GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling

The core novelty is the mechanism, not just the outcome: GSQ uses Gumbel-Softmax sampling during calibration to make the discrete quantization grid selection differentiable, which is what lets it optimize accuracy without the lookup-table overhead that makes vector quantization methods like AQLM expensive to deploy. The accuracy gain is real, but the deployment story is the actual argument.

This connects most directly to the inference-efficiency thread running through recent Modelwire coverage. The SpecGuard paper from April 16 ('From Tokens to Steps') attacked latency from the decoding side; the K-Token Merging paper from the same date attacked it from the sequence-compression side. GSQ attacks it from the weight-storage side. Together they sketch a picture of the field converging on a stack of complementary compression techniques rather than any single solution. The Latent Phase-Shift Rollback paper from April 20 is a reminder that inference-time correctness and inference-time efficiency are being pursued in parallel, and that a highly compressed model running faster may still need error-correction scaffolding on top.

The real test is whether GSQ's accuracy claims hold when applied to models beyond the standard LLaMA evaluation suite, particularly on instruction-tuned or RLHF-fine-tuned checkpoints where weight distributions differ. If third-party reproductions on Mistral or Qwen variants show similar perplexity gaps closing at 2-bit, the method is robust; if gains shrink, the calibration set sensitivity is the culprit.