Check token paper gives LLMs a <CHECK> marker to verify reasoning mid-stream

An arXiv preprint titled "Check Token: Real-Time Self-Verification and Precise Truncation in LLM Reasoning" starts from an analogy with ECC memory. ECC memory stores 8 parity bits for every 64 data bits and automatically detects and corrects errors on each read. The parity bits hold no data and only protect integrity, at about 12.5% overhead.

The authors say the reasoning chains of large language models have no such built-in safeguard. Once an error appears, it propagates along the chain, and existing methods can only verify externally after generation has finished.

Their proposal is the check token. A functional marker, , is added to the model's vocabulary. The model can output it at any position in its own generation, and doing so triggers self-checking: analysis, localization, truncation and rewriting of the reasoning so far. The token carries no reasoning content and is discarded after it triggers the check. Like ECC parity bits, it only safeguards correctness, and the authors put its overhead at about 13%. They claim this establishes built-in self-verification for language model generation streams for the first time.

Two further mechanisms are described. Speculative forking is said to eliminate the latency cost of false positives, meaning checks that fire when nothing was wrong. Segmented reward is said to make the timing of the trigger learnable end to end, so the model learns when to emit .

The experiments use Qwen3-32B and Qwen3-Next. The authors report that the check token achieves a +10.8 percentage point improvement on HMMT25 at 1.13x overhead, with token efficiency 88x that of Best-of-8. They also report that the advantage of precise truncation grows monotonically with chain length, so longer reasoning chains benefit more.

Key facts

  • The paper adds a special token to the vocabulary; the model emits it at any position to trigger analysis, localization, truncation and rewriting of its own reasoning.
  • The token carries no reasoning content and is discarded after triggering; the authors compare it to ECC parity bits (8 per 64 data bits, about 12.5% overhead) and cite about 13% overhead for their method.
  • Reported result on Qwen3-32B / Qwen3-Next: +10.8 pp on HMMT25 at 1.13x overhead, with token efficiency 88x that of Best-of-8.
  • Speculative forking is said to remove false-positive latency, and segmented reward is said to make trigger timing learnable end to end.
  • The authors report that the precise truncation advantage increases monotonically with chain length.

Why it matters

Errors in a long reasoning chain tend to carry forward, and the authors argue that existing verification only happens externally once generation is done. A model that can flag and cut its own faulty reasoning while it writes would catch mistakes earlier. The paper's pitch is that this costs little: about 13% overhead, set against a Best-of-8 baseline that the authors say is 88x less token-efficient. The claim of a first built-in self-verification mechanism for generation streams is the authors' own.

Who it affects

Mainly researchers and engineers working on reasoning models, reinforcement learning for reasoning, and inference-time compute. Teams that currently spend tokens on sampling several answers, as in Best-of-8, are the natural audience for a cheaper alternative. The results are reported on Qwen3-32B and Qwen3-Next.

How to use it

Nothing here is a product or a download. The recipe the abstract describes is to add a token to the vocabulary, train the model to emit it where its reasoning may have gone wrong, use speculative forking to avoid latency from false alarms, and use a segmented reward so the trigger timing is learned. No code or model release is mentioned in the source.

How solid is it

This is a single arXiv preprint, and the account here rests on its abstract. The headline figure is +10.8 pp on HMMT25 at 1.13x overhead. The source does not say which of Qwen3-32B or Qwen3-Next that figure refers to, nor the baseline accuracy on HMMT25. It gives no results on benchmarks other than HMMT25, and it does not say whether the paper is peer reviewed. No author names or institutions are given.

Risks and caveats

The 88x token efficiency is stated relative to Best-of-8 only, and the meaning of the 1.13x overhead (tokens, compute or latency) is not defined beyond the word overhead. The ECC comparison is the authors' framing, not evidence in itself. The claims about speculative forking, segmented reward and the growing truncation advantage are the authors' experimental findings and have not been independently checked here. Treat the numbers as provisional until the full paper and any replication are available.

“We propose the check token, establishing built-in self-verification for language model generation streams for the first time”

— From the paper's abstract