Research map / Architectures and efficiency / Quantization
WaterSIC: Information-Theoretically (Near) Optimal \\Linear Layer Quantization
Egor Lifar, Semyon Savkin, Or Ordentlich, Yury Polyanskiy
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 post-quantization finetuning of rescalers (ablation) | WaterSIC (Llama-3.2-1B, 1.00 bits) · WikiText-2 | WikiText-2 perplexity: 82.44 -> 30.47 | Improved |
| add post-quantization finetuning of rescalers (ablation) | WaterSIC (Llama-3.2-1B, 1.50 bits) · WikiText-2 | WikiText-2 perplexity: 30.73 -> 18.50 | Improved |
| add post-quantization finetuning of rescalers (ablation) | WaterSIC (Llama-3.2-1B, 2.00 bits) · WikiText-2 | WikiText-2 perplexity: 16.19 -> 13.59 | Improved |
| add post-quantization finetuning of rescalers (ablation) | WaterSIC (Llama-3.2-1B, 4.00 bits) · WikiText-2 | WikiText-2 perplexity: 9.92 -> 9.91 | Improved |
| add post-quantization finetuning of rescalers (ablation) | WaterSIC (Qwen3-8B, 2.125 bits) · WikiText-2 | WikiText-2 perplexity: 11.37 -> 10.18 | Improved |
| add post-quantization finetuning of rescalers (ablation) | WaterSIC (Qwen3-8B, 4.125 bits) · WikiText-2 | WikiText-2 perplexity: 9.79 -> 9.67 | Improved |
| use WaterSIC instead of Huffman-GPTQ | Llama-3.2-1B at about 2 bits · WikiText-2 | WikiText-2 perplexity: 86.80 -> 16.19 | Improved |
| use WaterSIC instead of Huffman-GPTQ | Qwen3-8B at 2.125 bits · WikiText-2 | WikiText-2 perplexity: 13.97 -> 11.37 | Improved |
| use WaterSIC instead of GPTQ | Qwen3-8B at 3.125 bits · WikiText-2 | WikiText-2 perplexity: 12.77 -> 10.03 | Improved |
| use WaterSIC instead of QTIP | Llama-3.2-1B at about 3 bits · WikiText-2 | WikiText-2 perplexity: 11.17 -> 10.57 | Improved |
| use WaterSIC instead of AWQ (g128) | Llama-3.2-1B at about 3.2 bits · WikiText-2 | WikiText-2 perplexity: 16.74 -> 10.35 | Improved |
Closest papers by meaning
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.84)
- QuIP$\#$: Even Better LLM Quantization with Hadamard Incoherence and Lattice Codebooks (ICML 2024 · similarity 0.84)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.83)
- Quantization without Tears (CVPR 2025 · similarity 0.83)
- LQER: Low-Rank Quantization Error Reconstruction for LLMs (ICML 2024 · similarity 0.83)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.83)
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.
