Walking the Tightrope: Autonomous Disentangling Beneficial and Detrimental Drifts in Non-Stationary Custom-Tuning

En Yu (Huazhong University of Science and Technology) · Jie Lu (University of Technology Sydney) · Xiaoyu Yang (Kuaishou- 快手科技)

autoregressive token streamsbiases in predictionschain-of-thought reasoning concept driftconcept graph-empowered llm expertscounterfact-aware reinforcement fine-tuningcounterfactual preference optimizationcounterfactual reasoning trajectoriescxr-counterfactlarge-scale dataset multi-modal large language modelsnon-stationary reinforcement fine-tuningpreference alignmentreasoning token distributionstemporal shiftstheoretical bridge

This paper uncovers a critical yet overlooked phenomenon in multi-modal large language models (MLLMs), especially for chest diagnosis: detrimental concept drift within chain-of-thought (CoT) reasoning during non-stationary reinforcement fine-tuning (RFT), where reasoning token distributions evolve unpredictably, thereby introducing significant biases in final predictions. To address this, we are pioneers in establishing the theoretical bridge between concept drift theory and RFT processes by formalizing CoT's autoregressive token streams as non-stationary distributions undergoing arbitrary temporal shifts. Leveraging this framework, we propose a novel autonomous counterfact-aware RFT that systematically decouples beneficial distribution adaptation from harmful concept drift through concept graph-empowered LLM experts generating counterfactual reasoning trajectories. Our solution, Counterfactual Preference Optimization (CPO), enables autonomous and stable RFT in non-stationary environments, particularly within the medical domain, through custom-tuning of counterfactual-aware preference alignment. Extensive experiments demonstrate our superior performance of robustness, generalization and coordination within RFT. Besides, we also contribute a large-scale dataset CXR-CounterFact (CCF), comprising 320,416 meticulously curated counterfactual reasoning trajectories derived from MIMIC-CXR. Our code and data are public at: https://github.com/XiaoyuYoung/CPO.