WM-SAR: World Model Sarcasm Reasoning

• The paper introduces WM-SAR, a modular framework that decomposes sarcasm into literal meaning, context, normative expectation, and intention.
• It leverages four specialized LLM agents whose outputs are integrated by a logistic regression classifier, achieving competitive accuracy on benchmark datasets.
• The framework provides interpretable decision rationales, enhancing transparency compared to traditional black-box deep-learning methods.

World Model inspired Sarcasm Reasoning (WM-SAR) is a computational framework for sarcasm detection that formulates the problem as a structured reasoning process over multiple dimensions of discourse. This approach explicitly decomposes the judgment of sarcasm into separate components—literal meaning, context, normative expectation, and speaker intention—each managed by specialized LLM-based agents. Deviations between these components are numerically quantified and synthesized by a lightweight logistic regression classifier, yielding both interpretable decision boundaries and competitive empirical results on benchmark sarcasm datasets (Inoshita et al., 30 Dec 2025).

1. Conceptual Foundations of WM-SAR

WM-SAR reformulates sarcasm understanding in natural language processing as a "world model" inference task. Sarcasm is viewed as arising from an interaction between linguistic surface meaning and underlying pragmatic or social cues, specifically the mismatch between a statement's semantic polarity and the expected sentiment based on social norms, together with the speaker's intent. Traditional deep-learning approaches often rely on black-box predictions, masking the cognitive factors underlying sarcasm and limiting interpretability. WM-SAR addresses this by architecting explicit modules that mirror key aspects of human sarcasm processing—semantic evaluation, context reconstruction, norm-based expectation, and Theory-of-Mind-based intention estimation—thus enabling traceable detection mechanisms and rationales for predictions.

2. Architecture and Agent Decomposition

WM-SAR consists of four LLM-based agents and a deterministic inconsistency detector, with all outputs integrated by a logistic-regression arbiter. The modular agents operate in parallel, providing low-dimensional signals and concise rationales:

• Literal Meaning Evaluator (A_lit): Computes surface polarity $M_{lit}(u) \in [-1,1]$ for utterance $u$, accompanied by a textual rationale.
• Context Model (A_ctx): Reconstructs a minimal background situation $C(u) = \langle$ social relation, scene, preceding event $\rangle$, outputting a hypothesized context and justification.
• Normative Expectation Evaluator (A_norm): Predicts normative sentiment $E_{norm}(C(u)) \in [-1,1]$ for context $C(u)$, with rationale.
• Intention Reasoner (A_int): Performs Theory-of-Mind inference to estimate intention $T_{sar}(u, C(u)) \in [0,1]$, returning scalar and ToM justification.
• Inconsistency Detector: Computes semantic discrepancy $D(u, C(u)) = M_{lit}(u) - E_{norm}(C(u))$, with polarity-flip indicator $$SD(u, C(u)) = \mathbb{I}[\operatorname{sgn}(M_{lit}(u)) \neq \operatorname{sgn}(E_{norm}(C(u)))]$$.

All features—absolute discrepancy $|D|$, flip indicator $SD$, and intention $T$—are standardized and input directly into a logistic regression classifier, avoiding cascading prompts. This design facilitates interpretable, low-dimensional decision statistics.

3. Mathematical Formulation

Sarcasm probability is modeled as:

$$P(\text{sarcasm}=1 \mid u) = \sigma(w_1\,|D(u,C(u))| + w_2\,T(u,C(u)) + w_3\,SD(u,C(u)) + b)$$

where $\sigma(x) = 1/(1+e^{-x})$ is the sigmoid function, and $(w_1, w_2, w_3, b)$ are weights obtained via regularized logistic regression. Interaction features such as $D+T$, $D \times T$, $\sqrt{D}$, and $\sigma(T)$ may be included, but the main determinants are $|D|$ (inconsistency magnitude) and $T$ (intentionality confidence). Hyperparameters are selected by stratified 5-fold cross-validation on train-validation data, using accuracy as the primary metric and macro-F1 as the tie-breaking criterion.

4. Experimental Setup and Baselines

WM-SAR is evaluated on three sarcasm detection benchmarks:

• IAC-V1: Political forum posts.
• IAC-V2: Expanded, more diverse set from IAC-V1.
• SemEval-2018 Task 3: English tweets.

Data splits are 80% training, 10% validation, 10% testing. All LLM agents leverage the GPT-4.1-mini backbone, with task-specific prompts enforcing structured JSON outputs containing required scalars and rationales. No in-domain fine-tuning occurs; reasoning operates under zero- or few-shot conditions. Compared baselines include deep learning models (MIARN, SAWS, DC-Net), fine-tuned BERT, various zero-shot and prompt-engineered GPT-4.1-mini variants, and CAF-I multi-agent architectures. Metrics are Accuracy and macro-F1.

5. Empirical Results

WM-SAR achieves state-of-the-art or near-state-of-the-art performance on all tested datasets:

Dataset	WM-SAR Accuracy	WM-SAR F1	Best Baseline Accuracy	Best Baseline F1
IAC-V1	0.745	0.745	0.725 (GPT-4.1-mini)	0.724
IAC-V2	0.791	0.791	0.791 (BERT)	0.790
SemEval	0.714	0.714	0.707 (BERT)	0.694

On average, WM-SAR outperforms the best non-WM-SAR model by 2–3% in Accuracy/F1 compared to zero-shot GPT variants and by ~4–5% over fine-tuned deep learning models.

Ablation studies demonstrate the criticality of the intention reasoning signal (T): removing $T_{sar}$ drops average Accuracy/F1 from 0.750/0.750 to ≈0.672/0.665 (–8% absolute). Semantic inconsistency $|D|$ and polarity-flip $SD$ further strengthen predictions, with their exclusion reducing performance by 2–4%. Omitting feature interactions also reduces accuracy, confirming the value of non-linear combinations of signals.

6. Interpretability and Analysis

WM-SAR provides explicit and interpretable decision rationales for each prediction. For example, in a non-sarcastic utterance (“…my apologies… has no idea… compelled to speak up…”), the system computes $M_{lit}=+0.10$, $E_{norm}=+0.60$, yielding $D=-0.50$, $SD=0$, and $T_{sar}=0.10$, resulting in a correct non-sarcasm classification with transparent justification. Conversely, for "positive sarcasm" cases with strong alignment between literal and normative signals and no intention cues, the model fails to detect sarcasm (e.g., “A truly charming publication… warm human goodness.” with $M_{lit}=+0.80$, $E_{norm}=+0.80$, $D=0$, $SD=0$, $T_{sar}=0.10$), highlighting systematic limitations.

7. Significance, Limitations, and Extensions

Decomposition of sarcasm reasoning into modular evaluators enables direct measurement of the "polarity reversal" central to sarcastic language. The Theory-of-Mind intention score filters out non-sarcastic inversions (e.g., exaggeration, joke), while logistic regression provides a compact and interpretable aggregation mechanism.

Key limitations include failure on "positive sarcasm" lacking reversed or ambiguous signals, restriction to single-turn text data, and reliance on the GPT-4.1-mini agent family. Prospective extensions involve integrating dialogue history or user profiles into context $C(u)$, expanding to multimodal inputs (images, audio), testing with alternative LLM backbones, and refining prompt strategies to capture subtler forms of irony beyond literal/norm inversion (Inoshita et al., 30 Dec 2025).

A plausible implication is that the modular world-model approach of WM-SAR may generalize to broader pragmatic phenomena beyond sarcasm, offering a template for structured agent decomposition and interpretable decision making in other socially nuanced NLP tasks.