Research map / Architectures and efficiency / Pruning and efficient tuning
A Robust Optimization Guided Pruning Framework for Vision and Large Language Models
Gabriel Afriat, Hussein Hazimeh, Dimitris Paparas, Rahul Mazumder
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 Frobenius-norm uncertainty set (adds gamma*I) instead of default tiny dampening | CAP on DeiT-Base · ImageNet-1K, 80% sparsity | accuracy: 45.19 -> 55.19 | Improved |
| use Frobenius-norm uncertainty set | CAP on DeiT-Small · ImageNet-1K, 70% sparsity | accuracy: 53.77 -> 60.04 | Improved |
| use Frobenius-norm uncertainty set | CAP on DeiT-Tiny · ImageNet-1K, 80% sparsity | accuracy: 8.22 -> 8.34 | No clear effect |
| use Frobenius-norm uncertainty set | CAP on DeiT-Base · ImageNet-1K, 2:4 sparsity | accuracy: 75.29 -> 77.34 | Improved |
| use eigenvalue-wise uncertainty set | ALPS on Llama-3.1-8B · WikiText2, 70% sparsity | perplexity: 30.16 -> 29.53 | No clear effect |
| use Frobenius-norm uncertainty set | ALPS on Llama-3.1-8B · WikiText2, 80% sparsity | perplexity: 113.18 -> 105.01 | Improved |
| use Frobenius-norm uncertainty set | ALPS on Llama-2-13b-hf · PTB, 80% sparsity | perplexity: 503.36 -> 529.50 | Hurt |
| use eigenvalue-wise uncertainty set | ALPS on Llama-3.1-8B · pruning time, 60% sparsity | minutes: 53.97 -> 56.52 | Hurt |
| use eigenvalue-wise uncertainty set | CAP on DeiT-Base · pruning time | minutes: 45.58 -> 81.51 | Hurt |
| use Frobenius-norm uncertainty set (gamma-based) | ALPS on Llama-3.1-8B · mean accuracy of 3 tasks, 70% sparsity | accuracy: 48.86 -> 48.99 | No clear effect |
| use gradient-proportional uncertainty set instead of constant-radius (ablation) | RobOP-CAP on DeiT-Base · ImageNet-1K, 80% sparsity | accuracy: 55.19 -> 47.36 | Hurt |
| tune gamma instead of default CAP dampening 1e-8 (ablation) | CAP on DeiT-Base · ImageNet-1K, 80% sparsity | accuracy | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ALPS — ALPS: Improved Optimization for Highly Sparse One-Shot Pruning for Large Language Models (NeurIPS 2024) · better on 6 of 7 shared results
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
- ECoFLaP: Efficient Coarse-to-Fine Layer-Wise Pruning for Vision-Language Models (ICLR 2024 · similarity 0.90)
- Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization (NeurIPS 2025 · similarity 0.90)
- OSSCAR: One-Shot Structured Pruning in Vision and Language Models with Combinatorial Optimization (ICML 2024 · similarity 0.89)
- Each Complexity Deserves a Pruning Policy (NeurIPS 2025 · similarity 0.89)
- MoPE-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric (CVPR 2024 · similarity 0.89)
- Collaborative Multi-Mode Pruning for Vision-Language Models (CVPR 2026 · similarity 0.89)
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.
