Anonymous

Anonymous release for the ICLR 2027 submission MemGUI-RL: Reinforcement Learning for Proactive Context Management in Long-Horizon Mobile GUI Agents. Project page: https://memgui-rl-anonymous.github.io/ MemGUI-8B-RL is MemGUI-8B-SFT (Qwen3-VL-8B-Instruct supervised on MemGUI-3K) post-trained for 100 optimizer steps with FARPO (Folding-Aware Reward-decoupled Policy Optimization, span-to-step ratio rho = 9). The policy speaks the ConAct (Context-as-Action) interface of MemGUI-Agent: every response contains a folding directive for its own history, an optional memory operation and the next GUI action. The checkpoint is a standard Qwen3VLForConditionalGeneration model (weights in bf16, ~17.5 GB).…

Evaluation logs of every model row reported in the ICLR 2027 submission MemGUI-RL (project page: https://memgui-rl-anonymous.github.io/). One zip archive per evaluation session: memguibench/.zip - per-task ConAct trajectories (traj.json, three attempts where applicable), MemGUI-Eval judge outputs (results.csv, per-attempt judgments, failure labels, information-retention analyses) and run metadata. Screenshots are omitted for size; the paper's case-study tasks are included under casestudies/ with down-scaled screenshots. mobileworld/.zip - per-task trajectories and the MobileWorld GUI-only evaluation report. manifest.json maps every archive to the row of the paper it supports.…

RL prompt corpus of the ICLR 2027 submission MemGUI-RL (project page: https://memgui-rl-anonymous.github.io/). Every record is one annotated state of a MemGUI-3K trajectory in the ConAct conversation format consumed by the FARPO trainer (https://github.com/memgui-rl-anonymous/MemGUI-RL): Each record has taskid, stepnumber, ispositive, badstep, rawresponse (the annotated ConAct response:,,,, ), conversations (system prompt, user prompt with the folded history / UI memory / recent step record and the screenshot, reference assistant turn) and metadata (impact / reasonableness labels of the step, trajectory ids). Screenshots are not duplicated here: image references have the form…