ACLArena compares ways to add agent skills across post-training stages without forgetting earlier ones

arXiv / HuggingFace Daily Papers·low signalXBlueskyLinkedInCopy link

ACLArena (arXiv 2609.23989, 09-21) builds a sequential agent post-training pipeline and analyzes forgetting at both the model and token level. It then compares multi-teacher on-policy distillation, self-distilled fine-tuning and model merging as ways to recover earlier capabilities. Its proposed recipe combines offline replay of high-quality trajectories with a routed network of RL-specialized LoRA experts, and it is tested on four reasoning and agentic tasks in and out of domain. The abstract gives no headline numbers, so treat it as a map of trade-offs rather than a proven recipe.