Data Mixing Agent: Learning to Re-weight Domains for Continual Pre-training
This addresses the problem of balancing performance across source and target domains for AI practitioners, offering an automated alternative to manual heuristics, though it is incremental as it builds on existing domain reweighting strategies.
The paper tackles catastrophic forgetting in continual pre-training of large language models by introducing Data Mixing Agent, a reinforcement learning framework that learns to re-weight domains automatically, outperforming baselines in math reasoning tasks and generalizing across unseen fields without retraining.
Continual pre-training on small-scale task-specific data is an effective method for improving large language models in new target fields, yet it risks catastrophic forgetting of their original capabilities. A common solution is to re-weight training data mixtures from source and target fields on a domain space to achieve balanced performance. Previous domain reweighting strategies rely on manual designation with certain heuristics based on human intuition or empirical results. In this work, we prove that more general heuristics can be parameterized by proposing Data Mixing Agent, the first model-based, end-to-end framework that learns to re-weight domains. The agent learns generalizable heuristics through reinforcement learning on large quantities of data mixing trajectories with corresponding feedback from an evaluation environment. Experiments in continual pre-training on math reasoning show that Data Mixing Agent outperforms strong baselines in achieving balanced performance across source and target field benchmarks. Furthermore, it generalizes well across unseen source fields, target models, and domain spaces without retraining. Direct application to the code generation field also indicates its adaptability across target domains. Further analysis showcases the agents' well-aligned heuristics with human intuitions and their efficiency in achieving superior model performance with less source-field data.