Streaming Hallucination Detection in Long Chain-of-Thought Reasoning

TL;DR

This paper introduces a method for real-time detection of hallucinations in long chain-of-thought reasoning by modeling hallucinations as an evolving latent state, enabling interpretable, streaming detection.

Contribution

It proposes a novel approach that treats hallucinations as a dynamic latent state and introduces a cumulative signal for real-time, interpretable detection in long reasoning chains.

Findings

Effective streaming hallucination detection demonstrated

Provides real-time, interpretable evidence of hallucinations

Models hallucinations as an evolving latent state

Abstract

Long chain-of-thought (CoT) reasoning improves the performance of large language models, yet hallucinations in such settings often emerge subtly and propagate across reasoning steps. We suggest that hallucination in long CoT reasoning is better understood as an evolving latent state rather than a one-off erroneous event. Accordingly, we treat step-level hallucination judgments as local observations and introduce a cumulative prefix-level hallucination signal that tracks the global evolution of the reasoning state over the entire trajectory. Overall, our approach enables streaming hallucination detection in long CoT reasoning, providing real-time, interpretable evidence.