Research map / Continual learning / Continual fine-tuning of LLMs
Protect to Adapt: Orthogonal Subspace Control with Ranked Negative-Prompt Curriculum for Few-Shot Action Recognition
Hantao Qi, Yan Yan, Junlong Gao, Hanzi Wang
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 OSC (orthogonal-subspace LoRA) to vanilla LoRA (ablation) | P2A · HMDB51 5-way 1-shot | accuracy (%): 78.6 -> 79.9 | Improved |
| add OSC to vanilla LoRA (ablation) | P2A · Kinetics-100 5-way 1-shot | accuracy (%): 90.7 -> 91.8 | Improved |
| add RNC curriculum negatives on top of OSC (ablation) | P2A · HMDB51 5-way 1-shot | accuracy (%): 79.9 -> 82.3 | Improved |
| add RNC curriculum negatives on top of OSC (ablation) | P2A · Kinetics-100 5-way 1-shot | accuracy (%): 91.8 -> 93.7 | Improved |
| use three-phase easy-medium-hard curriculum instead of a single difficulty level (ablation) | P2A · HMDB51 5-way 1-shot | accuracy (%): 80.7 -> 82.3 | Improved |
| add negative prompts without OSC to LoRA (ablation) | CLIP-FSAR with LoRA · HMDB51 5-way 1-shot | accuracy (%): 78.6 -> 77.3 | Hurt |
| add easy negative prompts to fully fine-tuned backbone (ablation) | CLIP-FSAR full fine-tuning · HMDB51 5-way 1-shot | accuracy (%): 75.8 -> 73.8 | Hurt |
| use entropy-based effective rank instead of fixed subspace dimension (ablation) | OSC · HMDB51 5-way 1-shot | accuracy (%): 79.6 -> 79.9 | Improved |
| apply OSC to upper layers 6-11 instead of bottom layers (ablation) | OSC · HMDB51 5-way 1-shot | accuracy (%): 77.8 -> 79.9 | Improved |
| apply OSC to all 12 layers instead of upper half (ablation) | OSC · HMDB51 5-way 1-shot | accuracy (%): 79.9 -> 79.5 | Hurt |
| add Wo projection to OSC on WqWvWk (ablation) | OSC · HMDB51 5-way 1-shot | accuracy (%): 79.9 -> 79.7 | No clear effect |
| vary subspace dimension K from 1 to 128 (ablation) | OSC · HMDB51 5-way 1-shot | accuracy (%): 79.6 -> 79.4 | No clear effect |
Closest papers by meaning
- Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · similarity 0.87)
- Efficient Transfer Learning for Video-language Foundation Models (CVPR 2025 · similarity 0.87)
- Frame Order Matters: A Temporal Sequence-Aware Model for Few-Shot Action Recognition (AAAI 2025 · similarity 0.87)
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.86)
- PhaseLoRA: Control-Regime-Conditioned Low-Rank Adaptation for Continuous-Action Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.86)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.86)
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.
