Paper 2601.02151
Entropy-Adaptive Fine-Tuning: Resolving Confident Conflicts to Mitigate Forgetting
- Published
- Jan 2026
- Research lab
- Independent
- Citations
- 12
- GitHub
- 105 stars
01 In brief
Summary
The paper investigates why Supervised Fine-Tuning (SFT) causes catastrophic forgetting while on-policy Reinforcement Learning (RL) preserves general capabilities.
The authors identify a distributional gap: SFT data contains 'Confident Conflicts'—tokens with low probability and low entropy, where the model is confident but forced to learn a divergent label, causing destructive gradient updates.
They propose Entropy-Adaptive Fine-Tuning (EAFT), which uses token-level entropy as a soft gating mechanism to down-weight gradients on conflicting tokens while learning from uncertain ones.
Experiments on Qwen and GLM models (4B-32B) across math, medical, and agent domains show EAFT matches SFT on target tasks while significantly mitigating forgetting.
The method is robust to gating variations and computationally efficient via a Top-20 entropy approximation.
Limitations include unsuitability for counterfactual training and dependence on base model calibration.
02 From the paper
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.