A method for repairing flawed AI reasoning drafts reported the strongest average score in a six-benchmark comparison, and it had the best score on five of the six individual tests. Called DCGC, it recorded an average accuracy score of 24.8. Its reported scores were 44.9 on GSM8K, 22.3 on MATH, 10.7 on MBPP, 13.1 on HumanEval, 35.7 on MMLU-STEM and 22.5 on MMLU-Pro.
A second pass over a first attempt
DCGC is designed for a model that starts with an imperfect answer from an upstream solver. It uses a masked diffusion model, which builds a response through repeated denoising, and asks the model to consider the problem alone as well as the problem together with the draft.
The key control, called Dynamic Dual-CFG, keeps separate branches for problem-only and joint problem-draft conditioning. It scales the draft-related adjustment according to the relative confidence gap, meaning the difference in confidence between those branches. When the joint branch offers no confidence gain, the method does not amplify that draft signal.
Training mixed standard solving pairs with refinement triplets: a problem, an imperfect draft and a refinement target. It covered three domains, targeted approximately 10,000 unique problems per domain and produced 55,440 training samples plus 2,913 validation samples.
The proposed guidance used a fixed alpha setting of 0.5 and a fixed beta setting of 1.0, calibrated only on the GSM8K validation split and then held constant across benchmarks. Constant-scaling baselines used a sweep from 0.5 to 2.0 in steps of 0.5.
Where the scores moved
The evaluation used separate hard sets with 216 examples from GSM8K, 282 from MATH, 216 from MBPP, 69 from HumanEval, 956 from MMLU-STEM and 6,675 from MMLU-Pro.
The standard-sampling comparison showed a large gap between models without and with supervised fine-tuning. Without SFT, the model had an average accuracy of 4.7 and scored 7.4% on GSM8K. With SFT, the corresponding figures were 18.9 and 32.4%, alongside 30.9% on MMLU-STEM and 19.2% on MMLU-Pro.
DCGC also outscored the simpler guidance variants in the reported average. Its score was 24.8, compared with 22.4 for Single-CFG using the problem branch, 20.1 for Single-CFG using the joint branch, 20.2 for static Dual-CFG and 21.5 for independent Dual-CFG. The paper reports that DCGC had the best score among the LLaDA variants across all six benchmarks.
The draft had to be relevant
Tests of draft relevance found that the original aligned draft performed better than altered versions. On MATH, accuracy was 22.3 with the original draft, compared with 10.3 when drafts were shuffled and 11.0 when they came from a different domain. The original condition also had the highest draft-reuse ratio.
The researchers also tried a gold-agnostic setting, meaning the system was not told whether a draft was right before deciding how to handle it. On the full benchmark sets, DCGC reported accuracy of 43.20 on MATH, 86.96 on GSM8K and 69.17 on MMLU-STEM. On the selected correction subsets, the scores were 23.5, 39.8 and 33.8. Relative to five-sample majority voting, the reported changes were +2.01, +0.16 and +0.51 for the full sets, and +3.1, +1.7 and +7.4 for the correction subsets.
Within the correction subsets, DCGC had the highest reported hard-subset accuracy on all three benchmarks: 31.51 on GSM8K, 18.11 on MATH and 31.45 on MMLU-STEM. Its easy-subset scores were 53.33, 44.62 and 40.35, respectively.
What the confidence signal can and cannot do
An analysis of token-level confidence offers a clue about the gate's behavior. On MATH, the mean confidence gap was 0.214 at positions where the two branches disagreed, compared with 0.026 where they agreed. Draft reuse was 0.321 at disagreement positions and 0.192 at agreement positions. The association between gate scale and branch-disagreement probability was modest, with a Spearman correlation of 0.167 and a 95% confidence interval from 0.150 to 0.182.
The method retained reported performance when the initial drafts came from different upstream solvers. On hard sets generated from Mistral-7B-v1, it scored 42.0% on GSM8K. On Qwen-2.5-32B hard sets, it scored 28.8% on GSM8K and 18.6% on MATH, versus 12.1% and 0.6% for Qwen self-refinement.
The authors also tested DREAM, a separate diffusion backbone, and describe the result as supporting evidence rather than proof of universal transfer. Standard sampling averaged 4.7 without SFT and 17.4 after SFT, while Dynamic Dual-CFG without SFT averaged 15.0. DCGC's reported scores were 42.6 on GSM8K, 18.1 on MATH, 13.4 on MBPP and 31.4 on MMLU-STEM.
The document is an arXiv preprint labeled arXiv:2608.25428v2 and dated 14 Oct 2026. Its reported results describe performance in the tested model and benchmark settings, not a guarantee that every flawed reasoning trace can be repaired.
Paper data and sources
Original title: DCGC: Draft-Conditioned Global Correction for Complex Reasoning with Masked Diffusion Models
Authors: Minhae Oh, Nakyung Lee, Jungwoo Lee
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text