A few weeks after MIT researchers showed GPT-6 Astra repairing a small model by rewriting its weights, a new arXiv preprint argues that the same repair can be done without any language model at all. Tree-Ising frames the fix as a maths problem over the stored bits of the weights and hands it to a solver. If the result holds, weight editing may turn out to be less about an agent's insight and more about asking the right optimisation question.
The target is a known failure in Qwen3.5-0.8B. In thinking mode it falls into short repeating cycles on 93% of MATH-500 problems under a 32,768-token cap, which leaves it at 7.9% accuracy. The Astra agent fixed this by tracing the loops to the MLP layers 15 and 19 and rewriting 64 columns in each. According to the agent's public logs, that took 158 turns, 89.7 million tokens, 50 forward-pass requests and 15.6 hours, which the authors cost at $312.
Tree-Ising allows only two kinds of change to each weight: a sign flip or a shift by a power of two, so the shared exponents in BF16 storage stay untouched. That turns every candidate change into a yes-or-no choice, and the objective splits exactly into one Ising problem per output row, which a GPU solves with simulated annealing. On the same two layers it reaches 54.6% on MATH-500 against 51.7% for the agent's edit under the same evaluation, and a second solver, simulated bifurcation, gets 55.1%. It changes 13% fewer weights, needs four forward-pass requests and zero LLM tokens, and runs in about 31 minutes for roughly $0.90. On the larger Qwen3.5-2B, with no reference edit to copy, GSM8K accuracy rises from 21.6% to 69.4%.
For readers following our GPT-6 Astra review, this is a useful correction to the story. The agent's achievement was finding the failure, the direction and the layers; Tree-Ising was designed afterwards by reading the agent's logs, reuses its two layers on the smaller model and calibrates the dose against its edit. The cost comparison therefore sets a full discovery run against the application of a known recipe, and the authors say so.
The limits are wide. The study covers one failure mode, two models from one family and maths benchmarks only; side effects on general language ability were not measured. The base score is taken from the MIT report rather than rerun, each benchmark used a single seed, and only the MATH-500 and GSM8K gains survive the strictest correction for multiple comparisons. It is an unreviewed preprint, and a cheap, auditable repair is worth more once someone shows it works on a model you would actually deploy.