LLM Reasoning Is Latent, Not the Chain of Thought, New Research Suggests

A new position paper published on arXiv argues that large language model (LLM) reasoning should be understood as latent-state trajectory formation rather than as a faithful surface chain of thought (CoT). The paper, titled "LLM Reasoning Is Latent, Not the Chain of Thought," suggests that this distinction is critical for claims about faithfulness, interpretability, reasoning benchmarks, and inference-time intervention.

The authors contend that the field has often confounded three key factors in LLM reasoning, leading to a misalignment between surface-level outputs and the underlying processes. By formalizing three competing hypotheses, they propose that reasoning is primarily mediated by latent states, which are not directly observable but shape the model's outputs. This perspective could significantly impact how researchers design and evaluate reasoning tasks in LLMs.

The implications of this research are far-reaching. If reasoning is indeed a latent-state trajectory, it suggests that current methods for interpreting and intervening in LLM reasoning may be incomplete. Future work could focus on developing new techniques to uncover and manipulate these latent states, potentially leading to more robust and interpretable AI systems. The paper also raises questions about the validity of existing reasoning benchmarks, which may need to be revisited in light of this new framework.