AdaMX recovers 83% of low-bit accuracy loss

AdaMX recovered 83% of MXFP4’s commonsense accuracy loss and 82% of its MMLU loss across language models from 3B to 70B parameters. The new adaptive-microscaling paper says its 22nm accelerator chooses a recovery scheme per block and a representation per operand without increasing equivalent bit width, adding roughly 1% system energy; on Gemma 4 12B, it retained as much as 96% of FP16 accuracy across four vision-language benchmarks. Builders weighing compression against more inference capacity should file this beside the archive’s warning that model-compression averages can hide task-specific failures: AdaMX makes a promising hardware case, but the authors’ preprint still needs workload-level replication.