DiagLoop: A Counterfactual Data Flywheel with Stage-Localized Reinforcement for Diagnostic LLMs

2026-08-04

AI Summary

Existing causal diagnostic models struggle with effective local deployment due to the absence of interpretable reasoning pathways, scarcity of real-world critical-case samples, and poor cross-scenario transferability. To address these challenges, this work proposes a counterfactual data flywheel mechanism that leverages family-level physical or clinical rules to automatically generate high-quality training scenarios. By iteratively identifying reasoning failure points in a student model through staged analysis and applying bounded repair combined with stage-wise localized reinforcement learning, the approach optimizes the diagnostic reasoning chain. Remarkably, it achieves high-fidelity causal inference using only synthetic data—without requiring expert annotations—and establishes a closed-loop synergy between data generation and model refinement. Evaluated on industrial system monitoring and medical diagnosis tasks, the method improves strict-path accuracy by 11.6 and 5.5 percentage points, respectively, substantially outperforming the strongest baselines and proprietary models.

Abstract

Causal diagnostic models must explain how conclusions follow from evidence because diagnoses guide repairs and treatments. Yet serious cases are scarce, records rarely contain reasoning paths, and data transfer poorly across configurations, complicating local deployment. We present DiagLoop, a counterfactual data flywheel that converts codified physical relations or clinical guidelines, authored once per mechanism family, into training supervision beyond recorded cases. A training-only teacher proposes counterfactual worlds by varying causes, contexts, and observations, while an independent hybrid checker admits only valid worlds. The student reasons through symptom abstraction, causal-chain construction, and root-cause attribution. Stage-specific criteria identify its earliest failure. For nonterminal failures, a bounded repair probes downstream competence, and the resulting weakness profile guides subsequent data generation. Stage-localized reinforcement learning updates only the model-generated continuation, while replay and preservation reduce forgetting. The same criteria govern admission, attribution, reward, and regeneration through checks separate from the proposer. Using only synthesized scenarios and no case-level expert reasoning annotations, the resulting 8B model improves strict path correctness over the strongest conventional baseline. Gains are 11.6 points across eight industrial systems and 5.5 points across ten disease categories. Gains over a deranged-routing control are 3.9 and 2.3 points, respectively. The model also exceeds the evaluated proprietary references in both domains, even when they receive few-shot examples or the specification in context.

Problem

• causal diagnostic models
• counterfactual data
• reasoning paths
• data scarcity
• local deployment

Innovation

• counterfactual data generation
• stage-localized reinforcement learning
• causal diagnostic reasoning
• data flywheel
• hybrid validation