ACLArena: Agent Continue Learning in Multi-stage Post-training
The paper proposes a continual-learning setup for agents that tries to preserve old skills while adding new ones across post-training stages.
ACLArena studies why agents forget or generalize as they move through sequential training, looking at both model-level and token-level behavior. The authors compare multi-teacher on-policy distillation, self-distilled fine-tuning, and model merging as ways to recover earlier capabilities without losing newer ones. Their proposed recipe combines offline replay on high-quality trajectories with multiple RL-specialized LoRA experts routed by task. The experiments cover four reasoning and agentic tasks in both in-domain and out-of-domain settings. HF Daily Papers' note
ACLArena studies why agents forget or generalize as they move through sequential training, looking at both model-level and token-level behavior. The authors compare multi-teacher on-policy distillation, self-distilled fine-tuning, and model merging as ways to recover earlier capabilities without losing newer ones. Their proposed recipe combines offline replay on high-quality trajectories with multiple RL-specialized LoRA experts routed by task. The experiments cover four reasoning and agentic tasks in both in-domain and out-of-domain settings. HF Daily Papers' note
score 5