MIRD: Multimodal Implicit Reasoning Distillation for Efficient Sarcasm Detection
Abstract
Multimodal Sarcasm Detection (MSD) identifies sarcastic intents by jointly comprehending visual and textual modalities, providing deeper contextual understanding than conventional text-only detectors. However, it remains challenging as it relies on capturing subtle, fine-grained pragmatic incongruities across heterogeneous modalities. Existing methods mainly adopt traditional feature fusion networks or Vision Large Language Models (VLLMs). Nevertheless, they encounter two key problems: (1) conventional representation learning focuses on explicit feature alignment but lacks the deductive logic necessary for deep pragmatic prediction, and (2) mainstream LLM-based generative approaches suffer from prohibitive computational latency due to autoregressive generation, limiting their viability under real-time constraints. To address this, we propose MIRD, a novel Multimodal Implicit Reasoning Distillation framework that balances profound cognitive reasoning with high efficiency requirements. Our method comprises three core designs. First, a specialized cross-modal alignment network is developed to realize early-stage, fine-grained feature interleaving, providing a robust perception foundation. Second, we introduce a Multi-Perspective Implicit Preference Distillation (MIPD) strategy, which utilizes structurally defective negative trajectories to explicitly internalize the powerful deductive logic of teacher LLMs into a compact implicit latent state. Third, through Implicit Reasoning-Driven Adaptation (IRDA), MIRD directly maps the converged implicit reasoning representations to task-specific classification boundaries, thereby completely bypassing the computational bottleneck of autoregressive decoding. Extensive experiments on three benchmark datasets (MMSD, MMSD2.0, and RedEval) demonstrate that MIRD achieves state-of-the-art accuracy while consistently delivering significant improvements in efficiency over native autoregressive baselines. Code is available at https://github.com/possion0/MIRD.