Joint stabilization and low-bit quantization at W3A3 and below

Determine how to stabilize the compensation solver while preserving acceptable accuracy at lower quantization bit widths such as W3A3, where the quantized baseline itself experiences severe degradation.

Background

The proposed method addresses numerical collapse under W4A4, where activation quantization can make the Gram matrix rank deficient and cause direct compensation solving to fail. The authors note that ill conditioning becomes worse as bit width decreases, but lower-bit settings such as W3A3 also substantially degrade the baseline model before stabilization is applied. Consequently, simultaneously maintaining baseline performance and stabilizing the compensation solver at these more aggressive quantization levels remains unresolved.

References

Second, ill conditioning worsens as bit width decreases, but at lower bit widths such as W3A3 the baseline itself degrades severely, so stabilizing both simultaneously remains unresolved.

— QuaSAR: Quantization Compensation via Stable Activation-Aware Rank Truncation  (2608.14149 - Lee et al., 14 Aug 2026) in Section Conclusion and Future Work