Hallucinations Leave Traces: Future Context-Based LLM Hallucination Detection

The Hallucination phenomenon in LLMs has the characteristic of causing a Snowball Effect, where incorrect information generated by the model spreads throughout the subsequent context. Existing uncertainty-based methods have the drawback of requiring internal model information (such as Logits), sampling-based methods are costly, and retrieval-based methods have limitations in information accessibility.

To address these constraints, Naver Cloud proposes a black-box-based detection method utilizing Future Context. The core idea is that when it is difficult to determine whether the current sentence is a hallucination, a separate Detector LLM first generates the Future Context that would follow that sentence, and uses this as an additional clue.

The proposed pipeline proceeds through the following steps:

• Select the sentence to be verified
• Sample future sentences via the Detector LLM
• Add the generated future context to the input of the existing detection method
• Make the final hallucination determination

This approach effectively catches hallucinations without needing the model's internal information or external retrieval, by jointly considering how consistent the current sentence is with the subsequent context.