Pretraining Curricula Enable Selective Fine-tuning

2026-07-06

AI Summary

This study investigates how pretraining curriculum design enhances the precision of fine-tuned models in suppressing undesirable behaviors. By comparing balanced versus imbalanced task sampling strategies in conflicting tasks, the authors find that imbalanced pretraining encourages neural networks to develop internally segregated circuits, thereby strengthening representational disentanglement. Analyses through ablation studies, activation patching, and synthetic linguistic tasks demonstrate that this strategy significantly improves in-context learning capabilities, enabling more selective refusal during fine-tuning and yielding more robust behavioral control in rule-consistency tasks.

Abstract

Transformers follow implicit curricula whereby some tasks are learned before others. However, how explicit pretraining curricula influence learning, generalization, and the selectivity of fine-tuning is unclear. This is important for AI safety, where fine-tuning is used to selectively suppress misaligned behaviors. Here, we compare curricula that pretrain tasks in a balanced (sampled uniformly) or an imbalanced (one task early, the other late) fashion. We show that imbalanced learning of two conflicting copy tasks promotes in-context learning and improves the selectivity of refusal fine-tuning. Ablations and activation patching show that this occurs because imbalanced pretraining encourages tasks to be disentangled in separable neural circuits, whereas balanced training routes both tasks through a common pathway. We extend these findings to a synthetic language learning task involving rule-consistent and rule-violating data, where imbalanced curricula similarly lead to more localized, less entangled rule representations, resulting in more robust rule-following behavior. Together, these results suggest that imbalanced pretraining curricula may be an important tool for promoting disentangled representations, with direct consequences for the precision and reliability of safety fine-tuning.

Problem

• pretraining curricula
• selective fine-tuning
• disentangled representations
• AI safety
• in-context learning

Innovation

• imbalanced pretraining curricula
• disentangled representations
• selective fine-tuning
• neural circuit disentanglement
• AI safety