ACLArena: Agent Continue Learning in Multi-stage Post-training

ACLArena studies why agents forget earlier skills when post-training proceeds through multiple stages: mathematical reasoning, search, e-commerce tool use, and instruction following. Using Qwen3-8B-Base, the benchmark compares sequential training with multi-teacher mixed on-policy distillation, self-distilled fine-tuning, and model merging, connecting the analysis to consolidation recipes used in Qwen3, DeepSeek-V4, and Microsoft’s MAI-Thinking-1. Sequential updates move parameters in only partially aligned task directions; token analysis shows that forgetting concentrates at high-entropy decision points rather than uniformly across generations. The effect is severe: AIME26 falls from 25.83 after math training to 6.04 after e-commerce training, while search performance also regresses. The proposed solution, Mixture of Low-Rank Experts, first uses filtered expert trajectories for SDFT, then freezes the shared backbone and trains a separate LoRA adapter with reinforcement learning for each domain, using environment context for routing. Across the four-stage pipeline, MLE reaches 49.7 on NQ, 38.6 on multi-hop search, and 42.4 on GPQA, while achieving 85.0 on IF-Eval. The study also uses DAPO-Math-17K and finds that SFT establishes valid tool-use behavior through larger parameter updates, whereas RL and mixed OPD make smaller, policy-local refinements. Overall, ACLArena argues that continual agent learning should share transferable behavior but isolate conflicting task-specific adaptations instead of forcing every capability into one parameter space.