Research map / Architectures and efficiency / Quantization
WinQ: Accelerating Quantization-Aware Training of Language Models Around Saddle Points
Dongyue Li, Zechun Liu, Kai Yi, Zhenshuo Zhang, Changsheng Zhao, Raghuraman Krishnamoorthi, Harshit Khaitan, Hongyang R. Zhang, Steven Li
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 noise injection (sigma=0.001) to WinQ (ablation) | WinQ · LLaMA-1B W1A16 WikiText2 | perplexity: about 4% improvement over sigma=0 | Improved |
| add weight re-initialization (alpha>0) to WinQ (ablation) | WinQ · LLaMA-1B W1A16 WikiText2 | perplexity: about 6.7% improvement over alpha=0 | Improved |
| add WinQ on top of ParetoQ | ParetoQ · LLaMA-1B W1A16 WikiText2 | perplexity: 16.9 -> 15.3 | Improved |
| add WinQ on top of ParetoQ | ParetoQ · LLaMA-1B W4A16 WikiText2 | perplexity: 10.3 -> 10.2 | No clear effect |
| add WinQ on top of ParetoQ | ParetoQ · LLaMA-3B W1.58A8 accuracy | accuracy: 55.9 -> 58.6 | Improved |
| add WinQ on top of ParetoQ with 2x training budget (40B tokens) | ParetoQ · LLaMA-1B W1A16 WikiText2 | perplexity: 16.2 -> 14.7 | Improved |
| add WinQ on top of ParetoQ for Llama-8B | ParetoQ · Llama-8B W2A16 zero-shot avg | accuracy: 64.7 -> 65.3 | Improved |
| increase noise sigma from 0.001 to 0.004 (ablation) | WinQ · Llama-1B W1A16 batch 64 WikiText2 | perplexity: 15.3 -> 18.5 | Hurt |
| reduce batch size from 64 to 16 at sigma=0.001 (ablation) | WinQ · Llama-1B W1A16 WikiText2 | perplexity: 15.3 -> 16.7 | Hurt |
| interpolate weights at alpha=0.4 before Hessian estimate (ablation) | LLaMA-1B 2-bit QAT · max absolute Hessian eigenvalue | max abs eigenvalue: 1.64 -> 3.09 | Improved |
| add noise injection sigma=0.001 to 2-bit QAT (ablation) | LLaMA-1B 2-bit QAT · max absolute Hessian eigenvalue at alpha=0.4 | max abs eigenvalue: 3.09 -> 3.95 | Improved |
| add WinQ to QuEST in 1-bit weight 1-bit activation | QuEST · 30M LLaMA SlimPajama | test perplexity: 5% reduction in test perplexity | 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.
- ParetoQ — ParetoQ: Improving Scaling Laws in Extremely Low-bit LLM Quantization (NeurIPS 2025) · better on 16 of 18 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
- QLLM: Accurate and Efficient Low-Bitwidth Quantization for Large Language Models (ICLR 2024 · similarity 0.90)
- QuEST: Stable Training of LLMs with 1-Bit Weights and Activations (ICML 2025 · similarity 0.90)
- GuidedQuant: Large Language Model Quantization via Exploiting End Loss Guidance (ICML 2025 · similarity 0.90)
- LeanQuant: Accurate and Scalable Large Language Model Quantization with Loss-error-aware Grid (ICLR 2025 · similarity 0.89)
- CBQ: Cross-Block Quantization for Large Language Models (ICLR 2025 · similarity 0.89)
- Quantization Error Propagation: Revisiting Layer-Wise Post-Training Quantization (NeurIPS 2025 · similarity 0.89)
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.
