{"slug": "echoes-seeing-beyond-the-current-image-via-recursive-history-compression-and-on", "title": "ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation", "summary": "A team of researchers posted arXiv paper 2610.22408v1, introducing ECHOES, a method that lets long-horizon GUI agents compress visual history into a recursively updated textual state instead of retaining past screenshots. ECHOES uses a Compression-Think-Action format and a Perception-Memory Fidelity Gap-guided on-policy distillation reinforcement learning framework, in which the same model given privileged multi-image inputs acts as a teacher over student rollouts. The authors report substantial gains in task performance and memory fidelity across four GUI benchmarks and their new ECHOESBMK diagnostic benchmark, with competitive performance and a favorable performance-inference-cost trade-off without keeping historical screenshots at deployment.", "body_md": "arXiv:2610.22408v1 Announce Type: new \nAbstract: Long-horizon GUI agents rely on visual history to maintain task states, but retaining historical screenshots increases visual-token usage and inference costs, while textual summaries can discard critical state information, undermining long-horizon reliability. We propose ECHOES, which leverages training-time privileged information to guide textual history compression and strengthen long-horizon reasoning. Specifically, ECHOES adopts a Compression-Think-Action (CTA) format to recursively update a textual history state and reason. The same model, given privileged multi-image inputs, serves as a teacher that provides dense token-level on-policy distillation on student-generated CTA rollouts. Our Perception-Memory Fidelity (PMF) Gap-guided OPD-RL framework dynamically balances privileged distillation and task-reward optimization, with periodic teacher updates supporting continued improvement. We further introduce ECHOESBMK, a fine-grained diagnostic benchmark derived from real GUI agent failure trajectories to jointly evaluate visual perception and cross-modal state retention. Experiments on four GUI benchmarks and ECHOESBMK demonstrate substantial gains in task performance and memory fidelity. Without retaining historical screenshots at deployment, ECHOES achieves competitive performance with a favorable performance-inference-cost trade-off and exhibits positive zero-shot transfer to unseen visual planning tasks.", "url": "https://wpnews.pro/news/echoes-seeing-beyond-the-current-image-via-recursive-history-compression-and-on", "canonical_source": "https://arxiv.org/abs/2610.22408", "published_at": "2026-10-19 04:00:00+00:00", "updated_at": "2026-10-19 04:18:10.782513+00:00", "lang": "en", "topics": ["ai-agents", "ai-research", "computer-vision", "large-language-models"], "entities": ["ECHOES", "ECHOESBMK", "Compression-Think-Action", "Perception-Memory Fidelity Gap", "arXiv"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/echoes-seeing-beyond-the-current-image-via-recursive-history-compression-and-on", "markdown": "https://wpnews.pro/news/echoes-seeing-beyond-the-current-image-via-recursive-history-compression-and-on.md", "text": "https://wpnews.pro/news/echoes-seeing-beyond-the-current-image-via-recursive-history-compression-and-on.txt", "jsonld": "https://wpnews.pro/news/echoes-seeing-beyond-the-current-image-via-recursive-history-compression-and-on.jsonld"}}