An Oct. 5 arXiv preprint introduces Guidance-TTT, which trains a compact model to suggest strategic code edits while a second model, whose weights remain unchanged, implements them. On a 70-case polyomino-packing benchmark, where solvers place irregular pieces, the team reports 91.89 out of 100. On that benchmark, a control that left the guidance model unchanged while testing the same number of candidate attempts scored 84.85. A previously reported four-agent CORAL system with web access scored 89.40, but the systems and test conditions differ, so the figures are not an equal-budget head-to-head test.
Automated code search often repeats a loop: propose a change, run the program against tests and score it. Training the model that writes complete programs during this loop can be costly. A single verifier score can also obscure whether a poor result came from a weak strategy or faulty code.
Guidance-TTT separates those jobs. In the main setup, Qwen3-8B, a compact guidance model, proposes a high-level change to a saved solution. GLM-5.2 rewrites the candidate program while its weights stay fixed; an executable verifier runs the program against task-specific checks and scores it. That score updates only the guidance model. The 30-round search makes 3,840 candidate attempts on each of the four benchmark tasks.
The paper also reports results on three other code-optimization tasks: Lasso regression-path solving, an AtCoder apple-production problem and a GPU-kernel benchmark.
The approach did not use web access, and each task supplied an executable verifier; these experiments therefore do not test open-ended research. Their trajectory-derived statistics come from one search per configuration, so the paper does not show how consistently the results would recur across runs.
Sources
Image: Figure 3 from Jiang et al., arXiv:2610.06269v1, reused under CC BY 4.0.