Reproduce “Reasoning Large Language Model Errors Arise from Hallucinating Critical Problem Features” 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
Alex Heyman, Joel Zylberberg
Researchers investigated reasoning Large Language Model errors and found they frequently "hallucinate" critical problem features explicitly given in the prompt, leading to incorrect solutions. This input-conflicting hallucination phenomenon consistently appears across different models, problem complexities, and logic tasks like graph coloring and stable matching.
Key insights
- RLLM reasoning errors frequently arise from "input-conflicting hallucinations," where models misrepresent explicitly provided problem features, a distinct phenomenon from fact-conflicting hallucinations.
- These hallucinations often involve a non-uniform bias towards higher-indexed entities (e.g., vertices, students), suggesting weaknesses in representing numerical order or relationships.
- The persistence of these hallucinations across different RLLMs, problem complexities, and reasoning tasks indicates a fundamental limitation in how models process and maintain fidelity to initial input information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.