The Pulse
DoubtLess Cuts Qwen3 Reasoning Tokens by 31% Without Retraining
Researchers introduce DoubtLess, a decoding method that uses a smaller model to shorten Qwen3-32B’s reasoning traces without changing its weights. In tests across seven benchmarks, the method cut thinking tokens by 31%, while a separate sma
AI.info Team · Written with AI ·
DoubtLess cut Qwen3-32B’s reasoning tokens by 31% across seven benchmarks, while the researchers’ pooled results showed no statistically significant accuracy change. The method, introduced in a paper submitted to arXiv on October 13, 2026, uses a much smaller model during decoding rather than retraining the larger one. In a separate, limited latency test, it also completed problems faster than plain decoding.
A small model steers the next token
DoubtLess asks a smaller model to read the larger model’s reasoning trace twice at every step: once under an instruction to be brief, and once under an instruction to be verbose. It subtracts the resulting next-token scores, scales that difference, and adds it to the large model’s scores. The large model still generates the response; the smaller one shifts which of its candidate tokens become more likely.
The approach is designed to target repeated checks after a model has already produced the correct answer, rather than ban particular words or stop reasoning when a fixed threshold is reached. The paper’s authors describe the result this way: “On seven benchmarks of competition mathematics, graduate-level science and competitive programming, DoubtLess shortens the CoT of Qwen3-32B by 31% with unchanged accuracy, outperforming five inference-time baselines.” Minghao Luo and Liang Chen, “DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning”
Seven benchmarks, with a small accuracy dip
The main evaluation pairs Qwen3-32B with Qwen3-0.6B and covers competition mathematics, graduate-level science, olympiad problems and programming. The seven benchmarks include AIME 2024 and 2025, HMMT 2025, AMC 2023, GPQA-Diamond, OlympiadBench and LiveCodeBench. The authors report 528 problems across the pooled evaluation, with three runs per problem.
Across those tests, mean thinking tokens fell from 10,254 under plain decoding to 7,120 with DoubtLess, a 31% reduction. Mean accuracy moved from 70.8% to 69.9%, a decline of 0.9 percentage points; the paper’s reported 95% interval for that change spans a decline of 2.8 points to an increase of 1.0 point. So the paper’s “unchanged accuracy” describes a result without a statistically significant pooled drop, not identical scores on every benchmark.
The method also carried over to two other model families in tests on AIME 2025 and AMC 2023. DoubtLess cut thinking tokens by 25% for EXAONE-4.0-32B and 40% for Granite-4.2-30B, with no reported accuracy loss in those comparisons. Those results cover a smaller pair of benchmarks than the main evaluation.
Fewer tokens translated into faster test runs
Extra model calls do not automatically mean faster inference. DoubtLess runs the small model twice at each reasoning step, but the authors measured average end-to-end latency of 119.4 seconds per problem, compared with 156.2 seconds for plain decoding. The test used one B200 GPU and 14 problems—two from each benchmark—and the authors estimate the small model’s two forward passes add about 4% to the large model’s floating-point operations.
A self-debugging coding agent provides a second test beyond standalone question answering. On 164 HumanEval+ problems, the agent used 20% fewer thinking tokens per turn and its pass rate rose by 2.0 percentage points against plain decoding. The system writes code, runs tests and gets one repair turn when a test fails, so the result reflects an agent workflow rather than a single model response.
The savings mostly come after an answer appears
In the authors’ analysis, DoubtLess cut reasoning after the correct answer first appeared by about half, while reducing reasoning before that point by about a fifth. The distinction matters: the system does not simply force the model to stop at the first answer. The paper gives an example where a useful algebraic check catches a wrongly excluded root, showing why removing every follow-up check would be risky.
The results are promising but leave a practical boundary: the main comparison centers on one 32-billion-parameter model paired with a 0.6-billion-parameter model, and the cross-family checks cover just two math benchmarks. The latency result also comes from a 14-problem sample on one GPU. The next test is whether DoubtLess can keep useful verification while reducing repetition across broader tasks, model sizes and inference setups.