Research map / Architectures and efficiency / Pruning and efficient tuning
Shapeshifter: a Parameter-efficient Transformer using Factorized Reshaped Matrices
Aliakbar Panahi, Seyran Saeedi, Tom Arodz
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 Kronecker-based reshaped factorization with standard low-rank factorization (rank 16, embedding rank 256) (ablation) | Shapeshifter · IWSLT'14 De-En | BLEU: 36.36 -> 26.68 | Hurt |
| replace Kronecker-based reshaped factorization with standard low-rank factorization (rank 16, embedding rank 256) (ablation) | Shapeshifter · WMT'18 En-Ro | BLEU: 22.12 -> 17.47 | Hurt |
| replace Kronecker-based factorization with standard low-rank, embeddings not factorized (ablation) | Shapeshifter 16/- · IWSLT'14 De-En | BLEU: 35.85 -> 32.01 | Hurt |
| replace Kronecker-based factorization with standard low-rank, embeddings not factorized (ablation) | Shapeshifter 16/- · WMT'18 En-Ro | BLEU: 21.93 -> 19.34 | Hurt |
| replace Kronecker-based factorization with standard low-rank (rank 16/256) (ablation) | Shapeshifter · IWSLT'14 De-En | train loss: 2.59 -> 2.97 | Hurt |
| replace Kronecker-based factorization with standard low-rank (rank 16/256) (ablation) | Shapeshifter · WMT'18 En-Ro | train loss: 2.19 -> 2.32 | Hurt |
| replace Kronecker-based factorization with standard low-rank (rank 16/256) (ablation) | Shapeshifter · IWSLT'14 De-En | dev loss: 2.67 -> 2.99 | Hurt |
| replace Kronecker-based factorization with standard low-rank (rank 16/256) (ablation) | Shapeshifter · WMT'18 En-Ro | dev loss: 3.09 -> 3.37 | Hurt |
| replace PHM layers (n=16) with Shapeshifter 16/256 (about four times fewer parameters) | T5-small Transformer · WMT'18 En-Ro | BLEU: 22.38 -> 22.12 | Mixed |
| replace PHM layers (n=16) with Shapeshifter 16/256 | T5-small Transformer · IWSLT'14 De-En | BLEU: 36.49 -> 36.36 | Mixed |
| compress Transformer with Shapeshifter 16/256 versus uncompressed | T5-small Transformer · WMT'18 En-Ro | BLEU: 24.30 -> 22.12 | Hurt |
| increase Shapeshifter rank from 16/256 to 24/256 (ablation) | Shapeshifter · WMT'18 En-Ro | BLEU: 22.12 -> 22.56 | 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.
- DeLighT — DeLighT: Deep and Light-weight Transformer (ICLR 2021) · better on 3 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
- Building on Efficient Foundations: Effective Training of LLMs with Structured Feedforward Layers (NeurIPS 2024 · similarity 0.89)
- MemoryFormer : Minimize Transformer Computation by Removing Fully-Connected Layers (NeurIPS 2024 · similarity 0.88)
- Transformer Block Coupling and its Correlation with Generalization in LLMs (ICLR 2025 · similarity 0.87)
- Transformers Get Stable: An End-to-End Signal Propagation Theory for Language Models (ICML 2024 · similarity 0.87)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.87)
- Bigram Subnetworks: Mapping to Next Tokens in Transformer Language Models (NeurIPS 2025 · 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.
