LoopQ: Quantization for Recursive Transformers

In three linesLoopQ is a post-training quantization (PTQ) framework for recursive language models (LoopLMs) that reuse Transformer blocks. It addresses three challenges: distribution shift across roles, state reuse between loops, and recursive error accumulation. Under W4A4 quantization, LoopQ improves downstream accuracy by 68.8% and reduces perplexity by 87.7% versus static PTQ baseline.

Summary generated by Claude — human-verified