Tree-Ising repairs Qwen reasoning loops with sparse BF16 edits
Original title: Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
On two layers it scores 54.6% on MATH-500 versus 51.7% for a GPT-6 Astra edit and 7.9% for the base model.
Tree-Ising treats weight-space repair of language-model reasoning loops as sparse discrete optimization over BF16 weights, solved as independent Ising problems by simulated annealing or quantum-inspired simulated bifurcation. On the two layers edited by a GPT-6 Astra agent for Qwen3.5-0.8B, both solvers beat the agent edit on five math benchmarks while changing fewer weights and using four forward passes and no LLM tokens. The same recipe, with layers chosen from the base model, raises Qwen3.5-2B greedy GSM8K accuracy, and an eight-variable subproblem ran with QAOA on an IBM Heron processor.
Unconfirmed: The 7.9% base MATH-500 score is reported by the authors under their evaluation protocol.
- Simulated bifurcation scores 55.1% on MATH-500, statistically close to annealing.
- Edits change 13% to 17% fewer weights and cut BF16 bit flips by 2.2 to 2.3.
- Annealing costs about $0.90 in GPU time versus 89.7 million agent tokens.
- Qwen3.5-2B greedy GSM8K rises from 21.6% to 69.4%.
- Vendors
- IBM
- Products
- IBM Heron
- Organizations
- MIT
- AI models
- Qwen3.5-0.8BQwen3.5-2BGPT-6 Astra
From the original article
1 min read at huggingface.co
Weight-space programming, introduced in recent MIT work by Takehi et al. , enables a frontier LLM agent to read a network's weights and write a repair. Their GPT-6 Astra agent corrected the reasoning loops of Qwen3.
By Hadi Salloum, Ammar Ali · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.