TechCapsules

AI-summarised brief · reviewed before publication

Researchers from QDeep and ITMO University introduced Tree‑Ising, a method that repairs repetitive reasoning in small language models by editing selected internal weights. The technique, framed as a mathematical optimization problem, ran on conventional GPUs, achieving 55.1% accuracy on the MATH‑500 benchmark for a Qwen3.5‑0.8B model and boosting GSM8K performance from 21.6% to 69.4% on a Qwen3.5‑2B model under specific conditions. A proof‑of‑concept eight‑variable repair was executed on an IBM quantum processor using QAOA, though no quantum advantage was shown.