Rethinking Thinking Tokens: Understanding Why They Underperform in Practice

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Vennam, Sreeram, Valente, David, Herel, David, Kumaraguru, Ponnurangam
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866929595280785408
author Vennam, Sreeram
Valente, David
Herel, David
Kumaraguru, Ponnurangam
author_facet Vennam, Sreeram
Valente, David
Herel, David
Kumaraguru, Ponnurangam
contents Thinking Tokens (TT) have been proposed as an unsupervised method to facilitate reasoning in language models. However, despite their conceptual appeal, our findings show that TTs marginally improves performance and consistently underperforms compared to Chain-of-Thought (CoT) reasoning across multiple benchmarks. We hypothesize that this underperformance stems from the reliance on a single embedding for TTs, which results in inconsistent learning signals and introduces noisy gradients. This paper provides a comprehensive empirical analysis to validate this hypothesis and discusses the implications for future research on unsupervised reasoning in LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2411_11371
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Rethinking Thinking Tokens: Understanding Why They Underperform in Practice
Vennam, Sreeram
Valente, David
Herel, David
Kumaraguru, Ponnurangam
Computation and Language
Machine Learning
I.2.6
Thinking Tokens (TT) have been proposed as an unsupervised method to facilitate reasoning in language models. However, despite their conceptual appeal, our findings show that TTs marginally improves performance and consistently underperforms compared to Chain-of-Thought (CoT) reasoning across multiple benchmarks. We hypothesize that this underperformance stems from the reliance on a single embedding for TTs, which results in inconsistent learning signals and introduces noisy gradients. This paper provides a comprehensive empirical analysis to validate this hypothesis and discusses the implications for future research on unsupervised reasoning in LLMs.
title Rethinking Thinking Tokens: Understanding Why They Underperform in Practice
topic Computation and Language
Machine Learning
I.2.6
url https://arxiv.org/abs/2411.11371