Reproduce “A Unified Definition of Hallucination: It's The World Model, Stupid!” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Emmy Liu, Varun Gangal, Chelsea Zou, Michael Yu, Xiaoqi Huang, Alex Chang, Zhuofu Tao, Karan Singh and 2 more
Researchers from DegenAI Labs and academic institutions propose a unified definition of hallucination in AI as "inaccurate world modeling observable to the user," characterized by a reference world model, view function, and conflict resolution policy. This framework subsumes diverse task-specific definitions and provides a structured approach for benchmark design and targeted mitigation strategies.
Key insights
- Diverse existing definitions of hallucination across various AI applications can be systematically unified and understood as specific instantiations of the proposed W, V, P framework.
- The framework provides a clear conceptual distinction between hallucination (incorrect beliefs about the reference world) and other model failures like planning errors or reward errors, allowing for more targeted research.
- Explicitly defining W, V, and P in benchmark design enables greater comparability and reproducibility in evaluating hallucination across different tasks and methodologies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.