Tree-Ising Repairs Qwen Models with Sparse Bit-Level Edits
Hadi Salloum and Ammar Ali’s Tree-Ising paper frames language-model repair as sparse discrete optimization over BF16 weights, solved as independent QUBO problems with GPU-based annealing or simulated bifurcation. The authors report 54.6% on MATH-500 for repaired Qwen3.5-0.8B layers versus 51.7% for a GPT-6 Astra agent edit, with only four forward passes and zero LLM tokens. On Qwen3.5-2B, the recipe reportedly raises greedy GSM8K accuracy from 21.6% to 69.4%, although no independent replication is supplied.
If independently validated, the method could offer a substantially cheaper alternative to LLM-agent weight editing, using discrete optimization instead of large token budgets and repeated forward passes.
Computer Science > Machine Learning
arXiv:2611.14210 (cs)
[Submitted on 10 Nov 2026]
Title:Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
Authors:Hadi Salloum, Ammar Ali
View a PDF of the paper titled Tree-Ising: Repairing Language Models with Bit-Level Ising Edits, by Hadi Salloum and 1 other authors