Tree-Ising preprint reports cheaper weight repairs that improve math accuracy

Researchers report in a preprint that Tree-Ising repairs language model weights with four forward-pass requests and zero LLM tokens. On Qwen3.5-0.8B, the method scored 54.6% on MATH-500, compared with 51.7% for an earlier agent-generated repair under the same evaluation protocol. Tree-Ising turns selected changes to stored weights into separate optimization problems solved on a GPU. Its simulated annealing recipe cost about $0.90 in GPU time and changed 13% to 17% fewer weights than the agent repair. On Qwen3.5-2B, the same recipe increased greedy GSM8K accuracy from 21.6% to 69.4%. The demonstrated repairs concern selected layers in these two models.