Research map / Video / Video-language understanding
SPIKE-RL: Video-LLMs meet Bayesian Surprise
Sahithya Ravi, Aditya Chinchure, Raymond T. Ng, Leonid Sigal, Vered Shwartz
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 |
|---|---|---|---|
| add GRPO belief optimization (SPIKE-RL) (ablation) | SPIKE · Oops! surprise localization | Acc@0.25s: 60.0 -> 62.9 | Improved |
| add GRPO belief optimization (SPIKE-RL) (ablation) | SPIKE · FunQA surprise localization | IoU: 65.7 -> 68.2 | Improved |
| add GRPO belief optimization (SPIKE-RL) (ablation) | SPIKE · Mr. Bean surprise localization | IoU: 54.8 -> 61.1 | Improved |
| add GRPO belief optimization (SPIKE-RL) (ablation) | SPIKE · belief diversity | 1-cos similarity: 33.5% -> 40.3% | Improved |
| add GRPO belief optimization (SPIKE-RL) (ablation) | SPIKE · Oops! human correlation | Spearman correlation: 0.84 -> 0.87 | Improved |
| replace uniform sampling with SPIKE surprise-weighted sampling | Qwen2.5-VL-7B · BlackSwan | accuracy: 67.2 -> 68.8 | Improved |
| replace uniform sampling with SPIKE-RL sampling | Qwen2.5-VL-7B · ExFunTube | LLM-Match: 68.7 -> 75.7 | Improved |
| replace uniform sampling with SPIKE-RL sampling | Qwen2.5-VL-7B · VideoMME-S | accuracy: 59.8 -> 62.5 | Improved |
| replace uniform sampling with SPIKE-RL sampling | Qwen2.5-VL-7B · NextQA | accuracy: 68.6 -> 70.3 | Improved |
| replace uniform sampling with SPIKE-RL sampling | Qwen2.5-VL-32B · VideoMME-S | accuracy: 69.9 -> 73.5 | Improved |
| replace uniform sampling with SPIKE-RL sampling | Qwen2.5-VL-32B · FunQA (Task 2) | LLM-Match: 72.7 -> 75.8 | Improved |
| use RGB histogram shot-boundary sampling instead of uniform | Qwen2.5-VL-7B · BlackSwan | accuracy: 67.2 -> 49.6 | Hurt |
Closest papers by meaning
- SURGE: Surprise-Guided Token Reduction for Efficient Video Understanding with VLMs (ICLR 2026 · similarity 0.87)
- Black Swan: Abductive and Defeasible Video Reasoning in Unpredictable Events (CVPR 2025 · similarity 0.86)
- OVO-Bench: How Far is Your Video-LLMs from Real-World Online Video Understanding? (CVPR 2025 · similarity 0.85)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.85)
- VES-RFT: Rewarding Visual Evidence Sensitivity to Mitigate Hallucinations in Large Vision-Language Models (CVPR 2026 · similarity 0.84)
- DELTAVID: Enhancing Fine-Grained Spatiotemporal Perception with Cross-Video Differences (NeurIPS 2026 · similarity 0.84)
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.
