Research map / Architectures and efficiency / Pruning and efficient tuning
Decompose, Mix, Adapt: A Unified Framework for Parameter-Efficient Neural Network Recombination and Compression
Nazia Tasnim, Shrimai Prabhumoye, Bryan A. Plummer
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 |
|---|---|---|---|
| add sigmoid (SiLU-style) gating to mixer matrix instead of none (ablation) | CRISP · CUB/CIFAR-100/FGVC-Aircraft (mixer regularization) | average accuracy: 81.8 -> 82.2 | Improved |
| use ReLU gating instead of SiLU (ablation) | CRISP · CUB/CIFAR-100/FGVC-Aircraft (mixer regularization) | average accuracy: 82.2 -> 61.5 | Hurt |
| use GELU gating instead of SiLU (ablation) | CRISP · CUB/CIFAR-100/FGVC-Aircraft (mixer regularization) | average accuracy: 82.2 -> 81.6 | No clear effect |
| add explicit orthogonal regularization instead of SiLU gating (ablation) | CRISP · CUB/CIFAR-100/FGVC-Aircraft (mixer regularization) | average accuracy: 81.8 -> 82.3 | No clear effect |
| add distillation stage after neural mimicry (ablation) | CRISP ViT-B/16 50% compression · Six fine-grained (50% compression, ablation) | average accuracy: 57.8 -> 88.8 | Improved |
| add SVD initialization of the compressed model (ablation) | CRISP ViT-B/16 50% compression · Six fine-grained (50% compression, ablation) | average accuracy: 85.1 -> 88.8 | Improved |
| use CRISP matrix mixer instead of RECAST vector mixer | PEFT on VTAB-1K · VTAB-1K (19 tasks, overall) | accuracy: 55.3 -> 59.2 | Improved |
| use CRISP instead of RECAST at 75% compression | ViT-B/16 · Six fine-grained (75% compression, classifier only) | average accuracy: 70.2 -> 80.5 | Improved |
| increase mixer rows r from 8 to 64 with s fixed at 16 (ablation) | CRISP · CIFAR-100/CUB-Birds/FGVC-Aircraft | accuracy: 22.1M -> 3.5M parameters | Hurt |
| increase mixer columns s from 8 to 32 with r fixed at 16 (ablation) | CRISP · CIFAR-100/CUB-Birds/FGVC-Aircraft | accuracy: 6.1M -> 22.1M parameters | Improved |
| increase mixer columns s from 32 to 64 (ablation) | CRISP · CIFAR-100/CUB-Birds/FGVC-Aircraft | accuracy: 22.1M -> 43.4M parameters | No clear effect |
| use CRISP single matrix multiply instead of RECAST K-fold averaging | training throughput ViT-B/16 · Efficiency (L40S) | training throughput (samples/s): 3.3 -> 163 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- RoSA — RoSA: Rotational Sparse Adaptation for Memory-Efficient Fine-Tuning (NeurIPS 2026) · better on 1 of 1 shared result
- Basis-Sharing — Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025) · better on 1 of 1 shared result
- DoRA — DoRA: Weight-Decomposed Low-Rank Adaptation (ICML 2024) · better on 1 of 1 shared result
- VBLoRA — VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks (NeurIPS 2024) · 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
- RECOMBINER: Robust and Enhanced Compression with Bayesian Implicit Neural Representations (ICLR 2024 · similarity 0.85)
- DARE the Extreme: Revisiting Delta-Parameter Pruning For Fine-Tuned Models (ICLR 2025 · similarity 0.85)
- Towards Meta-Pruning via Optimal Transport (ICLR 2024 · similarity 0.85)
- ModHiFi: Identifying High Fidelity predictive components for Model Modification (NeurIPS 2025 · similarity 0.84)
- In defense of parameter sharing for model-compression (ICLR 2024 · similarity 0.84)
- Compress to Impress: Efficient LLM Adaptation Using a Single Gradient Step on 100 Samples (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.
