State over Tokens: Characterizing the Role of Reasoning Tokens

Quick Overview

The paper "State over Tokens" characterizes the role of reasoning tokens in Large Language Models (LLMs) by demonstrating that the sequence of these tokens, rather than just the final output, dictates the model's internal state and computational process, revealing that LLMs often use these reasoning steps to maintain internal consistency and avoid fundamental errors, even if the final answer appears plausible.

Key Points: LLMs often generate intermediate reasoning steps (reasoning tokens) that are functionally separate from the final answer's linguistic meaning. The sequence of these tokens dictates the model's internal state and computational process, suggesting they are crucial for maintaining coherence. The paper cites an example where the sequence 1, 1, 2, 5, 14, 42 leads to the final answer 14, but the actual required calculation is a sum of products (15 + 241), indicating the sequence is not a direct representation of the math. LLMs can exhibit a tendency to skip crucial information or use arbitrary semantics (like a 'billion dollars' token) if the reasoning sequence is not carefully managed or explicitly constrained. When forced to only output the final answer without showing intermediate steps, models often fail to achieve high accuracy, suggesting the reasoning process itself is necessary for the computation. The core finding is that the reasoning token sequence acts as an external hard drive or a computational blueprint that the model follows, which is why explicit reasoning steps are critical for complex tasks.

Context: This video discusses a research paper titled "State over Tokens: Characterizing the Role of Reasoning Tokens," which investigates how Large Language Models (LLMs) utilize the intermediate steps they generate during complex problem-solving. The core concept is the distinction between the linguistic output (the final answer) and the internal computational process represented by 'reasoning tokens,' which function as a working memory or a step-by-step guide for the model.