Found first: a primary source the press has not covered yet.
A new paper by Shan Yang tests 20 released reinforcement-learning models across a 1.5B-to-32B parameter range and finds that most of the apparent capability gains from outcome RL are a token-budget artifact. At 4k and 8k token budgets, 82% of pass@1 improvements come from finishing, defined as answering before the budget runs out. The paper is at arxiv.org/abs/2610.34894.
What the source says
Yang evaluates each RL model against the base model it was trained from, separating gains into two components: finishing (answering before the token budget expires) and reach (problems RL solves that the base model failed on all eight attempts, minus problems the base model solved that RL failed on all eight). Across four physics and math datasets, reach adds at most 2.1 percentage points on any single set. Of 144 physics problems that only RL models solved, 3.5% held up under resampling or human review. Nine of ten novel reasoning steps a human confirmed in RL outputs also appeared in base-model samples. The author's institutional affiliation is not listed on the arxiv page.
Why it matters
The RL-scaling narrative rests on the claim that outcome RL unlocks problems beyond the base model's reach. This study puts a number on how much of that is real across 20 publicly released models: at most 2.1 percentage points on physics and math. The 82% finishing figure means most benchmark gains disappear if you give the base model more tokens to complete its answer. The physics-problem collapse, 144 apparent RL-only solutions reduced to 3.5% on review, suggests that exclusive-solution counts overstate genuine capability gain. The finding that novel RL reasoning steps recur in base-model samples narrows the case for RL-specific reasoning emergence further still.