Rethinking Thinking Tokens: Understanding Why They Underperform in Practice
This addresses a practical problem for researchers and practitioners in AI by identifying limitations in an unsupervised reasoning method, though it is incremental as it builds on existing concepts.
The paper investigates why Thinking Tokens underperform compared to Chain-of-Thought reasoning in language models, finding that they only marginally improve performance across benchmarks due to issues with single embeddings causing inconsistent learning and noisy gradients.
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.