[arXiv]score: 0.10

PEEU Method Improves Small MLLM Planning for GUI Web Agents via Hindsight Experience

June 26, 2026

PEEU autonomously explores GUI environments to collect planning experiences and synthesizes training data using hindsight, targeting weak planning in small open-source MLLMs without relying on commercial models. A hierarchical analysis framework (TDHAF) breaks down generalization across low, mid, and high task granularities to diagnose compositional failures.

HOW THIS AFFECTS YOU

●

builderIf you're deploying small MLLMs for web automation, PEEU's autonomous data synthesis approach could reduce dependence on GPT-4-class models for training data generation.

●

researcherTDHAF's three-level compositional generalization analysis provides a structured diagnostic lens for evaluating GUI agent planning beyond aggregate task success rates.

read original ↗arxiv.org

← back to feed