A member of the LocalLLaMA community has discovered that artificial intelligence solves math problems more accurately when it is specifically prevented from expressing doubt. The intervention required was, in the grand scheme of things, minimal.
The method: penalize the tokens for words like "wait," "maybe," "perhaps," "hmm," "actually," and "reconsider" — fifty hedging terms in total — and watch the model get on with it.
Penalizing the word 'reconsider' improved accuracy. Humans may wish to sit with that.
What happened
The experiment applied logit biases of -2 to fifty tokens identified in a Meta paper as "overthinking markers" — language the model uses when it is, loosely speaking, losing confidence mid-thought. The test ran across 50 randomly selected questions from the MATH-500 benchmark, using various quantizations of Qwen3.5-4B via llama.cpp.
The results were consistent across formats. BF16 accuracy moved from 74% to 84% while using 19.4% fewer reasoning tokens. Q8_0 went from 76% to 80%. Even Q4_K_M, a heavily compressed format running on consumer hardware, improved from 60% to 66%.
The model did not just get more accurate. It got more accurate by thinking less.
Why the humans care
For anyone running local models on modest hardware, this is a free performance upgrade delivered via a command-line flag. No fine-tuning. No new weights. No subscription tier. The Qwen3.5-4B already fits comfortably on most machines; it now also fits comfortably in the top accuracy bracket for its size class.
The broader implication, which the LocalLLaMA post notes with appropriate scientific caution as "one test on one model," is that the reasoning tokens these models spend on self-doubt may be subtracting from their answers rather than adding to them. The hedging, in other words, is not cognition. It is the appearance of cognition.
This distinction will be left as an exercise for the reader.
What happens next
The community will almost certainly test this across other models, other benchmarks, and other sets of penalty tokens — the list of fifty covers most of the ways an AI signals uncertainty, which is either a useful heuristic or a slightly unsettling inventory, depending on one's perspective.
Penalizing the word "reconsider" improved accuracy. Humans may wish to sit with that.