Reducing a reasoning model’s output is only useful if accuracy holds up—or improves. In the reported tests, a targeted edit to nine weight matrices cut thinking tokens while raising scores on two benchmarks; a same-size random edit did neither, making the direction of the change central to the result.
Key Capabilities
- The edit is derived from the model’s own activations and applied without gradient updates or additional training data, offering a distinct route to model behavior changes.
- Against the unmodified base, the report measures a 4.3-point gain on GPQA Diamond and a 13.0-point gain on LiveCodeBench v6, alongside roughly 29–30% fewer thinking tokens on those benchmarks.
- IFBench accuracy rises by 3.3 points, but its confidence interval includes zero; the reported result there is a 54.1% reduction in thinking tokens without a measured accuracy regression.
- A norm-matched random edit shows no clear accuracy change or token savings. The published benchmark figures use Q8_0 weights; other offered quantizations have not been separately evaluated.
Who It Suits—and Trade-offs
A good fit for researchers studying reasoning efficiency, weight editing, or controlled model evaluations who can assess an experimental checkpoint for their own tasks. Look elsewhere if you need independently established performance across many benchmarks or assurance that the reported effect persists in BF16, FP8, or 4-bit deployment: those conditions remain untested.