Less is more: UC Berkeley and Google unlock LLM potential through simple sampling
1 min read
Summary
A new paper by Google Research and the University of California, Berkeley demonstrates that by simply scaling up sampling-based search, AI’s language models can boost reasoning skills.
The technique works by producing numerous responses and using the model to verify the answers.
Even with a minimalist implementation, it was found to surpass the performance of deep learning models that specialised in reasoning when tested on popular benchmarks.
This stems from the common misconception that complicated architectures and highly specialised training are required to achieve the best results.
However, the costs associated with this type of sampling-based search can be prohibitive, with some questions costing up to $650 to answer.
Yet, the researchers suggest using simpler verification methods and smaller models to get these costs down.