Hadi Salloum, Ammar Ali · Weight-Space LLM Repair
Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
Tree-Ising treats repairing an LLM as a sparse discrete optimisation over BF16 weights, using sign flips and exponent shifts solved as QUBO (Ising) problems on a GPU via simulated annealing or simulated bifurcation. On the two layers edited, it reaches 54.6% on MATH-500, against 51.7% for the edit made by a prior frontier-agent method and 7.9% for the base model. The method needs four forward-pass requests and no LLM tokens, and the paper also reports a proof-of-concept run on an IBM Heron quantum processor.
Open the source ↗Similar Research & Safety releases
- Weighted State Souping for Efficient Long-Context Retrieval
Apple Machine Learning Research (with Harvard University co-authors) · Weighted SSM Souping · Week 47, 2026
- Evaluating the Evaluators: Investigating LLM Judges for Personalized Writing Style Assessment
Apple Machine Learning Research (Natalie Mackraz, Andrew Silva, Aini Putkonen, Brihi Joshi, Joelle Alcaidinho, Arnav Arora, Barry-John Theobald, Katherine Metcalf) · LLM Judge Study · Week 47, 2026
- WikiTrace: Evidence-Linked Memory for LLM Agents
Yukyeong Lee, Yumin Choi, Jinheon Baek, Sung Ju Hwang (KAIST AI) · Evidence-Linked Memory · Week 47, 2026
- On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models
Apple Machine Learning Research (with Harvard University authors) · Gating in SSMs · Week 47, 2026
- Coachable health age and SensorWM
Google Research · Wearable health world model · Week 47, 2026
- TOME-Math: Scaling Mathematical Proof Supervision from Structured Textbooks
Intern Large Models (Youqing Fang et al.) · Math Proof Dataset · Week 47, 2026
Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.