ResearchBenchmarksLanguage Understanding on MMLU High tier (Reward and Tool Cost)Follow77.4Reward+RL Free71.57673.08874.676.112Jun 5, 2026Evaluation ResultsMethodMethodLinksRewardTool Cost+RL FreeBackbone=Qwen3-4B, RL...Backbone=Qwen3-4B, RL strategy=Free2026.0677.472+RL GateBackbone=Qwen3-4B, RL...Backbone=Qwen3-4B, RL strategy=Gate, Protocol=confidence-gated GSPO2026.0674.523+RL OTCBackbone=Qwen3-4B, RL...Backbone=Qwen3-4B, RL strategy=OTC2026.0673.80+RL FlatBackbone=Qwen3-4B, RL...Backbone=Qwen3-4B, RL strategy=Flat2026.06730BaselineBackbone=Qwen3-4BBackbone=Qwen3-4B2026.0671.80