glossary
Terms and phrases
Every term is defined the way the research note that uses it does, tied to a verified claim where one exists. 10 terms from 1 notes.
All terms
AXTree
The accessibility tree of a web page, which Region4Web reorganizes into functional regions through hierarchical decomposition and semantic abstraction.
canonical sub-instructions
Standardized smaller steps into which WebAgent decomposes the original task instruction before execution begins.
end-to-end success
A measure of whether an agent completes a multi-step web task from start to finish, which offers limited insight into where failures arise because it does not separate planning, execution, or replanning errors.
flat-planning baseline
A single agent that receives the task instruction and produces actions directly, without an explicit decomposition step, used as a reference point to measure the effect of hierarchical decomposition.
global planning agent
In CoAct, the agent responsible for high-level planning, which collaborates with a local execution agent to complete tasks.
hierarchical decomposition
A method of structuring the planning layer of an LLM-based web agent by splitting a task, a role, an observation, or a process into parts, as opposed to flat planning where a single agent handles the entire task without such splitting.
high-level planning
One of three core capabilities of LLM-based web agents, referring to deciding what to do at a strategic level, as distinguished from low-level execution and replanning.
local execution agent
In CoAct, the agent responsible for carrying out the steps determined by the global planning agent.
low-level execution
One of three core capabilities of LLM-based web agents, referring to carrying out individual steps correctly, identified as the dominant bottleneck in current LLM web agents.
replanning
One of three core capabilities of LLM-based web agents, referring to noticing that something went wrong and adjusting the plan accordingly.