ACLArena is a framework for studying Agent Continual Learning — integrating new agent capabilities across sequential post-training stages, an area with no well-established recipe. Using a sequential training pipeline, the authors analyze the mechanisms of forgetting and generalization at both the model level and the token level. They then systematically compare multi-teacher on-policy distillation, self-distilled fine-tuning, and model merging as strategies for recovering previously learned capabilities while preserving newly acquired ones.
