrating multiple reasoning paradigms into a single deployable model is non-trivial. Simply mixing dialogues for Supervised Fine-Tuning (SFT) blurs paradigm-specific learning. Conversely, training separate models (or adapters) for each paradigm leads to "gradient conflict," where independently trained branches update in incompatible directions, making them difficult to merge into a single model without performance degradation.

2. Methodology

The authors propose PACT (Periodic Anchor Consensus Training), a framework that couples a novel data synthesis module with a consensus-based training strategy.

A. DPS: Doctor-Patient-Supervisor Data Synthesis

To address the data synthesis challenge, the authors introduce DPS, a distillation-based module inspired by Learning Using Privileged Information (LUPI).

• State Separation: Each EMR is partitioned into patient-visible information ($s_{vis}$: chief complaint, symptoms) and privileged clinical information ($s_{priv}$: hidden exams, target diagnosis).
• Role-Separated Generation: The Doctor agent generates questions based only on the accumulated trajectory and $s_{vis}$, never seeing $s_{priv}$. The Patient agent answers based on $s_{vis}$.
• Supervisor Validation: Two supervisors (Doctor-Supervisor and Patient-Supervisor) utilize the full EMR ($s$) to validate the dialogue. They perform a "PASS-or-REWRITE" check to correct issues like information leakage, pacing, or missing high-value questions without rewriting the entire response.
• Paradigm Conditioning: This process generates validated dialogues under four distinct clinical reasoning paradigms:
  1. Serial: Step-by-step hypothesis testing.
  2. Tree: Maintaining competing diagnoses for differential inquiry.
  3. Heap: Prioritizing questions by urgency and expected information value.
  4. Feedback: Performing self-critique before responding.

B. PACT: Periodic Anchor Consensus Training

To address the training challenge, PACT avoids mixing data or post-hoc merging. Instead, it trains a single deployable model through periodic aggregation.

• Branch Training: The model trains one LoRA Branch (adapter) per reasoning paradigm.
• Anchor Regularization: A shared Anchor ($\Delta_{anc}$) is maintained as the deployable consensus. During local training, each Branch minimizes a loss function comprising the task loss and an L1 regularization term ($\lambda \|\Delta^{(k)} - \Delta_{anc}\|_1$) to keep branches close to the Anchor.
• Periodic Aggregation: Every $T^*$ epochs, the Anchor is updated via a sign-consensus merging process:
  1. Trim: Low-magnitude updates (noise) are removed.
  2. Elect: A majority vote determines the sign of each parameter dimension across branches.
  3. Merge-Project: Only sign-consistent values are averaged, and the result is projected back to the LoRA rank via truncated SVD to maintain the low-rank structure.
• Deployment: After training, all Branch adapters are discarded. The final model is simply the base model plus the final Anchor ($M + \Delta_{anc}$), requiring no architectural changes or routing mechanisms during inference.

3. Key Contributions

1. DPS Module: A data-construction method that synthesizes validated, multi-paradigm dialogues while strictly preserving doctor-patient information asymmetry, preventing label leakage.
2. PACT Framework: A consensus-based training method that learns paradigm-specific LoRA branches and periodically aggregates them into a single deployable model, avoiding the interference of mixed SFT and the incompatibility of independent training.
3. Dynamic Benchmark: The introduction of a dynamic, multi-turn Chinese medical diagnosis benchmark designed to evaluate agents on interactive consultation capabilities (evidence collection, diagnosis, and exam recommendation) rather than static single-turn QA.

4. Experimental Results

The authors evaluated PACT on a dynamic multi-turn benchmark using a GPT-4o-powered patient simulator.

• Performance: PACT (trained on Qwen3-8B) achieved 55.31% Strict Diagnosis Accuracy (DA) and 76.16 Soft Score, significantly outperforming the strongest zero-shot proprietary baseline (GPT-4.1) and medical-specialized LLMs (e.g., Doctor-R1, HuatuoGPT-o1).
  • Compared to GPT-4.1, PACT improved Strict DA by +18.61 points and Exam F1 by +44.06 points.
  • Compared to the Qwen3-8B base model, PACT improved Strict DA by +22.78 points.
• Ablation Studies:
  • Data: Removing any of the four reasoning paradigms from the training data reduced Exam F1 by 5–6 points, indicating the paradigms provide complementary supervision.
  • Training: PACT outperformed Mixed-SFT, Sequential Curriculum, and various post-hoc merging techniques (TIES, DARE-TIES). Notably, standard post-hoc merging (TIES) degraded performance significantly (40.41% Strict DA), validating the need for periodic aggregation during training.
• Generalization: The approach demonstrated cross-architecture generalization on Llama3.1-8B and Ministral-3-8B, and showed transferability to an English-translated setting.

5. Significance and Claims

The paper claims that PACT successfully addresses the dual challenges of data asymmetry and paradigm interference in medical AI.

• Efficiency: Unlike methods requiring complex routing or multiple models, PACT produces a standard single-LoRA model with no additional inference overhead.
• Effectiveness: By explicitly modeling diverse clinical reasoning strategies and enforcing consensus during training, PACT achieves state-of-the-art performance in interactive diagnostic scenarios, surpassing both general-purpose and specialized medical LLMs.
• Scalability: The framework demonstrates that diverse diagnostic strategies can be learned within a single model without the "catastrophic forgetting" or "gradient conflict" typically associated with multi-task learning in medical domains.

The authors emphasize that while symptom gathering is important (where GPT-4.1 performed slightly better), the structured diagnostic reasoning acquired through multi-paradigm consensus training is the critical factor driving superior diagnostic accuracy and exam recommendation quality.