Tree-Ising Repairs Qwen3.5-0.8B for $0.90, 350x Under Agent

Two researchers claim they repaired a language model's reasoning loops for ninety cents.

The paper, posted to Hugging Face on November 11, is from Hadi Salloum at QDeep in Riyadh and Ammar Ali at ITMO University's LEAD Lab for Efficient AI and Data. It compares their method, Tree-Ising, against an earlier repair of Qwen3.5-0.8B driven by the GPT-6 Astra agent. The abstract is flat about the baseline: "The prior GPT-6 Astra agent corrected reasoning loops in Qwen3.5-0.8B using 158 agent turns, 89.7M tokens, and 50 forward-pass requests." Tree-Ising used four forward passes, no LLM tokens, finished in 31 minutes, and cost $0.90 against the agent's $312.

On MATH-500@4, Qwen3.5-0.8B scored 7.9% out of the box. The agent's edits pushed it to 51.7%. Tree-Ising's simulated-annealing variant landed at 54.6%, and the simulated-bifurcation variant at 55.1%. GSM8K@1 went from 13.8% to 70.8%. The authors report a +2.9±1.0 point advantage over the agent baseline on MATH-500, statistically significant at p≈0.003.

The method "recasts repair as a sparse discrete optimization problem over BF16 weights," decomposing two targeted MLP layers into row-level QUBO/Ising problems and solving each row independently. Moves are restricted to sign flips and exponent shifts inside BF16 groups, so shared exponents and mantissas are never rewritten. It is the latest entry in a crowded corner on our fine-tuning tracker, which ran a Shanghai AI Lab weight-space-programming paper last week.

Transferred to Qwen3.5-2B with no reference edit, the same recipe automatically selected layers 15 and 19 and took GSM8K from 21.6% to 69.4%, modifying 292,845 weights, or 1.2% of the targeted matrices. Of the test problems, 669 were solved only by the edited model, 38 only by the base, 247 by both.

The paper lists its own caveats: 100 of the 500 MATH-500 problems were seen during development, the recipe was formulated retrospectively from the agent's logs, and the quantum demonstration is a single 8-variable subproblem on IBM Heron with no downstream accuracy measurement.