Procedure without Judgment: What Small Language Models Learn under an Exact Verifier
A 4B model reached 92% accuracy on its training policy, but its learned checking routine failed to generalize.
The task was to identify the smallest set of facts explaining why an authorization policy blocked a tool call, using an exact verifier. Training on 2,000 algorithm-generated traces beat frontier-model-generated training data in all three seeds. On unseen policies, the student’s explanations remained sufficient in 94% of cases but were minimal in only 48%; Haiku 4.5, prompted to shrink its answers, found the exact reason in all 47 episodes. Reinforcement learning with exact rewards did not improve held-out accuracy.
ArXiv · AI/CL/LG's note
The task was to identify the smallest set of facts explaining why an authorization policy blocked a tool call, using an exact verifier. Training on 2,000 algorithm-generated traces beat frontier-model-generated training data in all three seeds. On unseen policies, the student’s explanations remained sufficient in 94% of cases but were minimal in only 48%; Haiku 4.5, prompted to shrink its answers, found the exact reason in all 47 episodes. Reinforcement learning with exact rewards did not improve held-out accuracy.
ArXiv · AI/CL/LG's note
score 4