Active Confusion Expression in Large Language Models: Leveraging World Models toward Better Social Reasoning
TLDR
LLMs struggle with social reasoning; proposed adaptive world model tracks states and intervenes, improving accuracy and reducing tokens.
Reasoning
Strengths: addresses a gap in social reasoning, proposes a novel world model mechanism, shows significant improvements on benchmarks. Weaknesses: limited to textual world model, no discussion of real-world deployment or broader applicability.
Read-first score
Read-first score 43.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 12.
Field roles
Rank sensitivity
Stability: volatile; rank range: 409.
Keyword Scores
Deep Analysis
Innovations
- Adaptive world model-enhanced reasoning mechanism that constructs a dynamic textual world model to track entity states and temporal sequences
- Dynamic monitoring of reasoning trajectories for confusion indicators and prompt intervention by providing clear world state descriptions
- Mimics human implicit world models to distinguish between external events and internal beliefs
Methodology
The paper analyzes DeepSeek-R1's reasoning trajectories to identify confusion patterns, then proposes a mechanism that builds a dynamic textual world model tracking entity states and temporal sequences. It monitors reasoning for confusion indicators and intervenes by providing clear world state descriptions. The approach is evaluated on three social benchmarks using accuracy and token reduction metrics.
Key Results
The mechanism achieves significant accuracy improvements (e.g., +10% on Hi-ToM) and reduces computational costs by up to 33.8% token reduction.