More tokens in an AI reasoning chain does not mean better answers — what's inside them does.
Researchers tested chain-of-thought prompting across 25 language models by pairing shorter and longer reasoning traces that followed the same logical plan, so neither was artificially rewritten. The extra tokens produced no meaningful accuracy gain on their own. When gains did appear, they tracked back to validation and checking steps buried in the longer text — not to length itself. A separate controlled experiment confirmed this: two traces covering identical facts and operations produced different outcomes only when the verbose version contained higher-quality reasoning prose, not simply more of it.
This matters because the field has been arguing about why chain-of-thought works — whether it buys useful computation by stretching the token sequence, or whether it forces the model to articulate useful intermediate steps. The answer, per this work, is neither cleanly. Length alone is not the mechanism, but neither is semantic content a full explanation — the quality and structure of that content appears to be the lever.
The practical implication cuts against a tempting shortcut: you cannot make a model reason better just by padding its scratchpad. Researchers also found that replacing verbose traces with length-matched nonsense filler recovered none of the accuracy gains, which rules out the pure compute account. For anyone building prompting pipelines or evaluating reasoning models, the signal here is that what you ask a model to write out matters more than how much you ask it to write — a finding that will be easy to ignore and expensive to misapply.