AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

S³-MSD: Large Vision-Language Model for Explainable and Generalizable Multi-modal Sarcasm Detection

Zhihong Zhu, Fan Zhang, Yunyan Zhang, Jinghan Sun, Guimin Hu, Hao Wu, Yuyan Chen, Bowen Xing, Xian Wu

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Instruction-tuned multimodal assistants · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces S³-MSD, a large vision-language model for explainable and generalizable multimodal sarcasm detection on image-text pairs.

Problem

Multimodal sarcasm detection (MSD) on image-text pairs suffers from lack of explainability and weak generalizability across datasets.

Approach

Proposes an LVLM-based approach (S³-MSD) aimed at producing explainable sarcasm predictions while generalizing better than prior MSD methods; specific mechanism not stated in the available abstract.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add self-training (iterative bootstrapping of explanations from correct answers) (ablation)S3-MSD (LLaVA-1.5-7B) · MMSD2.0accuracy (%): 86.65 -> 88.69Improved
add self-calibrating (condition on incorrect explanation to produce correct one) (ablation)S3-MSD (LLaVA-1.5-7B) · MMSD2.0accuracy (%): 87.64 -> 88.69Improved
add self-focusing (image-only DPO with noised entity regions) (ablation)S3-MSD (LLaVA-1.5-7B) · MMSD2.0accuracy (%): 87.31 -> 88.69Improved
add self-training (ablation)S3-MSD (LLaVA-1.5-7B) · RedEval (OOD)accuracy (%): 84.12 -> 85.98Improved
add self-focusing (ablation)S3-MSD (LLaVA-1.5-7B) · RedEval (OOD)accuracy (%): 84.64 -> 85.98Improved
add self-calibrating (ablation)S3-MSD (LLaVA-1.5-7B) · RedEval (OOD)accuracy (%): 85.14 -> 85.98Improved
replace self-calibrating with textual DPO (ablation)S3-MSD · MMSD2.0 and RedEvalaccuracy, P, R, F1Hurt
train with explanations via S3-MSD instead of golden labels onlyLLaVA-1.5-7B · MMSD2.0accuracy (%): 85.18 -> 88.69Improved
use S3-MSD instead of V-STaR (k = 5)LLaVA-1.5-7B · RedEval (OOD)accuracy (%): 84.73 -> 85.98Improved
apply S3-MSD to Qwen2-VL-7B and Qwen2.5-VL-7B backbonesS3-MSD · MMSD2.0* subtest (500 instances)accuracyImproved
use S3-MSD instead of untuned closed-source LVLMs (GPT-4o, o3, Gemini)MMSD2.0* subtest · MMSD2.0* subtestaccuracyImproved
fine-tune S3-MSD on MORE for explanation generationS3-MSD · MOREBLEU-4: 34.25 vs TURBO 35.26Mixed

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy