Repairing language models with bit-level Ising edits

The problem

Small reasoning language models frequently fail by falling into short cycles of repeated text when they reach their token budget, a problem known as reasoning loops (arxiv.org). For instance, Qwen3.5-0.8B hits this loop on 93% of MATH-500 problems under a 32,768-token cap, limiting its accuracy to 7.9% (arxiv.org). Recent MIT work by Takehi et al. introduced weight-space programming, where a frontier LLM agent reads network weights and writes a repair (arxiv.org). Their GPT-6 Astra agent corrected these loops by modifying 64 columns in the output projections of layers 15 and 19 (arxiv.org). While successful, the agent-based approach is extremely expensive, requiring 158 agent turns, 89.7 million tokens, 50 forward-pass requests, and 15.6 hours of compute for a single edit, with no mathematical guarantee of finding an optimal solution (arxiv.org, arxiv.org).

The idea

The authors propose Tree-Ising, a method that casts model repair as a sparse discrete optimization problem over the model's native BF16 weights (arxiv.org). The method operates in five steps: representing weights as state trees, selecting target layers through a single forward scan, defining an objective function based on forward statistics, formulating the problem as independent QUBO (quadratic unconstrained binary optimization) or Ising models, and solving them using a GPU (arxiv.org). The objective is designed to weaken loop directions, strengthen concluding directions, and preserve output on productive tokens while restricting the total number of modified weights (arxiv.org). Allowed edits are sign flips and exponent shifts that preserve shared group exponents and mantissas (arxiv.org). The authors solve these row-level problems simultaneously using simulated annealing or a quantum-inspired simulated bifurcation solver (arxiv.org).

The results

Evaluated on the two layers targeted by the agent, Tree-Ising achieves 54.6% accuracy on MATH-500, outperforming the agent's edit of 51.7% under the same evaluation protocol, whereas the base model scores 7.9% (arxiv.org). A simulated bifurcation solver yields a statistically indistinguishable repair at 55.1% (arxiv.org). Both solvers outperform the agent edit across all five math benchmarks, with statistically significant improvements on MATH-500 and GSM8K after correcting for multiple comparisons (arxiv.org). Tree-Ising modifies 13% to 17% fewer weights and reduces BF16 bit flips by a factor of 2.2 to 2.3 compared to the agent edit (arxiv.org). Furthermore, the method requires only four forward-pass requests and zero LLM tokens, with the annealing recipe costing about $0.90 in GPU time and executing in 31 minutes (arxiv.org, arxiv.org). When extended to Qwen3.5-2B using layers and doses selected solely from the base model, the recipe increases greedy GSM8K accuracy from 21.6% to 69.4% (arxiv.org). As a proof of concept, the authors also executed an eight-variable repair subproblem using QAOA on an IBM Heron processor (arxiv.org).

Why it matters

This work demonstrates that expensive, natural-language-driven LLM agents are not strictly necessary to program neural network weights. By framing weight repair as a mathematical optimization problem over discrete state trees, models can be repaired faster, cheaper, and with greater precision. The approach opens up new pathways for utilizing quantum-inspired and quantum hardware solvers to fix faulty model behaviors directly in weight space.

Caveats

The authors note that their cost and comparison scope has limitations: while the agent incurred a high cost to discover failure modes and select layers from scratch, their recipe reused the agent's selected layers and calibrated its dose based on the agent's edit for Qwen3.5-0.8B (arxiv.org). Thus, the cost comparison contrasts an exhaustive discovery process with the application of a fixed, optimized recipe (arxiv.org). Additionally, this document is a preprint available on arXiv and has not yet undergone formal peer review.