Linear-attention and state-space LLMs carry long context in a fixed-size recurrent state. Quantizing that state is a natural way to cut its cost during long decode. Unlike a weight tensor, however, the state is rewritten at every step, which makes it a moving target for quantization.

The paper reports that dynamic scaling fails for this recurrent state. Instead, the authors identify a 'scale-state feedback' that governs how the state should be quantized. The abstract does not spell out the mechanism in detail, but the title and framing suggest that the state's own scale must feed back into the quantization rule.

If correct, the finding points to a more principled way to compress recurrent state in linear-attention and SSM LLMs, potentially lowering the cost of long-context inference. As a single arXiv preprint, it has not yet been peer-reviewed.