The model had the answer. It couldn’t stop thinking. An independent audit of MIT CSAIL's "Sparks of Weight-Space Programming" result found that most of the reported MATH-500 accuracy gain for Qwen3.5-0.8B came from fixing a looping failure, not from the weight edit itself. The base model scored 8.2% with a 32,768-token budget in the rerun, with 92.4% of samples exhausting the budget, and only 152 of 2,000 samples finished on their own — all 152 correct; a simple loop detector flagged 1,696 of the 1,848 budget-exhausted samples, with typical loops starting about 6,600 tokens in and burning another 26,000 tokens on repeated lines. Cutting the thinking short and writing the first few words of the answer scored 49.8% without touching weights, close to the 48.1% GPT-6 Astra reported after changing 131,072 weights (0.016% of the model), and the authors' own report attributed 576 of 579 rescued practice samples to running out of budget. The model had the answer. It couldn’t stop thinking. https://julesbelveze.github.io/the-model-had-the-answer/ This is Qwen3.5-0.8B, about 4,000 tokens on an algebra problem. $45 + 2 = 47$. $47 - 5 = 42$. $42 + 5 = 47$. Okay, I am confident in 47. Wait, I should check if there is any hidden trick. Maybe the expression is $a \star b = 9a + 2b - ab + 5$. Wait, let me check if the term is $-ab$ or $+ab$. ... Wait, let me check if the term is $-ab$ or $+ab$. Wait, let me check if the term is $-ab$ or $+ab$. And then it writes that last line again and again. For 28,000 more tokens, until the budget runs out. Since it never gives an answer it actually gets marked wrong. Even though it had 47 the whole time. Why care about one small model stuck on one algebra problem? Because of a result you may have seen two weeks ago. A group at MIT CSAIL published Sparks of Weight-Space Programming https://github.com/RikiyaT/lm understand-logs . They gave GPT-6 Astra direct access to this “little” model's weights and let it experiment for 50 rounds. Astra ended up changing 131,072 weights 0.016% of the model , and MATH-500 accuracy jumped from 7.9% to 48.1%. Spoiler alert: most of the 40 points between 7.9% and 48.1% are caused by a stopping problem. The model often has its answer within a few thousand tokens and then can't end its own thinking. If you end it yourself and write the first few words of the answer for it, it scores 49.8%, without touching weights. That's about where the edit lands. The trick only works because every answer here goes in a “box”, but we'll get to that. The first three rows are all cut from the same 2,000 traces 500 problems, 4 samples each . The last row is the authors' edited checkpoint, run through the same pipeline. I get 51.8% there, a bit above the 48.1% they report. A bit of context if you haven't used these models. In thinking mode the model writes a scratchpad first, then emits a special