Research map / Vision-language and multimodal / Multimodal LLMs
KATANA: Knowledge-Aligned Topology-Aware Neural Agents for RL-Driven Vision-Language Model Compression
Nafew Azim, Mir Ali, Fuad Rahman, Nabeel Mohammed
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 |
|---|---|---|---|
| use cubic sparsity schedule instead of linear (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 78.2 -> 84.6 | Improved |
| use cubic sparsity schedule instead of constant (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 76.5 -> 84.6 | Improved |
| add structural norms to DNA metric (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 80.1 -> 84.6 | Improved |
| add activation term to DNA metric (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 79.5 -> 84.6 | Improved |
| use DNA instead of magnitude-only importance (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 77.3 -> 84.6 | Improved |
| add low-rank reconstruction (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 81.3 -> 84.6 | Improved |
| add novelty reward to RL (ablation) | KATANA · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 79.8 -> 84.6 | Improved |
| add diversity reward to RL (ablation) | KATANA · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 80.4 -> 84.6 | Improved |
| add curriculum weighting to RL (ablation) | KATANA · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 77.9 -> 84.6 | Improved |
| use GRPO policy updates instead of random evolutionary search (ablation) | KATANA · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 76.2 -> 84.6 | Improved |
| reduce search from 1,000 to 100 iterations | KATANA · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 84.6 -> 78.2 | Hurt |
| reduce calibration set from N=128 to N=64 (ablation) | KIRI · MSCOCO (70% sparsity, LLaVA-1.5-7B) | CIDEr: 1.8 CIDEr drop | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SparseVLM — SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025) · better on 3 of 3 shared results
- SCOPE — SCOPE: Saliency-Coverage Oriented Token Pruning for Efficient Multimodel LLMs (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.90)
- LLMC+: Benchmarking Vision-Language Model Compression with a plug-and-play Toolkit (AAAI 2026 · similarity 0.89)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.88)
- Unified Knowledge Maintenance Pruning and Progressive Recovery with Weight Recalling for Large Vision-Language Models (AAAI 2025 · similarity 0.88)
- UniCompress: Token Compression for Unified Vision-Language Understanding and Generation (CVPR 2026 · similarity 0.88)
- Hybrid Token Compression for Vision-Language Models (CVPR 2026 · similarity 0.88)
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.
