ResearchBenchmarksKnowledge-intensive reasoning on MMLU-CF first 1,000 samples (test)Knowledge-intensive reasoning on MMLU-CF first 1,000 samples (test)Follow74.2Exact Match AccuracyMGRS69.31270.58171.8573.119Nov 28, 2025Evaluation resultsMethodMethodLinksExact Match AccuracyMGRSBackbone=GPT-4o-miniBackbone=GPT-4o-mini2025.1174.2CoT-SCBackbone=GPT-4o-mini, n=5Backbone=GPT-4o-mini, n=52025.1171.1AoTBackbone=GPT-4o-miniBackbone=GPT-4o-mini2025.1170.9FoTBackbone=GPT-4o-mini, n=8Backbone=GPT-4o-mini, n=82025.1170.6Self-refineBackbone=GPT-4o-miniBackbone=GPT-4o-mini2025.1169.7CoTBackbone=GPT-4o-miniBackbone=GPT-4o-mini2025.1169.6AFlowBackbone=GPT-4o-miniBackbone=GPT-4o-mini2025.1169.5