Nathan reformulated weight editing for repetitive thinking in Qwen3.5-0.8B as a linear program. With directions and layers fixed from prior experiments, optimization replaced roughly eight hours of search with about a minute of solving and changed about eight times fewer weights for similar evaluation results. Like the preceding work, it aims to remove a repetitive bottleneck that obscures existing capability. This is a single-model result on selected benchmarks, not proof of broadly improved reasoning or acquisition of new skills.
Back to Tech
Linear programming speeds up a weight-editing search
The experiment reduces repetitive reasoning in one model; it does not demonstrate learning new knowledge.