FLOP-Efficient Training: Early Stopping Based on Test-Time Compute Awareness

TL;DR

FLOP-Efficient Training addresses the high cost of training large language models by introducing TTC-aware training, which jointly optimizes training FLOPs and test-time compute (TTC). The method collects intermediate validation points, fits an exponential learning-curve model to forecast final accuracy under a budget, and uses a sigmoid model to select an optimal TTC budget $K^*$ that satisfies both accuracy and compute constraints, enabling early stopping. A break-even bound links training and inference costs to quantify when TTC overhead is outweighed by training savings, with an efficient, online curve-fitting approach that requires only small-$K$ evaluations. Empirically, TTC-aware training achieves up to 92% training FLOP reductions while maintaining or improving accuracy across TinyLlama, Pythia, FineMath, and frontier models, yielding faster deployment cycles and more frequent model refreshes.

Abstract

Scaling training compute, measured in FLOPs, has long been shown to improve the accuracy of large language models, yet training remains resource-intensive. Prior work shows that increasing test-time compute (TTC)-for example through iterative sampling-can allow smaller models to rival or surpass much larger ones at lower overall cost. We introduce TTC-aware training, where an intermediate checkpoint and a corresponding TTC configuration can together match or exceed the accuracy of a fully trained model while requiring substantially fewer training FLOPs. Building on this insight, we propose an early stopping algorithm that jointly selects a checkpoint and TTC configuration to minimize training compute without sacrificing accuracy. To make this practical, we develop an efficient TTC evaluation method that avoids exhaustive search, and we formalize a break-even bound that identifies when increased inference compute compensates for reduced training compute. Experiments demonstrate up to 92\% reductions in training FLOPs while maintaining and sometimes remarkably improving accuracy. These results highlight a new perspective for balancing training and inference compute in model development, enabling faster deployment cycles and more frequent model refreshes. Codes will be publicly released.

Figures (10)

• Figure 1: Solve rate (%) versus training FLOPs for TinyLlama on HumanEval. The solid blue line shows the baseline checkpoints without TTC, while the solid orange line shows checkpoints using TTC (Pass@8). The red circle indicates that TTC can achieve accuracy at least equivalent to a fully trained checkpoint without TTC, but with significantly fewer training FLOPs. Discussion and results for other $K$ values (e.g., Pass@4) are in Section \ref{['sec:exp1']}. Based on our TTC estimation method for K, the predicted value is $K^{*}$ = 8.
• Figure 2: Flowchart of the proposed TTC-aware training procedure. The method efficiently selects the intermediate checkpoint with TTC configuration that is superior in terms of FLOPs and accuracy than the fully trained checkpoint, enabling early stopping. Please see Algorithm \ref{['alg:ttc_algo']} and the associated explanation in Appendix \ref{['app:ttc_algo_details']}.
• Figure 3: TinyLlama on HumanEval across Pass@$K$. Intermediate checkpoints can match or exceed the fully trained baseline with far fewer FLOPs, sometimes outperforming the baseline at higher $K$, while the final checkpoint is better only at the cost of much larger FLOPs.
• Figure 4: TinyLlama on Math-500 dataset with different Pass@K. The dotted horizontal line represents the score of the fully trained TinyLlama without TTC. Challenging setup given the TinyLlama available checkpoints accuracy on Math-500 and our compute resources.
• Figure 5: Qwen3-30B-A3B-Instruct on GSM8K (1 epoch). Pass@K over intermediate checkpoints shows non-monotonic behavior; TTC selects an intermediate checkpoint that outperforms the final checkpoint.