Research map / Architectures and efficiency / RNNs and normalization
Representation Drift Compensation: A Near-Zero Inference Cost Enhancement for LLM Decomposition
Xinhao Huang, You-Liang Huang, Zeyi Wen
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 learned block-level compensation bias to SVD decomposition (ablation) | SVD-LLM style low-rank decomposition · WikiText2 PPL, 30% compression | perplexity: 29.22 -> 16.16 | Improved |
| add learned compensation bias at high compression (ablation) | OPT-6.7B · WikiText2 PPL, 40% compression | perplexity: 153 -> 14.25 | Improved |
| add learned compensation bias (ablation) | OPT-13B · common-sense avg accuracy, 40% compression | avg accuracy: 42.6 -> 52.7 | Improved |
| add learned compensation bias (ablation) | OPT-30B · WikiText2 PPL, 40% compression | perplexity: 13.8 -> 10.75 | Improved |
| add compensation reduces Wasserstein distance between latent distributions (ablation) | LLaMA-3-8B · latent representations | Wasserstein distance: 0.1324 -> 0.0614 | Improved |
| use learned bias optimized on block output instead of static mean correction c = R mu (ablation) | Decomper · low-rank decomposition | qualitative | Improved |
| vary calibration set size (512 to 2048) and type (WikiText2 vs C4) (ablation) | Decomper · LLaMA-2-13B 30% compression, WikiText2 PPL | perplexity: 6.87 to 7.43 range | No clear effect |
| combine 20% Decomper with 4-bit quantization vs 3-bit GPTQ | LLaMA-2-7B · WikiText2 PPL | perplexity: 6.96 -> 6.55 | Improved |
| apply compensation to PCA-based decomposition (AFM) (ablation) | AFM on LLaMA-2-7B · WikiText2 PPL | perplexity: 18.82 -> 10.44 | Improved |
| compensation bias instead of recovery fine-tuning or least-squares update | LLaMA-2-7B 30% compression · WikiText2 PPL | perplexity: 9.21 -> 8.07 (vs FT LLM-Pruner) | Improved |
| Decomper vs FLAP on LLaMA-2-7B 20% perplexity | LLaMA-2-7B · WikiText2 PPL, 20% compression | perplexity: 6.764 -> 6.762 | No clear effect |
| Decomper vs SVD-LLM on a few individual tasks (WinoGrande, MMLU for LLaMA-3-8B vs Basis Sharing) | LLaMA-3-8B · WinoGrande, 20% compression | accuracy: 69.42 (Basis Sharing) vs 69.97 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- SVD-LLM — SVD-LLM: Truncation-aware Singular Value Decomposition for Large Language Model Compression (ICLR 2025) · better on 7 of 7 shared results
- Basis Sharing — Basis Sharing: Cross-Layer Parameter Sharing for Large Language Model Compression (ICLR 2025) · better on 4 of 4 shared results
- Dobi-SVD — Dobi-SVD: Differentiable SVD for LLM Compression and Some New Perspectives (ICLR 2025) · better on 2 of 2 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
- 3BASiL: An Algorithmic Framework for Sparse plus Low-Rank Compression of LLMs (NeurIPS 2025 · similarity 0.88)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.88)
- ESPACE: Dimensionality Reduction of Activations for Model Compression (NeurIPS 2024 · similarity 0.87)
- DISC: Dynamic Decomposition Improves LLM Inference Scaling (NeurIPS 2025 · similarity 0.87)
- S$^{2}$FT: Efficient, Scalable and Generalizable LLM Fine-tuning by Structured Sparsity (NeurIPS 2024 · similarity 0.87)
- The Silent Hyperparameter: Quantifying the Impact of Inference Backends on LLM Reproducibility (NeurIPS 2026 · 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.
