Check Token: Real-Time Self-Verification and Precise Truncation in LLM Reasoning
The authors report in this preprint: Researchers propose the 'check token,' a functional marker inspired by ECC memory parity bits that enables real-time self-verification and truncation during large language model generation.

Original image temporarily unavailable
The authors report in this preprint: Inspired by error-correcting code (ECC) memory, researchers introduce the 'check token' (<CHECK>) to provide built-in self-verification for large language model reasoning streams. Analogous to parity bits that carry no data but safeguard integrity at a ~12.5% overhead, the check token carries no reasoning content and is discarded after triggering, operating at roughly 13% overhead. When outputted by the model, it initiates self-checking, error localization, truncation, and rewriting. The authors implement speculative forking to remove false-positive latency and use segmented rewards to make trigger timing end-to-end learnable. Experiments on Qwen3-32B and Qwen3-Next models show a +10.8 percentage point improvement on the HMMT25 benchmark at 1.13x overhead, yielding significantly higher token efficiency than Best-of-8 sampling. As a preprint, code availability and broader model generalization remain to be independently verified.
View selection rationale
Importance 83/100
About this brief
- Content type
- Research summary
- Source type
- First-party / official source
- Original source
- arXiv Machine Learning ↗
- Source published
- Process
- AI-assisted discovery, translation, summarisation, classification and scoring; standards and corrections are maintained by the editorial desk.
- Verification
- The original URL and publication metadata are retained. This brief is not independently fact-checked unless explicitly stated.
- Published by
- Polyemergence Editorial Desk · PolyEmergence Pte Ltd.
- Conflict disclosure
- No financial or commercial conflict declared.