DeepMind names the problem everyone building with reasoning models has hit: "reasoning collapse"
The same model can ace a benchmark question and fail an equivalent one stated differently — DeepMind's new training method cuts that brittleness by roughly 40%.
DeepMind published research this month formally naming a failure mode most practitioners have encountered informally: 'reasoning collapse,' where a model solves a problem correctly, then fails on a logically identical version simply because it's phrased differently or wrapped in irrelevant context.
The paper proposes a training methodology — built around exposing models to systematically reworded variants of the same underlying problem during fine-tuning — that reduces this brittleness by approximately 40% in DeepMind's internal evaluations. It's a meaningfully different target than the raw-capability chase most model releases optimize for.
It fits a broader pattern researchers are pointing to across August's papers: the field's center of gravity is shifting from 'can the model do this at all' toward 'can it do this reliably, under rephrasing, at scale' — a much harder and much more commercially relevant bar.