Research map / Architectures and efficiency / Pruning and efficient tuning
Rethinking Parameter Sharing as Graph Coloring for Structured Compression
Boyang Zhang, Daning Cheng, Yunquan Zhang, Fangming Liu
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 |
|---|---|---|---|
| replace Basis Sharing heuristic with Geo-Sharing coloring | LLaMA-7B 20% · WikiText-2 | perplexity: 7.74 -> 7.07 | Improved |
| replace Basis Sharing with Geo-Sharing | LLaMA-7B 50% · WikiText-2 | perplexity: 19.99 -> 18.95 | Improved |
| replace Basis Sharing with Geo-Sharing | LLaMA-7B 50% · PTB | perplexity: 126.35 -> 117.23 | Improved |
| replace Basis Sharing with Geo-Sharing | LLaMA-7B 20% · 7 zero-shot tasks | average accuracy: 48.0 -> 49.1 | Improved |
| replace Basis Sharing with Geo-Sharing | LLaMA2-7B 50% · WikiText-2 | perplexity: 21.3 -> 20.16 | Improved |
| replace Basis Sharing with Geo-Sharing | LLaMA-7B 80% · C4 | perplexity: 2465.999 -> 995.33 | Improved |
| replace Basis Sharing with Geo-Sharing | OPT-6.7B 20% · WikiText-2 | perplexity: 11.79 -> 11.68 | Improved |
| replace FiPS with Geo-Sharing | Swin-L · ImageNet | Top-1 drop: 0.03 -> 0.01 | Improved |
| share parameters with Geo-Sharing | DeiT-B · CIFAR-100 transfer | accuracy: 72.80 -> 70.30 | Hurt |
| share parameters with Geo-Sharing | Swin-B · CIFAR-100 transfer | accuracy: 67.39 -> 68.05 | No clear effect |
| increase number of short-axis eigenvalues t (ablation) | Geo-Sharing · LLaMA perplexity | perplexity | Improved |
| increase amplitude factor beta beyond 1e-1 (ablation) | Geo-Sharing · LLaMA perplexity | perplexity | 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.
- Basis Sharing — Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025) · better on 17 of 17 shared results
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 4 of 4 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
- Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology (NeurIPS 2026 · similarity 0.85)
- Forget the Data and Fine-Tuning! Just Fold the Network to Compress (ICLR 2025 · similarity 0.85)
- In defense of parameter sharing for model-compression (ICLR 2024 · similarity 0.85)
- Effective Structural Encodings via Local Curvature Profiles (ICLR 2024 · similarity 0.85)
- Subspace Networks: Scaling Decentralized Training with Communication-Efficient Model Parallelism (NeurIPS 2025 · similarity 0.85)
- Distilling Graph Geometry: Knowledge Gap from GNNs to MLPs (NeurIPS 2026 · 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.
