OptiMer: Optimal Distribution Vector Merging Is Better than Data Mixing for Continual Pre-Training

arXiv cs.AI / 4/1/2026

💬 OpinionSignals & Early TrendsIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper introduces OptiMer, a continual pre-training approach that removes the need to choose fixed data-mixture ratios before training by training per-dataset models and then optimizing composition weights post-hoc.
OptiMer extracts a “distribution vector” from each dataset-specific CPT model to represent the parameter shift induced by that data, and uses Bayesian optimization to find optimal weights for combining these vectors.
Experiments with Gemma 3 (27B) on multiple languages (Japanese, Chinese) and domains (Math, Code) show OptiMer improves performance over data mixing and model averaging baselines.
The method reduces search cost by 15–35× and yields interpretable weights that can be used as effective mixture ratios for improved data-mixture CPT through retraining.
The same pool of distribution vectors can be re-optimized for different objectives without retraining, enabling target-tailored models on demand.

Abstract

Continual pre-training is widely used to adapt LLMs to target languages and domains, yet the mixture ratio of training data remains a sensitive hyperparameter that is expensive to tune: they must be fixed before training begins, and a suboptimal choice can waste weeks of compute. In this work, we propose OptiMer, which decouples ratio selection from training: we train one CPT model per dataset, extract each model's distribution vector, which represents the parameter shift induced by that dataset, and search for optimal composition weights post-hoc via Bayesian optimization. Experiments on Gemma 3 27B across languages (Japanese, Chinese) and domains (Math, Code) show that OptiMer consistently outperforms data mixture and model averaging baselines with 15-35 times lower search cost. Key findings reveal that 1) the optimized weights can be interpreted as data mixture ratios, and retraining with these ratios improves data mixture CPT, and 2) the same vector pool can be re-optimized for a given objective without any retraining, producing target-tailored models on demand. Our work establishes that data mixture ratio selection, traditionally a pre-training decision, can be reformulated as a post-hoc optimization over distribution vectors, offering a more flexible paradigm for continual pre-training.