ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation

ECHOES learns recursive textual task memory through privileged screenshot-history distillation and GRPO, then predicts GUI actions from the current screenshot and memory.

AI-assisted editorial review against a primary source. This is a catalog annotation, not an independent reproduction or endorsement.

What is this resource?

Resource type

Research papers

Environment

android

Topics

memory, interaction-history, state-tracking, long-horizon, imitation-learning, reinforcement-learning, efficiency, evaluation

Record publication date

2026-10-14

Evidence retrieved

2026-10-20T13:12:54.814500Z

Resource ID

arxiv:2610.22408

Recorded paper version

arXiv v1

Version updated: 2026-10-14T14:39:41Z.

This identifies the retained arXiv metadata snapshot; newly observed changes require separate review.

Authors

Tengda Guo, Jiawang Hu, Yifei Tao, Yin Wang, Haotian Hu, Fanyi Wang, Liujuan Tang

Cite this paper

Download BibTeX · Download CSL JSON

Citations preserve the selected source edition, its authors and recorded date precision. Unrecorded venues and identifiers are omitted.

Explore related topics

Scope and limitations

Privileged screenshot history is used during training, while deployment retains self-generated textual state. GUI results use action/step metrics and model-specific history settings; online Sokoban transfer is separate. GRPO rewards action correctness, valid output format and compression length. Diagnostic memory replay is distinct from the validation gap used to schedule training. No independent reproduction.

Primary evidence and verification

Source retrieval is separate from experiment reproduction. Use the original publication or publisher page when citing claims about the resource.

1. arxiv api

   Retrieved 2026-10-20T12:58:20.217517Z. Supports: name, authors, published_at, updated_at, identifiers.arxiv, details.arxiv_snapshot.

   Content fingerprint

   f275c47a8b6d349a70e39a9e55eea5a09d99d5f35cc896100f3bcfd5b0923e0b
2. editorial primary review

   Retrieved 2026-10-20T13:12:54.814500Z. Supports: name, kind, summary, environments, topics, scope, notes.

   Content fingerprint

   0389d96a8f6c4975129f5c526e8df9ca4690f31364a2f4e66aaea4a75f0eabbc

Suggest a correction

Open a source-backed correction. Include this record ID and the original source. Read our editorial policy and coverage boundaries.