Diverse reasoning traces teach LLMs to make better decisions
Source
amazon.science
Date
Why it matters
Self-consistency only helps when sampled reasoning paths actually differ. Training with global forking tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → gives a single model explicit handles for distinct reasoning modes, so diversity comes from the model rather than temperature.
Terms in this piece · Glossary
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
