Research map / Video / Video-language understanding
Learning to Deny: Action Denial in Multimodal Large Language Models
Raiyaan Abdullah, Shehreen Azad, Yogesh Rawat
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| preserve context but remove defining motion (action-denial negatives) vs action-presence clips | 20 MLLMs · UCF101-AD | accuracy (%): Action-Presence 97.5 -> Overall-AD 34.1 (Ovis2.5-9B) | Hurt |
| replace generic None option with explicit denial statement (ablation) | Ovis2.5-9B · UCF101-AD Overall-AD | accuracy (%): 34.1 -> 74.7 | Improved |
| remove the primary distractor from the options (ablation) | Ovis2.5-9B · UCF101-AD Overall-AD | accuracy (%): 34.1 -> 93.9 | Improved |
| reduce answer options from 11 to 4 (ablation) | 8 high-performing MLLMs · UCF101-AD Overall-AD | accuracy (%): VideoLLaMA3-7B 51.5 -> 43.7; Qwen2.5-VL-7B 25.1 -> 30.5 | Mixed |
| use reasoning (thinking) variants | MLLMs · UCF101-AD Overall-AD | accuracy (%): Ovis2.5-9B 34.1 -> 40.4; Kimi-VL 29.6 -> 20.9 | Mixed |
| increase model scale (ablation) | Qwen2.5-VL family · UCF101-AD Overall-AD | accuracy (%): 19.8 (3B) -> 25.1 (7B) -> 45.7 (72B) | Improved |
| add CausalAct-0 causal graph prompt | Ovis2.5-9B · UCF101, K400, HMDB51, SSv2, Diving48, FineGym99 denial | denial accuracy (%): HMDB51 51.2 -> 74.3; SSv2 16.0 -> 37.4; UCF101 78.7 -> 86.1 | Improved |
| add CausalAct zero-shot prompt | VideoLLaMA3 · UCF101-AD | Overall-AD | Hurt |
| finetune small MLLMs on graph-based VQA tasks (CausalAct) | 2B-3B MLLMs · UCF101-AD Action-Denial test | accuracy | Improved |
| use pruned or random graph instead of full causal graph (ablation) | CausalAct finetuned · UCF101-AD | accuracy | Hurt |
| use pruned or random graph in zero-shot (ablation) | CausalAct-0 · UCF101-AD | accuracy | No clear effect |
| freeze vision encoder during graph finetuning (ablation) | CausalAct finetuning · UCF101-AD | accuracy: about -15% on average | Hurt |
Closest papers by meaning
- Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs (NeurIPS 2026 · similarity 0.88)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.87)
- Look, Remember and Reason: Grounded Reasoning in Videos with Language Models (ICLR 2024 · similarity 0.87)
- Can't Make an Omelette Without Breaking Some Eggs: Plausible Action Anticipation Using Large Video-Language Models (CVPR 2024 · similarity 0.87)
- [Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning (NeurIPS 2026 · similarity 0.87)
- Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.87)
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.
