Video Humor Reasoning (Indic-SMILE)
Dataset + benchmarking LLMs on culturally grounded Hindi humor understanding.
Developed Indic-SMILE, a dataset for generating plausible reasons behind humor in short video clips.
- Extracted audio, visual, and textual features for multimodal reasoning.
- Benchmarked open-source and proprietary LLMs (e.g., InternVL3, VideoLLaMA3, Gemini 2.5 Flash, Claude Sonnet, Gemma-3, Sarvam-M) on humor understanding in Hindi, testing whether models can grasp culturally grounded humor that requires deep knowledge of Indian culture and language nuances.
- LoRA fine-tuned open-source LLMs and VLMs for the study.
- Work accepted as a workshop paper at the C3NLP Workshop, ACL 2026.