Tag
This paper provides a theoretical explanation for why the optimal sampling temperature for pass@k increases with the budget, deriving sufficient conditions and analyzing the empirical pattern without model training or queries.
The paper introduces AUF (Accept-Until-Fail), a simple modification to the cross-entropy loss for masked block drafters in speculative decoding that restricts supervision to the prefix up to the first predicted failure, improving average emitted length across benchmarks without changing inference.