21 upvotes
Bit-Level Weight Edits Let LLMs Self-Repair Without Gradients
What is Tree-Ising: Repairing Language Models with Bit-Level Ising Edits?
Tree-Ising lets an autonomous agent read a model’s binary weight matrix and flip individual bits to fix logical errors. It repaired the reasoning loops of Qwen3.5-0.8B in 158 interaction steps, using 89.7 million tokens and only 50 forward passes, all without gradient updates.
Who it is for
Researchers and developers of safety-critical language models can patch faulty reasoning quickly without retraining.
The story so far
Early Signal first spotted Tree-Ising: Repairing Language Models with Bit-Level Ising Edits on November 11, 2026, with 21 upvotes. Early Signal files it under LLMs, Agents.