Decomposing Reasoning Efficiency
The workshop version. Tokens per solved puzzle on CogniLoad as a first efficiency metric next to accuracy.
Abstract
Building on the CogniLoad benchmark, this work introduces an efficiency metric for reasoning LLMs, tokens generated per solved puzzle, to evaluate computational cost alongside accuracy across controlled difficulty conditions.
Cite
@inproceedings{kaiser2025decomposing,
title = {Decomposing Reasoning Efficiency},
author = {Kaiser, Daniel and Frigessi, Arnoldo and Ramezani-Kebrya, Ali and Ricaud, Benjamin},
booktitle = {NeurIPS 2025 Workshop on Efficient Reasoning},
year = {2025},
url = {https://openreview.net/forum?id=mym1D6mIZR}
}