An exploratory study of hallucination in Qwen 3.8-27B and GPT-OSS-20B (github.com)

A recent exploratory study examined the hallucination and abstention phenomena in two open-weight large language models (LLMs), Qwen 3.8-27B from Alibaba and GPT-OSS-20B from OpenAI. The study identified two distinct failure modes: Qwen exhibited "rigid confabulation," where it consistently fabricated information (citing the same false year, 1991, for different universities), while GPT-OSS experienced "deliberation collapse," resulting in a significant number of empty responses. This research highlights critical challenges in LLMs' reliability and consistency, raising concerns over the potential dissemination of misinformation. The implications of this study are significant for the AI/ML community as they underscore the need for better understanding and mitigation of hallucination in LLMs, which can impact user trust and application in critical domains. Findings revealed that authority pressure led to a notable increase in reasoning tokens within GPT-OSS, indicating model behavior can be influenced by prompts. However, emotional prompts did not yield increased fabrication, suggesting limitations in empathy-based prompts. The study, while not peer-reviewed and limited by its small sample size, provides valuable insights and sets the groundwork for further research into LLM reliability and user interaction dynamics.