Register Tokens for Bounded-State Reasoning in Diffusion Language Models¶
L1 mechanism diagnostic. Bounded continuous carry only, without LLaDA/Dream checkpoints or diffu-GRPO.
Paper information¶
| Field | Value |
|---|---|
| Paper | Register Tokens for Bounded-State Reasoning in Diffusion Language Models |
| Affiliations | University of Wisconsin–Madison, first author's affiliation |
| First public release | 2026-09-14 (arXiv v1) |
| Authors' code | No public author repository found as of 2026-09-16. |
| Adapter / method | register-tokens-dllm |
| Local code | src/auto_research/foundation_latest_20260916_followup.py |
Original paper¶
Background and contribution¶
Discard previous chunk text and carry a fixed number of continuous register hidden states into the next diffusion chunk, supporting cross-chunk reasoning without retaining all text.
flowchart LR
I[Chunk state] --> R[Fixed register bank] --> C[Clear previous text] --> N[Next chunk with carried registers]Key figure from the paper¶
Original Figure 5: register-token attention during denoising. From the paper; copyright belongs to the authors. Not a local result.
Core mechanism¶
The reference exposes auditable state transitions. Three-seed fixtures test invariants and bounds, not capability from random mini-suite scores.
Paper results¶
The authors' accuracy, speed, loss, and benchmark results are separate from local diagnostics or online gains.
Local reproduction¶
Three-seed mechanism metrics carry diagnostic_only=true and are not used for formal capability rankings.
Reproduction limits¶
No LLaDA/Dream checkpoint or diffu-GRPO run.