Research map / Vision-language and multimodal / Human and embodied modeling
Task-Aware Structured Memory for Dynamic Multi-modal In-Context Learning
Zhirui Chen, Ziwei Chen, Ling Shao
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 |
|---|---|---|---|
| task-vector scoring instead of hard pruning (ablation) | hard-pruning baseline, 20% budget · ImageNet-100 | accuracy: 45.3 -> 55.2 | Improved |
| add semantics-aware token merging (ablation) | hard pruning + task vector scoring · ScreenSpot | score: 12.1 -> 18.8 | Improved |
| add dynamic retrieval from latent bank (ablation) | pruning + task vector + merging · YouCook2 | score: 98.5 -> 109.5 | Improved |
| layer-adaptive gating instead of uniform lambda 0.5 (ablation) | TASM · ScreenSpot | score: 16.2 -> 19.5 | Improved |
| task-vector scoring instead of answer-aware attention (ablation) | EMLoC · average over IN-100, MME, OK-VQA | average score: 54.8 -> 56.2 | Improved |
| task-vector scoring instead of attention (H2O) (ablation) | H2O scoring · average over IN-100, MME, OK-VQA | average score: 45.1 -> 56.2 | Improved |
| JS-divergence trigger instead of periodic-dense retrieval (ablation) | periodic-dense retrieval · YouCook2 | score: 108.9 -> 109.5 | Improved |
| token merging instead of hard pruning (EMLoC) at 10% retention (ablation) | EMLoC · ScreenSpot | score: 12.5 -> 15.9 | Improved |
| transfer task vector from source classes to disjoint target classes (ablation) | TASM oracle · ImageNet-100 | accuracy: 65.0 -> 64.2 | No clear effect |
| replace task vector with random projection (ablation) | TASM · ImageNet-100 | accuracy: 65.0 -> 15.3 | Hurt |
| TASM instead of EMLoC | EMLoC · V-NIAH 200 images | accuracy: 18.9 -> 45.5 | Improved |
| TASM instead of EMLoC | EMLoC · Qwen2-VL-7B efficiency at 32k | GPU memory (GB): 10.2 -> 11.5 | Hurt |
Closest papers by meaning
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.88)
- Efficient Large Multi-modal Models via Visual Context Compression (NeurIPS 2024 · similarity 0.88)
- ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference (NeurIPS 2025 · similarity 0.88)
- Filter, Correlate, Compress: Training-Free Token Reduction for MLLM Acceleration (AAAI 2026 · similarity 0.87)
- OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models (NeurIPS 2026 · similarity 0.87)
- In-context Autoencoder for Context Compression in a Large Language Model (ICLR 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.
