Research map / Continual learning / Continual fine-tuning of LLMs
HypCL: Adapting CLIP in Hyperbolic Space for Continual Learning
Quan Cheng, Hao Yu, Da-Wei Zhou, Lijun Zhang
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 |
|---|---|---|---|
| replace Euclidean adapter and cosine logits with hyperbolic update composition and hyperbolic distance logits (no prototypes) (ablation) | HypCL · CIFAR100 B0 Inc10 | final accuracy: 74.25 -> 79.77 | Improved |
| replace Euclidean adapter and cosine logits with hyperbolic composition and distance logits (no prototypes) (ablation) | HypCL · ImageNet-R B0 Inc20 | final accuracy: 81.02 -> 83.00 | Improved |
| add visual prototype calibration (GDA logits fused at inference) (ablation) | HypCL · CIFAR100 B0 Inc10 | final accuracy: 79.77 -> 82.95 | Improved |
| add visual prototype calibration (GDA logits fused at inference) (ablation) | HypCL · ImageNet-R B0 Inc20 | final accuracy: 83.00 -> 83.67 | Improved |
| increase prototype fusion weight eta from 0.5 to 0.8 (ablation) | HypCL · ImageNet-R B0 Inc20 | final accuracy: 83.67 -> 82.33 | Hurt |
| increase adapter bandwidth b from 1 to 5 (ablation) | HypCL · ImageNet-R B0 Inc20 | average accuracy: 88.39 -> 88.98 | Improved |
| increase adapter bandwidth beyond 5 (b = 7, 9) (ablation) | HypCL · ImageNet-R B0 Inc20 | final accuracy: 83.67 -> 83.43 (b=7), 83.33 (b=9) | No clear effect |
| use HypCL (no exemplars) instead of exemplar-based PROOF | CLIP ViT-B/16 · Aircraft B0 Inc10 | average accuracy: 61.00 -> 71.34 | Improved |
| use HypCL instead of Engine | CLIP ViT-B/16 · UCF B0 Inc10 | average accuracy: 94.35 -> 93.99 | Hurt |
| use HypCL instead of Engine on ObjectNet | CLIP ViT-B/16 · ObjectNet B0 Inc20 | average accuracy: 59.11 -> 62.31 | Improved |
Closest papers by meaning
- Hyperbolic Displacement-Constrained Adaptation for CLIP-Based Class-Incremental Learning (NeurIPS 2026 · similarity 0.91)
- Mind the Gap: Preserving and Compensating for the Modality Gap in CLIP-Based Continual Learning (ICCV 2025 · similarity 0.88)
- Hierarchical Semantic Tree Anchoring for CLIP-Based Class-Incremental Learning (NeurIPS 2026 · similarity 0.88)
- TreeLoRA: Efficient Continual Learning via Layer-Wise LoRAs Guided by a Hierarchical Gradient-Similarity Tree (ICML 2025 · similarity 0.87)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.87)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · 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.
