WebATLAS: An LLM Agent with Experience-Driven Memory and Action Simulation

TL;DR

WebATLAS introduces an inference-time, memory-augmented LLM web agent that uses a multi-layer memory (Working Memory, Cognitive Map, Semantic Memory) and look-ahead action simulation within an actor–critic planning loop. By building a cognitive map through curiosity-driven exploration and simulating action trajectories before execution, it achieves robust, goal-directed navigation across unseen websites without website-specific fine-tuning. Ablation studies confirm the complementary roles of memory, planning, and look-ahead in improving performance. On WebArena-Lite, WebATLAS sets a new state-of-the-art at 63% success, signaling a practical, scalable path toward reliable long-horizon web automation using modular, inference-time components.

Abstract

Large Language Model (LLM) web agents often struggle with long-horizon web navigation and web task completion in new websites, producing inefficient action sequences unless fine-tuned on environment-specific data. We show that experience-driven memory, combined with look-ahead action simulation, is sufficient for LLM agents to adapt to unseen web environments by remembering past failures and predicting the consequences of future actions. We introduce WebATLAS (Actor-Critic Task-completion with Look-ahead Action Simulation), a memory-augmented LLM web agent that learns a lightweight internal model of the environment from interaction experience and performs hypothetical action rollouts before acting in the real world. WebATLAS builds a persistent cognitive map via curiosity-driven exploration, stores interaction outcomes as experience-based memory, and evaluates candidate actions in cognitive space using a planner--simulator--critic loop. This enables the agent to reuse past experience, avoid previously unsuccessful behaviors, and generate more efficient plans. We evaluate WebATLAS on the WebArena-Lite benchmark for autonomous web navigation and demonstrate a success rate of 63%, outperforming the previous state-of-the-art at 53.9%. Unlike previous systems, our modular architecture requires no website-specific LLM fine-tuning. Ablation studies confirm that experience-driven memory, look-ahead action simulation, and hierarchical replanning play complementary roles in enabling robust, training-free web agents.

Figures (1)

• Figure 1: Architecture of WebATLAS. (a) Overall flow of WebATLAS: The raw observation $o_t$ is summarized to lower cognitive load. Then the planner makes a plan $P_t$ based on the summarized observation $o'_t$. The actor proposes $N$ possible candidate actions for next step. The critic provides judgment of action candidates and finalizes the best action $a_t$ to take by considering action outcomes obtained from the cognitive map. (b) Memory construction with curiosity-driven exploration: We build cognitive map by employing exploratory lightweight agents to interact with the environment. (c) Look-ahead Action Simulation (LAS): At each step, WebATLAS simulates all candidate actions with the observation from the cognitive map, providing ability to look-ahead. We employ the memory agent to learn from LAS trajectories to make a better plan and update memory if necessary.