Multiverse Computing described Quantization-Aware Healing (QAH) in a paper submitted to arXiv on August 21 and a longer August 25 write-up. The method aims to recover performance when an LLM is both reduced in size and stored with lower-precision weights.
The team started with GPT-OSS 120B, compressed it to 60 billion parameters and recovered a 16-bit version of that smaller model. For QAH, the researchers trained a 4-bit MXFP4 version using output probabilities from the original 120B model as its teacher. The larger teacher provides information that the already-compressed 16-bit checkpoint may have lost.
In Multiverse Computing's evaluation, the 4-bit 60B model beat its 16-bit 60B counterpart on seven of nine benchmarks. It gained 7.4 points on the AA-LCR long-context reasoning test and 5.6 on AIME 2025 mathematics. It performed slightly worse on MMLU-Pro and SciCode. Most of the comparisons do not put the smaller model ahead of the original 120B teacher.
That comparison gives the 4-bit checkpoint an additional distillation pass against the larger teacher. The published table does not include a 16-bit checkpoint given the same extra training, so it cannot isolate whether quantization itself improved accuracy.
A separate experiment on GPT-OSS 9B compared QAH with conventional quantization-aware training. QAH reached its reported peak in about 100 steps, versus roughly 700 for the other method. Those are results from the team's specified training and test setup.