Entropy-Adaptive Fine-Tuning: Resolving Confident Conflicts to Mitigate Forgetting

TL;DR

This paper introduces Entropy-Adaptive Fine-Tuning (EAFT), a method that uses token-level entropy to reduce catastrophic forgetting during domain adaptation by selectively learning from uncertain samples.

Contribution

EAFT is a novel fine-tuning approach that leverages entropy gating to distinguish between knowledge conflicts and uncertainty, improving retention of general capabilities.

Findings

EAFT matches standard SFT performance on various tasks.

EAFT significantly reduces catastrophic forgetting.

Validated on models from 4B to 32B parameters across multiple domains.

Abstract

Supervised Fine-Tuning (SFT) is the standard paradigm for domain adaptation, yet it frequently incurs the cost of catastrophic forgetting. In sharp contrast, on-policy Reinforcement Learning (RL) effectively preserves general capabilities. We investigate this discrepancy and identify a fundamental distributional gap: while RL aligns with the model's internal belief, SFT forces the model to fit external supervision. This mismatch often manifests as "Confident Conflicts" tokens characterized by low probability but low entropy. In these instances, the model is highly confident in its own prediction but is forced to learn a divergent ground truth, triggering destructive gradient updates. To address this, we propose Entropy-Adaptive Fine-Tuning (EAFT). Unlike methods relying solely on prediction probability, EAFT utilizes token-level entropy as a gating mechanism to distinguish between epistemic uncertainty and knowledge conflict. This allows the model to learn from uncertain samples while suppressing gradients on conflicting data. Extensive experiments on Qwen and GLM series (ranging from 4B to 32B parameters) across mathematical, medical, and agentic domains confirm our hypothesis. EAFT consistently matches the downstream performance of standard SFT while significantly mitigating the degradation of general capabilities.