Researchers introduced ScaleQ-1.58, a post-training quantization framework that compresses reasoning language models to 1.58 bits using a technique called Attend to Your Own Thoughts (AYOT), which calibrates on the model’s own reasoning traces and answers rather than discarding them. The method achieves strong results on math and coding benchmarks with minimal calibration data: a quantized Qwen3-1.7B model retains 90.52% of baseline performance using only 4 million calibration tokens, while a 4B variant gains an 8.97 percentage-point absolute improvement. The approach showed consistent gains as calibration data increased, across different model scales, architectures, and task types.
