QUASAR says the 2-bit QAT loss floor is a weighting bug, not a law of physics

A write-up published on dev.to summarises an August 2026 paper, QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction, which locates the accuracy floor of low-bit QAT in a specific mechanical choice rather than in the arithmetic itself ^[2]. The reported payoff is up to 29% lower held-out KL divergence at 2-bit for a 1.4% increase in training step time ^[3][4], a ratio that makes this a cheap line item in an existing pipeline rather than a research programme.

The diagnosis is worth restating because it is not exotic. In standard QAT the forward pass runs on reconstructed weights, the quantized-then-dequantized tensor, while the optimizer updates the latent full-precision weights underneath ^[7]. Gradients are taken with respect to the reconstruction via the straight-through estimator and then applied to the latent weights, and because the two are linked by a lossy function, that signal is not the best descent direction for what is actually being updated ^[7]. The model still converges; it converges higher than a full-precision run, and the authors call that residual the loss floor gap ^[6][7].

QUASAR's theoretical claim is narrow and therefore useful: the saliency-weighted distance between the reconstruction and the ideal reconstruction is the only reconstruction-dependent term in the QAT convergence bound ^[8]. If that holds, improving the reconstruction is the whole lever. The implementation is three parts, all of them cheap. Per-parameter importance is approximated by an exponential moving average of squared gradients instead of a Hessian, using quantities the training step has already computed ^[9]. The affine dequantizer, scale and zero point, is then fitted in closed form to minimise saliency-weighted reconstruction error, so weights that move the loss get priority ^[10]. Finally a small candidate set of clipping ranges is searched, picking the one with the lowest weighted error before the dequantizer is fitted ^[11].

On cost: a 1.4% step-time increase turns a nominal 100-hour training run into 101.4 hours ^[13]. Compared against the headline 29% KL improvement, the benefit is roughly twenty times the cost in percentage terms, with the caveat that the two percentages measure unrelated quantities ^[14]. The context is that post-training methods such as GPTQ and AWQ need no training infrastructure at all but degrade sharply at 2-bit and 3-bit, which is exactly where QAT earns its keep ^[5][6], and where 16-bit to 2-bit compression is an eightfold reduction in weight memory ^[1][15].

Now the limits of what has actually been supplied. This is a secondary write-up, not the paper, and its text breaks off at the point where the 2-bit results are enumerated, so the per-model numbers behind the 29% figure are not in front of us ^[16]. The evaluation is described as covering the Qwen3 and Llama-3.1 families at 2, 3 and 4 bits against QAT and PTQ baselines ^[12], but "up to 29%" is doing unspecified work across that grid, and KL divergence against the full-precision teacher is not the same as downstream task accuracy.

What to watch: whether the 1.4% overhead holds at production tensor-parallel scale, since the closed-form fit and range search run per weight tensor per step ^[10][11]; and whether the improvement survives translation into task metrics rather than KL.