Comparing a Hesitation-Word Logit Penalty with presence_penalty in Qwen3.5-0.8B Thinking Mode: A Short Measurement Note

The Qwen3.5-0.8B model card notes that, at the recommended settings, the model is prone to thinking loops in thinking mode even with a presence penalty (presence_penalty) of 1.5 [3]. In this note, we compared a constant logit penalty applied to hesitation words (W3; λ = 3) with the card-recommended presence_penalty = 1.5 baseline (P). We used 200 questions from the GSM8K test set, 2 seeds, and a 4096-token thinking budget. When the budget filled up before ` ` was produced, ` ` was force-inserted, after which an answer of at most 512 tokens was generated. The primary hypothesis H1′ (an increase of at least +10 points (percentage points) in natural close), registered before the full run but after the speed test, was not supported: the difference was +5.25 points, with a paired-bootstrap (resampling) 95% confidence interval (CI) of 2.5–8.0. The non-inferiority hypothesis H2′ (a loss of at most −3 points in forced-closing accuracy) was met: W3 − P = +15.5 points (95% CI 10.5–20.75). That the difference is greater than zero does not rest on a preregistered superiority test; it is exploratory. The answer hits the 512-token limit in 70.25% of cases under P and 48% under W3; accuracy on completed answers is similar (89.92% and 87.98%). The accuracy difference largely co-occurs with the difference in truncation; this is not a causal claim. In an AI-assisted blind manual review of the 11 outputs flagged by the 30-gram loop detector, only 6 met the "true repetition" criterion. The threshold in the plan is 27/30 for 30 samples; the 10/11 threshold used for the 11 candidates, and the criterion, were written into the deviation log after the run outputs were obtained but before labeling. Because 6/11 fell below the threshold, the decision label became `olcer_gecersiz` (detector invalid); the result is sensitive to how the criterion is read (approximately 3/11–10/11). H1′ and H2′ do not depend on the loop detector; the loop rates should not be interpreted. The protocol was revised after the speed test, the run was resumed after an interruption, and it was carried out in fp16 on T4 GPUs; the results are narrow in scope.