Research map / Continual learning / Continual fine-tuning of LLMs
BMU-MoCo: Bidirectional Momentum Update for Continual Video-Language Modeling
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 bidirectional momentum update with local momentum encoders (ablation) | Base-MoCo · CVLM five video-text tasks | overall R@1: 30.82 -> 34.65 | Helped |
| Add bidirectional momentum update with local momentum encoders (ablation) | Base-MoCo · CVLM five video-text tasks | overall FR: 43.40 -> 25.79 | Helped |
| Add global momentum encoders on top of BMU local (ablation) | BMU-MoCo (local) · CVLM five video-text tasks | overall R@1: 34.65 -> 35.47 | Helped |
| Add global momentum encoders on top of BMU local (ablation) | BMU-MoCo (local) · CVLM five video-text tasks | overall FR: 25.79 -> 22.63 | Helped |
| Use both local and global encoders without BMU instead of one (ablation) | Base-MoCo · CVLM five video-text tasks | overall R@1: 31.47 -> 32.35 | Helped |
| Apply BMU on top of local and global encoders (ablation) | Base-MoCo + local + global · CVLM five video-text tasks | overall HM: 35.67 -> 37.59 | Helped |
| Update encoders iteration by iteration instead of ensembling at task end (ablation) | BMU-MoCo · CVLM five video-text tasks | overall HM: 35.67 -> 37.59 | Helped |
| Do not fix momentum encoders during each task (ablation) | BMU-MoCo · CVLM five video-text tasks | overall R@1: 34.07 -> 35.47 | Helped |
| Fix momentum encoders during each task (ablation) | BMU-MoCo · CVLM five video-text tasks | overall FR: 22.63 -> 13.81 | Mixed |
| Lower momentum coefficient m_hat from 0.99 to 0.9 (ablation) | BMU-MoCo · CVLM five video-text tasks | overall HM: 37.59 -> 36.82 | Hurt |
| Raise momentum coefficient m_hat from 0.99 to 0.999 (ablation) | BMU-MoCo · CVLM five video-text tasks | overall HM: 37.59 -> 36.19 | Hurt |
| Use BMU-MoCo without memory instead of ER-ring with 10% memory | BMU-MoCo vs ER-ring · CVLM five video-text tasks | overall R@1 (final): 32.54 -> 35.47 | Helped |
Closest papers by meaning
- ViLCo-Bench: VIdeo Language COntinual learning Benchmark (NeurIPS 2024 · similarity 0.89)
- Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025 · similarity 0.87)
- CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · similarity 0.87)
- Embracing Language Inclusivity and Diversity in CLIP through Continual Language Learning (AAAI 2024 · similarity 0.86)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.85)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.85)
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.