Leveraging the collective intelligence of large language models (LLMs)-based multiagent collaboration has led to significant advancements in intelligent applications across multiple domains. However, due to the untruthful content generated by LLMs, these collaborations face the challenge of continuously amplifying hallucinations, causing hallucination snowballing effect. Currently, existing research only discussed this concern in the context of a single model without analyzing or addressing it in agent collaborations. To tackle these challenges, this article proposes a context-aware hallucination analysis framework that captures token-level dependencies, leveraging semantic reasoning to validate and mitigate the snowballing effect in sequential multiagent collaboration. Specifically, we first propose a contextually embedded probabilistic modeling enabled hallucination analysis framework that systematically identifies and analyzes how collaborative processes propagate hallucinations. In addition, we construct a token-level disruption sequence detection approach for different task sequences to recognize and validate this effect across different domains. Finally, to mitigate hallucination snowballing without modifying the model architecture, we design a semantic reasoning empowered mitigation strategy based on a more effective bidirectional entailment clustering which mitigates hallucination propagation caused by the model itself, alleviate it caused by external knowledge deficiencies. Our extensive experiments with real datasets validate the existence of this effect in multiagent collaborations across various domains and demonstrate that our proposed mitigation strategy effectively reduces the propagation of hallucinations.
Xu et al. (Thu,) studied this question.