Dr. Post-Training: A Data Regularization Perspective on LLM Post-Training

2026-05-07

AI Summary

This work addresses the challenge of balancing scarce high-quality target data against abundant but imperfectly aligned general-purpose data in post-training large language models, aiming to mitigate overfitting. The authors propose the Dr. Post-Training framework, which reformulates general data as a data-induced regularizer. At each optimization step, the method constructs a set of feasible update directions defined by the general data and projects the gradient from the target data onto this set, enabling a flexible bias-variance trade-off. This approach transcends conventional data selection paradigms, unifying existing strategies as special cases while allowing fine-grained control over regularization strength. Experiments demonstrate consistent superiority over state-of-the-art methods across supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and reinforcement learning with verifiable rewards (RLVR), all with low computational overhead and high system efficiency.

Abstract

Data selection methods address a critical challenge in LLM post-training: effectively leveraging scarce, high-fidelity target data alongside abundant but imperfectly aligned general training data. In this work, we move beyond the data-selection framing and introduce Dr. Post-Training (Data-Regularized Post-Training), a novel framework that reconceptualizes general training data as a data-induced regularizer that prevents overfitting to the scarce target objective, rather than serving as a pool for selection. Specifically, our framework proposes that at each training step, construct a feasible set of model update directions using the general training data, and project the model update direction specified by the scarce target data onto that feasible set. Standard training and existing data selection methods arise as special cases with different choices of the data-induced regularizer, and these methods correspond to different points on a bias--variance spectrum with different regularization strength. Building on this view, we propose a family of methods offering a richer design space and more flexible bias--variance tradeoffs. For practical LLM-scale use, we introduce careful system optimizations that realize these methods with minimal overhead. Extensive experiments across SFT, RLHF, and RLVR show that our methods consistently outperform state-of-the-art data selection baselines, and system benchmarks confirm their efficiency.

Problem

• post-training
• data regularization
• large language models
• overfitting
• data selection

Innovation

• Data-Regularized Post-Training
• LLM post-training
• data-induced regularizer
• bias-variance tradeoff
• projection-based optimization