Repairing the damage 4-bit compression does, rather than avoiding it, surfaced in both a platform blog and a preprint the same week.
A Hugging Face post described Quantization-Aware Healing, reporting a 4-bit model that beats its full-precision original, while an arXiv preprint proposed activation-weighted seeded residual coding to repair low-bit weights. Only two items, but the framing overlaps.
The quantization conversation is shifting from how far weights can be cut to how the loss is healed afterwards. The claim of beating full precision comes from a blog-stage report and has not been independently reproduced.
Whether healing steps land in mainstream inference runtimes and quantization libraries, and whether the full-precision-beating result is reproduced independently.