A new preprint reports that DoubtLess, a decoding method for reasoning models, cut generated reasoning tokens by 31% across seven benchmarks while keeping accuracy within the authors’ reported uncertainty. It uses a small helper model to steer a larger model as it writes, rather than retraining the larger model.
Reasoning models often produce a chain of thought: the text they generate while working through a problem. At each step, DoubtLess asks a 0.6-billion-parameter model to read the prompt and reasoning so far twice, once under an instruction to be brief and once to be thorough. At each step, the helper assigns a score to each possible next token. Tokens it scores higher under the brief instruction get a relative boost in the larger model; tokens it favors more under the thorough instruction lose weight. That difference nudges which token the larger model chooses. The larger model still selects each token; the method neither bans particular words nor forces the reasoning to stop.
The main test paired Qwen3-32B with Qwen3-0.6B on seven math, science and coding benchmarks, including AIME 2024 and 2025, GPQA-Diamond and LiveCodeBench. The paper reports a 31% aggregate reduction in reasoning tokens with no statistically significant accuracy loss. Each problem was run with three random seeds, and the authors report 95% bootstrap intervals, so “unchanged” means the measured difference was not statistically significant, not that every score was identical.
The authors’ analysis found that most of the removed tokens came after the correct answer first appeared in the reasoning trace, where models often repeated checks. They also tested a self-debugging coding agent on 164 HumanEval+ problems. The agent ran tests and had one chance to repair a failed program; with DoubtLess, it used 20% fewer thinking tokens per turn and its pass rate rose by 2 percentage points against ordinary decoding. That result applies to this test setup, not to software agents in general.
The token reduction does not translate directly into the same percentage reduction in serving cost. DoubtLess runs the helper model twice at each reasoning step, adding computation even though the helper is smaller. The paper’s timing comparison covered 14 problems on one NVIDIA B200, a narrow test rather than evidence about production workloads. The study remains a preprint with benchmark evidence, not a deployment report.
The researchers have released code under Apache 2.0. The repository currently documents a Linux and NVIDIA GPU setup using vLLM 0.29.0, so trying it requires compatible model pairs and inference infrastructure. An earlier GTN report covered ConfSFT, which trains models to use confidence signals to shorten reasoning; DoubtLess changes token selection at inference time instead.
