Event

SACQ Ultra-Low-Bit VLA Quantization Framework Released

1 reports · 1 sources

Haesung Lee and Insu Han propose SACQ, an ultra-low-bit weight quantization framework for vision-language-action (VLA) models, combining structure-aware codebook fitting, action-aware mixed-bit allocation, and two-stage post-quantization refinement. According to the authors' report, on GR00T N1.7, with language/action weights at 1.5 bits, SACQ achieves an average success rate of 93.85% across four LIBERO suites, with model memory reduced by 4.83x; after additionally applying INT8 activation quantization and low-precision attention, the success rate remains 93.55%, and runtime is 1.51x faster than BF16. On SimplER, with language/action weights at 2 bits, SACQ achieves success rates of 63.76% and 52.27% on the Visual Matching and Variant Aggregation tasks, respectively, compared with 58.90% and 53.45% for BF16.