AuthorsJiali Cheng, Anjishnu Kumar, Roshan Lal et al.
TL;DR
WebATLAS uses experience-driven cognitive maps plus look-ahead action simulation to reach 63.0% success on WebArena-Lite, +9.1 points over Plan-and-Act.
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
LLM web agents often struggle with long-horizon web navigation and task completion in new websites, producing inefficient action sequences unless fine-tuned on environment-specific data.
Systems like Plan-and-Act achieve only 53.9% success on WebArena-Lite, leaving many realistic multi-step tasks unsolved and risking irreversible actions and dead-end states on complex web interfaces.
HOW IT WORKS
WebATLAS centers on a Planner, Actor, Critic, and Multi-layered Memory that includes Working Memory, a Cognitive Map, and Semantic Memory to guide web actions.
You can think of WebATLAS like a brain with short term RAM, a long term map of the site, and a mental simulator that plays out moves before acting.
This design lets WebATLAS reuse past experiences, simulate multi step futures, and avoid risky actions that a plain context window or reactive agent would miss.
DIAGRAM
This diagram shows how WebATLAS simulates candidate actions using the cognitive map before choosing a real web action.
DIAGRAM
This diagram shows how WebATLAS is evaluated on WebArena Lite and how ablations modify components like the cognitive map and planner.
PROCESS
Planner
The Planner in WebATLAS analyzes the natural language goal and initial observation to produce a structured plan P0 with subtasks and success predicates.
Actor
The Actor in WebATLAS conditions on the current plan, observation, and multi layered memory to propose a small set of executable next step candidates Ct.
Critic
The Critic in WebATLAS uses the cognitive map and semantic memory to simulate outcomes, compute V(a), and select the safest goal advancing action.
Look ahead Action Simulation
Look ahead Action Simulation in WebATLAS rolls out candidate actions over depth D using the cognitive map, updates plans via dynamic replanning, and writes new experience into memory.
KEY CONTRIBUTIONS
Actor critic planner with LLM based look ahead
WebATLAS introduces an actor critic planner where the Actor proposes candidates and the Critic uses look ahead simulation with a value function V(a) to choose safe actions, achieving 63.0% success on WebArena Lite.
Multi layered memory with cognitive map
WebATLAS builds a multi layered memory including Working Memory, a Cognitive Map of transitions, and Semantic Memory of environment rules via curiosity driven exploration and agentic summarization.
Modular architecture without fine tuning
WebATLAS offers a practical modular architecture that integrates planning, memory, and simulation to handle long horizon web tasks without any website specific LLM fine tuning.
RESULTS
Avg w/ Multi-site
63.0%
+9.1 over Plan-and-Act
Avg w/o Multi-site
67.1%
+9.6 over Plan-and-Act
Gitlab
73.3%
+20.0 over AgentOccam
Shopping Admin
77.1%
+22.8 over Plan-and-Act
These metrics come from WebArena Lite, a 165 task benchmark for realistic multi step web navigation across Gitlab, Reddit, shopping, admin, maps, and multi site tasks. The 63.0% average success shows that WebATLAS converts experience driven memory and look ahead simulation into a +9.1 point gain over Plan-and-Act without environment specific training.
BENCHMARK
These metrics come from WebArena Lite, a 165 task benchmark for realistic multi step web navigation across Gitlab, Reddit, shopping, admin, maps, and multi site tasks. The 63.0% average success shows that WebATLAS converts experience driven memory and look ahead simulation into a +9.1 point gain over Plan-and-Act without environment specific training.
BENCHMARK
Average success rate with multi site tasks on WebArena Lite.
BENCHMARK
Average success rate with multi site tasks for WebATLAS ablations on WebArena Lite.
KEY INSIGHT
Adding a direct HTML cognitive map (Base plus CM Raw) to WebATLAS initially reduced average success from 47.9% to 44.8% on WebArena Lite.
This is surprising because richer state storage seems helpful, but it shows that unfiltered raw HTML can overload LLM reasoning and that agentic summarization is crucial.
WHY IT MATTERS
WebATLAS unlocks web agents that can build and reuse cognitive maps, simulate future action sequences, and adapt to new sites without retraining.
Builders can now deploy long horizon web agents that safely explore, remember hazards, and refine plans online, making realistic multi step automation practical across diverse websites.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xiaohui Zhang, Zequn Sun et al.
· 2026
ActMem transforms dialogue history into atomic facts via Memory Fact Extraction, groups them with Fact Clustering, links them through a Memory KG Construction module, and uses Counterfactual-based Retrieval and Reasoning for action-aware answers. On ActMemEval, ActMem reaches 76.52% QA accuracy with DeepSeek-V3, beating LightMem’s 63.97% by 12.55 points and NaiveRAG’s 61.54%.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Answers use this explainer on Memory Papers.
Checking…