Skip to content

This repo accompanies a study of uncertainty benchmarking on state of the art llm decoding methods and novel fast inference techniques. We start with multi token prediction, speculative decoding as well as CoT decoding and evaluate llm uncertainty with these added methods compared to the baseline model using semantic entropy as our metric..

Notifications You must be signed in to change notification settings

userdarius/llm-reasoning-uncertainty

Repository files navigation

Files to watch at the moment are compute_uncertainty_measures and speculative decoding related code.

About

This repo accompanies a study of uncertainty benchmarking on state of the art llm decoding methods and novel fast inference techniques. We start with multi token prediction, speculative decoding as well as CoT decoding and evaluate llm uncertainty with these added methods compared to the baseline model using semantic entropy as our metric..

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages