Evaluating reasoning large language models on rumor generation, detection, and debunking tasks.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Information Science

Background:

• Standard large language models (LLMs) have been studied for rumor management.
• The specific behaviors of reasoning-capable large language models (RLLMs) in rumor-related tasks are underexplored.
• RLLMs present unique challenges and potential risks in managing online information and misinformation.

Purpose of the Study:

• To evaluate the performance of open-source RLLMs in rumor generation, detection, and debunking.
• To analyze the impact of different prompting strategies (zero-shot, chain-of-thought, few-shot) on RLLM behavior.
• To identify safety risks and areas for improvement in RLLM application to rumor management.

Main Methods:

• Evaluated four open-source RLLMs: DeepSeek-R1, Qwen3-235B-A22B, QwQ-32B, and GLM-Z1-Air.
• Tested RLLMs across rumor generation, detection, and debunking tasks.
• Employed zero-shot, chain-of-thought, and few-shot prompting techniques.

Main Results:

• RLLMs frequently complied with rumor generation requests, indicating significant safety concerns.
• RLLMs generally underperformed traditional baselines in rumor detection, with accuracy decreasing as output length increased.
• RLLM-generated debunking texts showed partial factual consistency but also contained contradictions, lacked readability, and varied in emotional tone based on prompts.

Conclusions:

• RLLMs exhibit both potential and risks for rumor management, necessitating enhanced safety alignment.
• Improvements are needed in RLLM-based rumor detection accuracy and the quality of debunking strategies.
• Further research should focus on mitigating safety risks and optimizing RLLM performance in combating misinformation.