Summary of Large Language Monkeys: Scaling Inference Compute with Repeated Sampling, by Bradley Brown et al.
Large Language Monkeys: Scaling Inference Compute with Repeated Sampling
by Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V. Le, Christopher Ré, Azalia Mirhoseini
First submitted to arxiv on: 31 Jul 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper explores the concept of inference compute in language models. Traditionally, models are limited to making one attempt at solving a problem during inference. The authors propose repeatedly sampling candidate solutions from a model to increase coverage, which is defined as the fraction of problems that can be solved by any generated sample. Across multiple tasks and models, the authors observe that coverage scales with the number of samples over four orders of magnitude, following a log-linear relationship. This paper highlights the importance of inference compute in scaling language models’ capabilities. The authors demonstrate improved performance in domains like coding and formal proofs where answers can be automatically verified. They also investigate common methods for picking from a sample collection, such as majority voting and reward models, which plateau beyond several hundred samples. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about improving how language models work during the process of solving problems. Normally, these models only try once to solve a problem. The authors suggest trying multiple times to find solutions. They found that when they do this, it helps solve more problems across different tasks and models. This can be useful in areas where answers can be easily checked, like coding and math proofs. The paper also looks at how people usually choose the best solution from a set of options and finds that these methods stop working well after trying many times. |
Keywords
* Artificial intelligence * Inference
