Tag
This paper proposes using register tokens in diffusion language models to maintain bounded-state reasoning across generation chunks, showing substantial gains on math and code benchmarks without full context retention.