ARCHER: From Rationale Imitation to Executable Reasoning in Small Language Models
ARCHER reports a 5.51-percentage-point accuracy gain over the strongest tool-augmented baseline across five models under 10 billion parameters.
The framework converts free-form reasoning into structured traces with calculator calls, runtime observations, verification actions and final answers. A second training stage uses reinforcement learning to reward correctness, tool grounding and adherence to the reasoning process. Reported gains include 8.95 points on MATH-500 and 5.33 and 5.34 points on AIME-2024 and AIME-2025, respectively. Tests on two additional reasoning-specialized 7B models also found benefits from executable grounding.
ArXiv · AI/CL/LG's note
The framework converts free-form reasoning into structured traces with calculator calls, runtime observations, verification actions and final answers. A second training stage uses reinforcement learning to reward correctness, tool grounding and adherence to the reasoning process. Reported gains include 8.95 points on MATH-500 and 5.33 and 5.34 points on AIME-2024 and AIME-2025, respectively. Tests on two additional reasoning-specialized 7B models also found benefits from executable grounding.
ArXiv · AI/CL/LG's note
score 5