Research map / Continual learning / Continual fine-tuning of LLMs
ConStruct-VL: Data-Free Continual Structured VL Concepts Learning
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 |
|---|---|---|---|
| Use Layered-LoRA with Adversarial Pseudo-Replay instead of the strongest data-free baseline | BLIP · VG+VAW, 7 tasks | Final accuracy: 79.39 -> 85.4 | Helped |
| Use Layered-LoRA with Adversarial Pseudo-Replay instead of the strongest data-free baseline | BLIP · VG, 7 tasks | Final accuracy: 81.30 -> 86.88 | Helped |
| Use Layered-LoRA with Adversarial Pseudo-Replay instead of the strongest baseline | BLIP · VAW, 4 tasks | Final accuracy: 86.00 -> 87.50 | Helped |
| Add Adversarial Pseudo-Replay to Layered-LoRA (ablation) | BLIP, VG+VAW · VG+VAW, 6 tasks | Final accuracy: 73.68 -> 87.11 | Helped |
| Replace APR with LwF distillation from past models (ablation) | Layered-LoRA on BLIP · VG+VAW, 6 tasks | Final accuracy: 87.11 -> 70.83 | Hurt |
| Generate positive rather than negative pseudo-replay examples (ablation) | Layered-LoRA with APR · VG+VAW, 6 tasks | Final accuracy: 87.11 -> 76.3 | Hurt |
| Use positive pseudo-replay with only the previous task model (ablation) | Layered-LoRA with APR · VG+VAW, 6 tasks | Final accuracy: 87.11 -> 80.98 | Hurt |
| Generate adversarial images instead of adversarial text (ablation) | Layered-LoRA with APR · VG+VAW, 6 tasks | Final accuracy: 87.11 -> 81.2 | Hurt |
| Generate adversarial text by randomly sampling other batch texts rather than attacking the past model (ablation) | Layered-LoRA with APR · VG+VAW, 6 tasks | Final accuracy: 87.11 -> 82.71 | Hurt |
| Apply APR only with the previous task model rather than all past task models (ablation) | Layered-LoRA with APR · VG+VAW, 6 tasks | Final accuracy: 87.11 -> 84.69 | Hurt |
Closest papers by meaning
- Seeing Clearly, Reasoning Confidently: Plug-and-Play Remedies for Vision Language Model Blindness (CVPR 2026 · similarity 0.85)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.85)
- VILA: On Pre-training for Visual Language Models (CVPR 2024 · similarity 0.85)
- Masking Teacher and Reinforcing Student for Distilling Vision-Language Models (CVPR 2026 · similarity 0.85)
- Probing Visual Language Priors in VLMs (ICML 2025 · similarity 0.85)
- Unified Reinforcement and Imitation Learning for Vision-Language Models (NeurIPS 2025 · 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.