PATS: Policy-Aware Training Scaffolding for Agentic Reinforcement Learning
PATS treats skills as temporary training scaffolds, then removes that guidance at deployment.
The paper says weak long-horizon LLM agents often generate repetitive failed rollouts, limiting reinforcement learning signal. PATS turns recent rollout groups into evidence cards, uses task-specific evaluation to adjust future rollout context, and pares back redundant guidance as the policy improves. The policy is still trained with environmental rewards through standard RLVR. The authors report gains of up to 18.6% on ALFWorld and WebShop, plus competitive results on seven search-augmented QA benchmarks with 32.1% fewer prompt tokens than the baseline. ArXiv · AI/CL/LG's note
The paper says weak long-horizon LLM agents often generate repetitive failed rollouts, limiting reinforcement learning signal. PATS turns recent rollout groups into evidence cards, uses task-specific evaluation to adjust future rollout context, and pares back redundant guidance as the policy improves. The policy is still trained with environmental rewards through standard RLVR. The authors report gains of up to 18.6% on ALFWorld and WebShop, plus competitive results on seven search-augmented QA benchmarks with 32.1% fewer prompt tokens than the baseline. ArXiv · AI/CL/LG's note
score 5