World model inspired sarcasm reasoning with large language model agents

arXiv 25.12 2025 39.5 method

TLDR

Reformulates sarcasm understanding as world model inspired reasoning with LLM agents, decomposing literal meaning, context, expectation, intention.

Reasoning

The paper introduces a novel decomposition of sarcasm reasoning into interpretable components using LLM agents, which is a strength for explainability. However, the 'world model' terminology is used loosely as inspiration rather than a true predictive or interactive world model, and the approach is narrowly focused on sarcasm detection without broader world dynamics.

Read-first score

Read-first score 39.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 11.

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=ablation,benchmark,evaluation,experiment

Recency 6%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.48975472

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Topical relevance 29%

15.7

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 217.

Keyword Scores

world model

world dynamics prediction

world simulator

generative world model

interactive world model

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Reformulates sarcasm understanding as a world model inspired reasoning process
Decomposes sarcasm into literal meaning, context, normative expectation, and intention using specialized LLM-based agents
Explicitly quantifies discrepancy between literal evaluation and normative expectation as a deterministic inconsistency score
Integrates inconsistency and intention scores via a lightweight Logistic Regression model for interpretable final prediction

Methodology

WM-SAR decomposes sarcasm reasoning into four LLM-based agents that model literal meaning, context, normative expectation, and intention. It computes a deterministic inconsistency score from the discrepancy between literal evaluation and normative expectation, and an intention score, then combines these signals using a lightweight Logistic Regression model to output a sarcasm probability.

Key Results

WM-SAR consistently outperforms existing deep learning and LLM-based methods on representative sarcasm detection benchmarks. Ablation studies confirm that integrating semantic inconsistency and intention reasoning is essential for effective sarcasm detection.

Limitations

The decomposition into normative expectations and intentions may require careful prompt engineering and may not generalize across all sarcasm types or domains
The lightweight Logistic Regression model may oversimplify complex interactions between the decomposed components

Resources

Paper