ABARUAH at SemEval-2026 Task 1: Leveraging High-Resolution VLMs and Reasoning LLMs for Multimodal Humor Generation

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Emerging Technologies & Innovation · Depth: Expert, quick

Summary

The ABARUAH system, developed for SemEval-2026 Task 1: Humor Generation, addresses both unimodal text and multimodal GIF-based humor creation challenges. Its robust two-stage pipeline first employs a Multimodal Grounding stage, utilizing the Qwen2-VL model to extract detailed semantic descriptions from GIFs. Subsequently, a Humor Synthesis stage generates the final humorous output, leveraging the Qwen3-8B model for reasoning and generation. This integrated approach demonstrated competitive performance in the shared task, achieving Elo-like ratings of 1009 for Subtask A, 973 for Subtask B1, and 914 for Subtask B2. The system successfully handled diverse humorous constraints and secured a notable 4th place ranking in overall standings for both Subtasks A and B1, showcasing its effectiveness in complex humor generation tasks.

Key takeaway

For Machine Learning Engineers developing multimodal content generation systems, consider adopting a two-stage pipeline approach. This strategy, exemplified by ABARUAH's use of Qwen2-VL for visual grounding and Qwen3-8B for humor synthesis, effectively addresses complex tasks like GIF-based humor. Your team could explore similar VLM-LLM integrations to enhance performance and manage diverse constraints in creative AI applications.

Key insights

The ABARUAH system uses a two-stage VLM/LLM pipeline for multimodal humor generation, achieving competitive SemEval-2026 results.

Principles

• Two-stage pipelines can decompose complex multimodal tasks.
• VLMs excel at grounding visual content for LLMs.
• LLMs can synthesize creative text from structured inputs.

Method

A two-stage pipeline: Multimodal Grounding (Qwen2-VL extracts GIF semantics) followed by Humor Synthesis (Qwen3-8B generates humorous text from extracted descriptions).

In practice

• Combine VLMs and LLMs for multimodal content creation.
• Use Qwen2-VL for visual description tasks.
• Employ Qwen3-8B for text generation from structured inputs.

Topics

• Humor Generation
• Multimodal AI
• Vision-Language Models
• Large Language Models
• SemEval-2026 Task 1
• Qwen Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Grandes Modelos de Linguagem Multimodais (MLLMs): Da Teoria \`a Pr\'atica — MLLMs merge LLM language capabilities with multimodal perception, advancing AI's understanding of diverse data.
• I built a public voting benchmark where models have to make memes out of daily news — Memebench evaluates LLM humor and creativity by generating memes from daily news for public A/B voting.
• Claude Code Let's Build: The AI Video Oracle (Qwen3 TTS) — A multi-modal AI pipeline can generate short videos from text queries using Gemini, Qwen3 TTS, and Omnihuman.
• Learning to Think Like a Cartoon Captionist: Incongruity-Resolution Supervision for Multimodal Humor Understanding — Explicitly supervising reasoning structure, not just scale, improves multimodal humor understanding.
• ViQ: Text-Aligned Visual Quantized Representations at Any Resolution — ViQ balances semantic richness and detail preservation in discrete visual representations for efficient, unified multimodal modeling.
• Large Language Models for Mobile GUI Text Input Generation: An Empirical Study — LLMs effectively generate mobile GUI text inputs for testing, with performance varying by UI context and improving with feedback.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.