Seeing is Deciding: Evidence-Based Parallel Probing for LLM-based Web Agents

Abstract

Autonomous agents powered by Large Language Models (LLMs) can automate complex web tasks, but their reliability is limited by uncertain action outcomes: the consequence of interacting with a web element is often known only after execution. Existing sequential methods are vulnerable to error propagation after wrong decisions, while tree-search approaches rely on speculative pre-execution scoring and incur substantial exploration cost. We introduce WebTactix, an evidence-based parallel probing framework that shifts web navigation from speculative prediction to grounded decision making. Rather than scoring actions before execution, WebTactix executes multiple candidate plans in parallel, observes their successor pages, and compares the resulting evidence to select the next branch while accounting for future continuation. This design enables the agent to explore multiple alternatives without committing prematurely to uncertain decisions or incurring the overhead of extensive tree search. Experiments on WebArena and Online-Mind2Web demonstrate consistent improvements in both task success and efficiency, with WebTactix achieving 76.8% success on WebArena while reducing average completion time by 56.2% relative to the strongest tree-search baseline.