Research map / Video / Video-language understanding
EgoExo-Con: Exploring View-Invariant Video Temporal Understanding
Minjoon Jung, Junbin Xiao, Junghyun Kim, Byoung-Tak Zhang, Angela Yao
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 semantic reasoning reward r_sem to GRPO (ablation) | View-GRPO (Qwen2.5-VL-7B) · EgoExo-Con grounding | G-ExoEgo: 14.3 -> 18.3 | Improved |
| add semantic reasoning reward r_sem to GRPO (ablation) | View-GRPO (Qwen2.5-VL-7B) · EgoExo-Con verification | V-ExoEgo: 39.8 -> 44.7 | Improved |
| add structural consistency reward r_struct on top of r_sem (ablation) | View-GRPO (Qwen2.5-VL-7B) · EgoExo-Con grounding | G-ExoEgo: 18.3 -> 18.7 | Improved |
| add structural consistency reward r_struct on top of r_sem (ablation) | View-GRPO (Qwen2.5-VL-7B) · EgoExo-Con verification | V-ExoEgo: 44.7 -> 45.1 | Improved |
| use View-GRPO instead of plain GRPO | Qwen2.5-VL-7B · EgoExo-Con grounding | G-ExoEgo: 14.3 -> 18.7 | Improved |
| use View-GRPO instead of SFT | Qwen2.5-VL-3B · EgoExo-Con grounding | G-ExoEgo: 12.9 -> 14.9 | Improved |
| apply View-GRPO to InternVL3.5-8B | InternVL3.5-8B · EgoExo-Con verification | V-ExoEgo: 50.7 -> 62.4 | Improved |
| fine-tune on both viewpoints instead of exo only (ablation) | TimeSuite · CharadesEgo grounding | G-ExoEgo: 59.5 -> 51.3 | Hurt |
| fine-tune on both viewpoints instead of ego only (ablation) | VideoChat2 · CharadesEgo verification | V-ExoEgo: 36.4 -> 34.7 | Hurt |
| unfreeze visual encoder during fine-tuning (ablation) | VideoChat2 · CharadesEgo verification | V-ExoEgo: 34.69 -> 32.78 | Hurt |
| use Qwen2.5-0.5B as reasoning judge (ablation) | View-GRPO · EgoExo-Con grounding | G-ExoEgo: -3% | Hurt |
| use medium reasoning length (256 tokens) instead of 128 or 512 (ablation) | View-GRPO · reward stability | accuracy | Improved |
Closest papers by meaning
- EgoExoBench: A Benchmark for First- and Third-person View Video Understanding in MLLMs (NeurIPS 2025 · similarity 0.91)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.90)
- Omnia de EgoTempo: Benchmarking Temporal Understanding of Multi-Modal LLMs in Egocentric Videos (CVPR 2025 · similarity 0.89)
- Bootstrap Your Own Views: Masked Ego-Exo Modeling for Fine-grained View-invariant Video Representations (CVPR 2025 · similarity 0.89)
- TOMATO: Assessing Visual Temporal Reasoning Capabilities in Multimodal Foundation Models (ICLR 2025 · similarity 0.89)
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.89)
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.
