A new fine-tuning technique squeezes better math performance out of existing language models by teaching them to ignore what they already know.
Researchers describe the method, called Trimmed Logit-Gap SFT (TrimSFT), in a preprint posted September 11, 2026 to arXiv under the ID 2609.09707. The paper has not been peer-reviewed. Standard supervised fine-tuning grades every token in a training example the same way, whether the model already nails that token or is still guessing. TrimSFT instead measures the gap between the model's top two guesses for each token and downweights both extremes: tokens the model has already mastered and tokens it is too unsure about to learn from cleanly. The authors tested it on six base models from the Llama, Qwen, and DeepMath families across five math benchmarks, reporting gains as large as 26.9 points on MATH500 over standard fine-tuning.
The interesting part is not the raw score bump. It is the diagnosis: a lot of fine-tuning effort may be wasted sharpening tokens a model already gets right, while the tokens that actually need work get drowned out. It is also a cheap fix, since it needs no reference model or extra forward pass, which is why labs building reasoning models will likely test it fast.
Big gains on a single benchmark family from one uncorroborated paper are worth treating with the usual skepticism until someone outside the author list reproduces them.