{
  "id": 1515360,
  "title": "QUASAR: How Saliency-Weighted Reconstruction Closes the Loss Floor Gap in LLM Quantization-Aware Training",
  "url": "https://urgent.news/2026/08/17/quasar-how-saliency-weighted-reconstruction-closes-the-loss-floor-gap",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-17T16:20:19.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/prabhakar_chaudhary_7afe4/quasar-how-saliency-weighted-reconstruction-closes-the-loss-floor-gap-in-llm-quantization-aware-5akb"
  },
  "original_language": "en",
  "account": "QUASAR: How Saliency-Weighted Reconstruction Closes the Loss Floor Gap in LLM Quantization-Aware Training\n\nQuantization enables large language models (LLMs) to run on low-resource devices by reducing memory requirements and computational costs. However, extreme low-bit quantization (2-bit or 3-bit) can significantly degrade model accuracy due to rounding errors and the quantization loss floor gap. A recent paper, QUASAR: Lowering the Loss Floor of Quantization-Aware Training (QAT) with Loss-Aware Reconstruction, proposes a lightweight solution to address this issue.\n\nThe paper identifies the loss floor gap as the primary cause of accuracy loss in QAT. During QAT training, the forward pass uses reconstructed weights from the quantized-dequantized process, while the optimizer updates the full-precision weights. This mismatch creates a gap between the achievable loss with full-precision weights and the loss floor that QAT converges to.\n\nQUASAR proposes a lightweight reconstruction step to close the loss floor gap. At each training step, it estimates the importance of each weight parameter using the exponential moving average of squared gradients, fits the best affine dequantizer (scale and zero-point) in closed form, and searches for the optimal clipping range to minimize the saliency-weighted reconstruction error. This approach prioritizes the accurate reconstruction of parameters that matter most to the loss.\n\nExperiments on Qwen3 and Llama-3.1 models show that QUASAR achieves up to 29% reduction in held-out KL divergence at 2-bit quantization compared to standard QAT, while improving accuracy by 3.5–4.3 percentage points across eight downstream tasks. The method also improves accuracy over PTQ baselines like GPTQ and AWQ. At 3-bit and 4-bit, the gains are smaller but consistent, with QUASAR achieving the lowest held-out KL divergence among competitive QAT methods.\n\nQUASAR's implementation overhead is minimal, with only a 1.4% increase in training step time due to the saliency EMA computation and closed-form dequantizer fitting. The method does not introduce any inference-time overhead, making it a straightforward drop-in improvement to the QAT training loop without sacrificing runtime performance.",
  "summary": "QUASAR: How Saliency-Weighted Reconstruction Closes the Loss Floor Gap in LLM Quantization-Aware Training Quantization is one of the most practical tools in the LLM deployment toolkit. Shrinking a model from 16-bit to 4-bit or even 2-bit precision can cut memory requirements by 4–8×, making it possible to run large models on consumer hardware, edge devices, or cost-constrained cloud instances.…",
  "key_points": [
    "QUASAR addresses loss floor gap in 2-bit/3-bit quantization of LLMs",
    "Saliency-weighted reconstruction prioritizes accurate parameter reconstruction",
    "QUASAR achieves up to 29% reduction in KL divergence at 2-bit quantization"
  ],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}