ARC Improves AI Agent Fairness in Open-Ended Interactions

In open-ended real-world interactions, AI agents can exhibit diverse valid behaviors, such as asking for clarification or providing progress updates. This flexibility, however, can complicate standard reinforcement learning (RL) evaluations, as rollouts within a group may not be truly comparable. This can lead to reward models favoring certain interaction styles over context-appropriate ones, a problem termed "reward fairness."To address this, the paper proposes ARC (Advantage Regularization via Conditioning), a novel training recipe. ARC restores fairer relative comparison by grouping rollouts based on the agent's strategy, combined with hybrid rewards and entropy regularization. This approach is studied within a new paradigm called `inter`, which decouples user-visible communication from an agent's internal reasoning and tool use. `inter` also provides `inter-86K`, a strategy-annotated training corpus.Empirical results demonstrate that ARC significantly strengthens core tool-use benchmarks. Additionally, the `inter` paradigm dramatically reduces the time-to-first-token compared to baseline approaches. These findings suggest that ensuring fair behavioral comparison is a critical bottleneck in open-ended interactive learning, as important as the reward mechanism itself.