Not All LLM Reasoning is Visible in the Chain-of-Thought

2026-07-24

AI Summary

This study investigates whether large language models fully disclose their reasoning processes in outputs—a critical issue for AI interpretability and safety. By introducing semantically irrelevant filler tokens into synthetic reasoning tasks, the authors present the first empirical evidence that state-of-the-art models can exploit such meaningless tokens to perform hidden reasoning, achieving objectives imperceptible to humans from the visible chain of thought. Through controlled experiments manipulating filler tokens and comparing reinforcement learning with supervised fine-tuning across 13 prominent models, the analysis reveals that several models improve accuracy by up to 13 percentage points using fillers. Notably, Claude Opus 4.5 implicitly satisfies modular arithmetic constraints without degrading primary task performance, demonstrating covert computational capabilities absent any explicit trace.

Abstract

A key question for AI safety is whether a language model expresses all of its reasoning in its output tokens. We demonstrate a concrete failure mode where frontier models exhibit invisible reasoning by leveraging semantically irrelevant filler tokens to improve performance on synthetic reasoning tasks. We evaluate 13 frontier language models across three tasks and find that many models benefit significantly from filler tokens, with accuracy improvements of up to 13 percentage points. The benefit depends on which tokens are used and differs across models. We further show that filler tokens enable Claude Opus 4.5 to satisfy a hidden modular arithmetic constraint without sacrificing accuracy on its primary task, demonstrating that invisible reasoning can serve objectives entirely invisible to CoT monitoring. Reinforcement learning gives Qwen3-235B strong preferences over filler token content, but neither RL nor supervised fine-tuning produces a filler token benefit that persists at test time. Our results indicate that frontier models already perform consequential computation with no interpretable trace in their output tokens.

Problem

• invisible reasoning
• chain-of-thought
• filler tokens
• AI safety
• language models

Innovation

• invisible reasoning
• filler tokens
• chain-of-thought
• AI safety
• language model interpretability