Figure 1 - available via license: Creative Commons Attribution-ShareAlike 4.0 International
Content may be subject to copyright.
Online adaptive agent that induces and reuses programmatic skills as actions (bottom), as opposed to adding textual skills in memory (top).
Source publication
To succeed in common digital tasks such as web navigation, agents must carry out a variety of specialized tasks such as searching for products or planning a travel route. To tackle these tasks, agents can bootstrap themselves by learning task-specific skills online through interaction with the web environment. In this work, we demonstrate that prog...
Contexts in source publication
Context 1
... agent then induces higher-level skills (e.g., search product(name)) that wrap primitive actions or prior skills as executable programs, accompanied with corresponding test trajectories to verify their quality. Verified skills are then incorporated into the agent action space and can be directly called to solve future tasks with similar procedures, as depicted in Figure 1 (bottom). We first evaluate ASI on the WebArena benchmark ( Zhou et al., 2024a) ( §3) and demonstrate that our online, adaptive ASI surpasses its static agent baseline without adaptive components by 23.5% in success rate. ...
Context 2
... first evaluate ASI on the WebArena benchmark ( Zhou et al., 2024a) ( §3) and demonstrate that our online, adaptive ASI surpasses its static agent baseline without adaptive components by 23.5% in success rate. To validate the advantage of using programmatic representations for skills, we further compare to an adaptive agent, AWM ( , that represents skills in memory as non-executable texts (Figure 1 top); we find ASI scores 11.3% higher success rate by employing verifiable, programmatic skills (Figure 1 bottom). ...
Context 3
... first evaluate ASI on the WebArena benchmark ( Zhou et al., 2024a) ( §3) and demonstrate that our online, adaptive ASI surpasses its static agent baseline without adaptive components by 23.5% in success rate. To validate the advantage of using programmatic representations for skills, we further compare to an adaptive agent, AWM ( , that represents skills in memory as non-executable texts (Figure 1 top); we find ASI scores 11.3% higher success rate by employing verifiable, programmatic skills (Figure 1 bottom). ...
Context 4
... the scope of this work, we focus on language model (LM) based agents, where each agent policy consists of an LM backbone L, a memory M, and a skill library A, as illustrated in Figure 1 top and bottom. In the implementation, the memory M and the skill library A are provided as input context to the LM backbone. ...
Context 5
... AWM which represents skills in free-form text representations and can only augment agent memory via M t ∪ {d t } → M t+1 (Figure 1 top), we introduce ASI that represents skills as executable python programs, and directly integrate skills into the agent action space instead, via A t ∪ {d t } → A t+1 (Figure 1 bottom). ...
Context 6
... AWM which represents skills in free-form text representations and can only augment agent memory via M t ∪ {d t } → M t+1 (Figure 1 top), we introduce ASI that represents skills as executable python programs, and directly integrate skills into the agent action space instead, via A t ∪ {d t } → A t+1 (Figure 1 bottom). ...
