Research map / Architectures and efficiency / Pruning and efficient tuning
A Unified Framework for Knowledge Transfer in Bidirectional Model Scaling
Jianlu Shen, Fu Feng, Jiaze Xu, Yucheng Xie, Jiaqi Lv, Xin Geng
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 |
|---|---|---|---|
| wavelet-based initialization instead of weight selection | BERT-S (L2S) · BERT pre-training | FLOPs saving (%): 33.0 -> 52.8 | Improved |
| wavelet synthesis instead of LiGO/Mango | BERT-B (S2L) · BERT pre-training | FLOPs saving (%): 44.9 -> 67.1 | Improved |
| wavelet synthesis instead of Mango | DeiT-B (S2L) · ImageNet-1K pre-training | FLOPs saving: +5.3% | Improved |
| wavelet decomposition instead of WS | GPT-S (L2S) · GPT pre-training | FLOPs saving: +24.1% | Improved |
| wavelet synthesis instead of LiGO/Mango | GPT-B (S2L) · GPT pre-training | FLOPs saving: +10.4% | Improved |
| wavelet decomposition init instead of WS | BERT-S · GLUE (direct fine-tune) | GLUE avg: 60.90 -> 70.29 | Improved |
| wavelet decomposition init instead of WS | BERT-S · SQuAD (direct fine-tune) | SQuAD avg: 21.28 -> 48.25 | Improved |
| wavelet init instead of scratch | DeiT-S · 7 downstream tasks | Average accuracy: +18.9% | Improved |
| wavelet init instead of WS | DeiT-S · CUB | accuracy: 57.7 -> 61.3 | Improved |
| wavelet synthesis instead of LiGO/Mango | BERT-B · GLUE after pre-training | GLUE avg: 81.60 -> 82.70 | No clear effect |
| wavelet decomposition instead of WS | BERT-S · GLUE after pre-training | GLUE avg: 75.58 -> 75.71 | No clear effect |
| choice of wavelet family (ablation) | BoT · BERT/GPT validation loss | validation loss | Mixed |
Closest papers by meaning
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.86)
- LLaVA-MoD: Making LLaVA Tiny via MoE-Knowledge Distillation (ICLR 2025 · similarity 0.85)
- MergeNet: Knowledge Migration Across Heterogeneous Models, Tasks, and Modalities (AAAI 2025 · similarity 0.85)
- Scaling Laws Meet Model Architecture: Toward Inference-Efficient LLMs (ICLR 2026 · similarity 0.85)
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.85)
- Energy-Based Transformers are Scalable Learners and Thinkers (ICLR 2026 · similarity 0.85)
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.
