WebATLAS: Actor-Critic Task-Completion with Look-ahead Action Simulation

ACL ARR 2026 May Submission14906 Authors

Keywords: LLM Agent, Web Agent, Planning, Tree Search, Cognitive Map

Abstract: Current web agents lack the ability to anticipate action outcomes, leading to inefficient plans that fail to account for the structure and dynamics of unfamiliar environments. We introduce \textbf{\method} (\textbf{A}ctor-Critic \textbf{T}ask-completion with \textbf{L}ook-ahead \textbf{A}ction \textbf{S}imulation for \textbf{Web} Applications), a memory-augmented web agent that grounds its planning in environment reality by simulating action consequences in \emph{cognitive space} before execution. WebATLAS first constructs a multi-layered agentic memory, comprising a \emph{cognitive map} that encodes action transition dynamics, and an \emph{episodic memory} that captures procedural experience, through lightweight curiosity-driven exploration. During execution, \emph{Look-Ahead Action Simulation} (LAS) leverages this memory to simulate action outcomes over candidate actions using real, previously observed transitions rather than LLM-hallucinated outcomes. A critic evaluates the resulting trajectories, selects the most promising action sequence, and triggers dynamic replanning when observed states diverge from predictions. WebATLAS achieves state-of-the-art results on three web agent benchmarks (WebArena-Lite, WebChoreArena, WebVoyager) without any website-specific LLM fine-tuning. Ablations confirm that multi-layered memory, look-ahead simulation, and replanning serve complementary roles.

Paper Type: Long

Research Area: LLM agents

Research Area Keywords: planning in agents; agent memory; web agents

Contribution Types: NLP engineering experiment

Languages Studied: English

EMNLP 2026 AI Reviewing Experiment: yes

Submission Number: 14906