Compute Aligned Training: Optimizing for Test Time Inference
For practitioners using LLMs with test-time compute scaling, this work addresses a fundamental training-inference gap, though the novelty is incremental as it adapts existing training paradigms.
The paper identifies a misalignment between standard post-training (SFT/RL) and test-time inference strategies that aggregate outputs, and proposes Compute Aligned Training to align training objectives with test-time procedures, resulting in substantial improvements in test-time scaling.
Scaling test-time compute has emerged as a powerful mechanism for enhancing Large Language Model (LLM) performance. However, standard post-training paradigms, Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), optimize the likelihood of individual samples under a base policy, creating a misalignment with test time procedures that rely on aggregated or filtered outputs. In this work, we propose Compute Aligned Training, which aligns training objectives with test-time strategies. By conceptualizing inference strategies as operators on the base policy, we derive new loss functions that maximize performance when said strategies are applied. We instantiate such loss functions for SFT and RL across common test time strategies. Finally, we provide empirical evidence that this training method substantially improves test time scaling over standard training.