Are action-based preferences vital? One of many key elements of ACT is that the contrastive pairs spotlight variations between conversational actions. In “ACT w/ Random Actions”, we moreover look at the significance of motion choice by randomly sampling each the successful and dropping motion when developing the desire pair, and observe this underperforms regular ACT.
Do we want on-policy sampling? In “ACT w/o on-policy sampling”, we look at the significance of on-policy sampling by evaluating regular off-policy DPO on the dataset as constructed in Section 1. Whereas we do observe some enhancements over SFT (e.g., from 69.0 to 74.8 Macro F1), the general enhancements are a lot bigger when utilizing on-policy sampling as with full ACT. This can be resulting from the truth that the off-policy damaging responses usually are not assured to lie within the language manifold of the coverage mannequin, and distribution shift could also be too troublesome to beat with off-policy studying.
Is trajectory simulation vital? ACT is better-aligned with multi-turn conversations resulting from its trajectory simulation. With out multi-turn simulation, our method could be seen equally to on-policy DPO variants like IRPO, however with a conversation-specific reward sign which accounts for dialog actions and activity heuristics. In “ACT w/ sampling w/o simulation”, we discover that this trajectory-level simulation is crucial to enhancing multi-turn efficiency, particularly the coverage mannequin’s means to cause about its personal clarification questions.
Is ACT mannequin agnostic? The bottom mannequin in our fundamental experiments, Zephyr, is obtained by aligning Mistral. In “ACT with unaligned basis fashions” we observe a efficiency hole of 6.5 Motion F1 and 4.3 Trajectory F1 after ACT tuning for the 2 fashions. Nevertheless, our outcomes exhibit ACT can enhance efficiency no matter pre-existing alignment with human suggestions, though it could possibly assist as an improved mannequin initialization. Total, we discover that enhancing base mannequin efficiency with ACT is mannequin agnostic.
