Hugging Face shared a new case study last week showcasing how small language models (SLMs) can outperform larger models. In the post, the platform’s researchers claimed that instead of increasing the training time of artificial intelligence (AI) models, focusing on the test-time compute can show enhanced results for AI models. The latter is an inference strategy that allows AI models to spend more time on solving a problem and offers different approaches such as self-refinement and searching against a verifier that can improve their efficiency.
How Test-Time Compute Scaling Works
In a post, Hugging Face highlighted that the traditional approach to improving the capabilities of an AI model can often be resource-intensive and extremely expensive. Typically, a technique dubbed train-time compute is used where the pretraining data and algorithms are used to improve the way a foundation model breaks down a query and gets to the solution.
Alternatively, the researchers claimed that focusing on test-time compute scaling, a technique where AI models are allowed to spend more time solving a problem and letting them correct themselves can show similar results.
Highlighting the example of OpenAI’s o1 reasoning-focused model, which uses test-time compute, the researchers stated that this technique can let AI models display enhanced capabilities despite making no changes to the training data or pretraining methods. However, there was one problem. Since most reasoning models are closed, there is no way to know the strategies that are being used.
The researchers used a study by Google DeepMind and reverse engineering techniques to unravel how exactly LLM developers can scale test-time compute in the post-training phase. As per the case study, just increasing the processing time does not show significant improvement in outputs for complex queries.
Instead, the researchers recommend using a self-refinement algorithm that allows AI models to assess the responses in subsequent iterations and identify and correct potential errors. Additionally, using a verifier that models can search against can further improve the responses. Such verifiers can be a learned reward model or hard-coded heuristics.
More advanced techniques would involve a best-of-N approach where a model generates multiple responses per problem and assigns a score to judge which would be better suited. Such approaches can be paired with a reward model. Beam search, which prioritises step-by-step reasoning and assigning scores for each step, is another strategy highlighted by researchers.
By using the abovementioned strategies, the Hugging Face researchers were able to use the Llama 3B SLM and make it outperform Llama 70B, a much larger model, on the MATH-500 benchmark.