Reproduce “Active Confusion Expression in Large Language Models: Leveraging World Models toward Better Social Reasoning” 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
Jialu Du, Guiyang Hou, Yihui Fu, Chen Wu, Wenqi Zhang, Yongliang Shen, Weiming Lu
Researchers from Zhejiang University developed an adaptive world model-enhanced reasoning mechanism to improve large language models' social reasoning by intervening when cognitive confusion is detected. This method dynamically constructs and leverages structured world state information, leading to substantial accuracy improvements and reduced computational costs across social reasoning benchmarks.
Key insights
- LLM outputs containing "confusion words" like "tricky" or "ambiguous" are strong indicators of cognitive dilemmas, correlating with reasoning errors and increased token consumption.
- Dynamically constructing and injecting structured world state information directly into the LLM's reasoning process during detected confusion helps disambiguate objective reality from subjective beliefs.
- This adaptive intervention strategy achieves a dual optimization effect, simultaneously improving reasoning accuracy and significantly reducing computational costs by guiding the model away from inefficient thought loops.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.