Found first: a primary source the press has not covered yet.
A paper accepted at NeurIPS 2026 shows that zeroing fewer than 0.05% of a language model's weights, with no other modification, matches full fine-tuning across multiple post-training regimes. The method, Deletion Tuning, is described in "The Surprising Effectiveness of Deleting Weights in LLM Post-Training" by Jack Lu, Zhenbang Yang, Mike Lasby, Tejas Pote, Yani Ioannou, and Mengye Ren.
What the source says
Deletion Tuning learns a binary mask over the gate projection of transformer blocks and zeros the selected weights, leaving everything else unchanged. It matches full fine-tuning (FFT) performance on reinforcement learning with verifiable rewards (RLVR), supervised fine-tuning (SFT), and on-policy distillation. The resulting masks are hundreds to thousands of times smaller than a corresponding FFT checkpoint. Under reinforcement learning, prior-task accuracy stays near the base model level with Deletion Tuning, while FFT, LoRA, and random sparse fine-tuning lose up to 5.4 accuracy points.
Why it matters
Standard fine-tuning modifies a large fraction of model parameters; this paper shows a binary mask over fewer than 0.05% of weights can produce the same downstream performance. The more consequential finding is the continual learning result. Models deployed across multiple tasks typically degrade on earlier tasks as new fine-tuning runs are added; under RL, that degradation is near zero for Deletion Tuning and up to 5.4 accuracy points for FFT, LoRA, and random sparse fine-tuning. For teams that maintain fine-tuned variants of a base model across tasks, that gap is the number to examine.