Are action-based preferences mandatory? One of many key components of ACT is that the contrastive pairs spotlight variations between conversational actions. In “ACT w/ Random Actions”, we moreover study the significance of motion choice by randomly sampling each the successful and dropping motion when establishing the choice pair, and observe this underperforms regular ACT.
Do we’d like on-policy sampling? In “ACT w/o on-policy sampling”, we study the significance of on-policy sampling by evaluating regular off-policy DPO on the dataset as constructed in Part 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 on account of the truth that the off-policy destructive responses aren’t assured to lie within the language manifold of the coverage mannequin, and distribution shift could also be too tough to beat with off-policy studying.
Is trajectory simulation mandatory? ACT is better-aligned with multi-turn conversations on account of its trajectory simulation. With out multi-turn simulation, our strategy might 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 important to bettering multi-turn efficiency, particularly the coverage mannequin’s capacity to motive about its personal clarification questions.
Is ACT mannequin agnostic? The bottom mannequin in our essential 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. Nonetheless, our outcomes exhibit ACT can enhance efficiency no matter pre-existing alignment with human suggestions, though it will probably assist as an improved mannequin initialization. Total, we discover that bettering base mannequin efficiency with ACT is mannequin agnostic.