FLEET gets more distinct code samples for the same compute

TL;DR

• FLEET, a paper posted on 23 September, adds a memory to text generation so later samples avoid the paths earlier ones took.
• On LiveCodeBench with Llama 3.2-3B, Pass@32 went from 59.9% to 66.2% at the same budget; the authors also claim the baseline's accuracy at a 3x speedup.
• It needs white-box access to the model's hidden states, the README says.