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

TL;DR

This paper identifies and addresses the challenge of harmful concept drift in multi-modal large language models during non-stationary reinforcement fine-tuning, proposing a novel counterfactual approach to improve robustness and generalization.

Contribution

It introduces a theoretical framework linking concept drift theory with RFT, and proposes Counterfactual Preference Optimization for stable, non-stationary fine-tuning.

Findings

Enhanced robustness and generalization in RFT

Effective decoupling of beneficial and harmful distribution shifts

Large-scale counterfactual reasoning dataset CXR-CounterFact

Abstract

This paper uncovers a critical yet overlooked phenomenon in multi-modal large language models (MLLMs): 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 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 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 contributed 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.