Research map / Architectures and efficiency / Quantization
CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs
Shigeng Wang, Chao Li, Yangyuxuan Kang, Jiawei Fan, Anbang Yao
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 learnable modulation (three factors) over static approximation (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 40.16 -> 57.06 | Improved |
| add softened ternarization on top of learnable modulation (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 53.19 -> 57.06 | Improved |
| add softened ternarization on top of learnable modulation (ablation) | CAT-Q, Llama2-7B · Commonsense zero-shot | average accuracy (%): 55.97 -> 57.69 | Improved |
| learnable modulation instead of direct learning of alpha and threshold (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 46.94 -> 57.06 | Improved |
| hard ternarization stage after differentiable stage (gamma 0.8 vs 1.0) (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 55.02 -> 57.06 | Improved |
| set s0 to 30 instead of 1000 (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 50.82 -> 57.06 | Improved |
| drop mu from weight reconstruction (W ~ alpha T) (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 56.59 -> 57.06 | Improved |
| learnable threshold instead of hand-crafted 0.5 (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 55.53 -> 57.06 | Improved |
| increase calibration samples from 512 to 1024 (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 57.06 -> 58.80 | Improved |
| increase calibration samples from 512 to 1024 (ablation) | CAT-Q, Llama2-7B · Commonsense zero-shot | average accuracy (%): 57.69 -> 60.50 | Improved |
| use group size 128 instead of channel-wise (ablation) | CAT-Q, Qwen3-4B · Commonsense zero-shot | average accuracy (%): 55.06 -> 57.06 | Improved |
| use CAT-Q ternary format instead of Q4_K_M in llama.cpp | Qwen3-4B · GPU decoding throughput | tokens/s: 225.91 -> 320.34 | Improved |
Closest papers by meaning
- ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025 · similarity 0.91)
- QBB: Quantization with Binary Bases for LLMs (NeurIPS 2024 · similarity 0.91)
- GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling (NeurIPS 2026 · similarity 0.91)
- SpQR: A Sparse-Quantized Representation for Near-Lossless LLM Weight Compression (ICLR 2024 · similarity 0.90)
- Q-Palette: Fractional-Bit Quantizers Toward Optimal Bit Allocation for Efficient LLM Deployment (NeurIPS 2025 · similarity 0.90)
- ABQ-LLM: Arbitrary-Bit Quantized Inference Acceleration for Large Language Models (AAAI 2025 · similarity 0.90)
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.
