# Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters

**Authors:** Charlie Snell, Jaehoon Lee, Kelvin Xu, Aviral Kumar  
**arXiv:** [2408.03314](https://arxiv.org/abs/2408.03314)  
**Date:** 2024-08-06  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.03314v1/

## Key points

- Focus on enhancing performance of Large Language Models (LLMs) by utilizing more test-time computation
- Scaling of inference-time computation in LLMs and its impact on performance
- Analysis of two key mechanisms for scaling test-time computation
- Need for a "compute-optimal" scaling strategy to maximize efficiency
- Notable improvements in test-time compute scaling efficiency compared to baseline methods
- Exploration into sequential and parallel sampling methods for optimal results
- Future directions for improving test-time compute scaling by combining techniques like verifiers with revisions or PRM tree-search methods

## Summary

This paper focuses on enhancing the performance of Large Language Models (LLMs) by utilizing more test-time computation. This is crucial for developing self-improving agents capable of handling open-ended natural language tasks. The study delves into the scaling of inference-time computation in LLMs and explores its impact on performance when a fixed yet significant amount of compute is allocated at test time. By analyzing two key mechanisms for scaling test-time computation, the research sheds light on potential performance levels achievable by LLMs and influences future strategies for pretraining and tradeoffs between inference-time and pre-training compute resources. It also highlights the need for a "compute-optimal" scaling strategy that dynamically allocates test-time compute per prompt to maximize efficiency. Implementing this approach leads to notable improvements in test-time compute scaling efficiency compared to baseline methods. Additionally, an exploration into sequential and parallel sampling methods reveals that finding an ideal balance between these two approaches can yield optimal results. The discussion also touches upon future directions for improving test-time compute scaling by combining various techniques such as verifiers with revisions or PRM tree-search methods. Overall, the study demonstrates that leveraging additional test-time compute through simple methods like revisions and search can outperform investing equivalent FLOPs in pretraining, particularly on certain types of prompts. However, there are opportunities for further research to explore how different approaches can be combined to enhance test-time compute scaling across various scenarios.