README.md

metadata

license: apache-2.0
tags:
  - mobile-gui-agent
  - agent-memory
  - benchmark-results
  - memgui-bench
  - mobileworld

MemGUI-RL-Eval

Evaluation logs of every model row reported in the ICLR 2027 submission FARPO: Folding-Aware Reward-Decoupled Policy Optimization for Long-Horizon Mobile GUI Agents (project MemGUI-RL; project page: https://memgui-rl-anonymous.github.io/). The release has two tiers.

Tier 1 - compact bundles (about 110 MB in total)

• memgui_bench/<session>.zip - per-task ConAct trajectories (traj.json, three attempts where applicable), MemGUI-Eval judge outputs (_memgui_eval/results.csv, per-attempt judgments, failure labels, information-retention analyses), metrics_summary.json and run metadata; screenshots and raw model requests omitted.
• mobileworld/<session>.zip - per-task trajectories and the MobileWorld GUI-only evaluation report.

Tier 2 - complete sessions (684 GB in total)

Everything the evaluation harness wrote, including every device screenshot, the per-step request sent to the policy (prompt, folded history, memory state, response), the MemGUI-Eval judge inputs and outputs, thread logs and the results.csv ledgers.

• memgui_bench_full/<session>/<session>.tasks_AAA-BBB.zip - all files of tasks AAA..BBB (top-level task directory, _attempt_trajs/<task> for attempts 2-3, _memgui_eval/<task> judge files; _backup_<timestamp> directories are the harness's copies of interrupted or re-run tasks).
• memgui_bench_full/<session>/<session>.session_files.zip - everything that is not a task directory (run metadata, eval reports, results.csv, thread logs). For the two sessions that were archived from a local copy (session-memgui-v26050803, memgui-8b-26062202) this archive also carries the _attempt_trajs and _memgui_eval trees, which is why it is large.
• mobileworld_full/<session>/<session>.zip - the complete MobileWorld session.

All archives are stored uncompressed (screenshots do not compress); unzip them into one directory per session to rebuild the original layout, e.g.

huggingface-cli download memgui-rl/MemGUI-RL-Eval --repo-type dataset \
    --include "memgui_bench_full/memgui_8b_on_mg_26072401/*" --local-dir ./eval
cd eval/memgui_bench_full/memgui_8b_on_mg_26072401 && for z in *.zip; do unzip -q -o "$z"; done

Paper row	Benchmark	Session	Full archives	Full size (GB)
MemGUI-8B-SFT (MemGUI-Agent release; RL initialisation)	MemGUI-Bench	session-memgui-v26050803	9	46.1
MemGUI-8B-SFT (MemGUI-Agent release; RL initialisation)	MobileWorld	memgui_8b_26050801	1	2.2
FARPO rho=9 = MemGUI-8B-RL (128 tasks: -check1 is the session reported in the paper, 26072401 the earlier session used for the case studies; 26062901 = 40-task development subset)	MemGUI-Bench	memgui_8b_on_mg_26072401-check1	10	83.9
FARPO rho=9 = MemGUI-8B-RL (128 tasks: -check1 is the session reported in the paper, 26072401 the earlier session used for the case studies; 26062901 = 40-task development subset)	MemGUI-Bench	memgui_8b_on_mg_26072401	25	91.3
FARPO rho=9 = MemGUI-8B-RL (128 tasks: -check1 is the session reported in the paper, 26072401 the earlier session used for the case studies; 26062901 = 40-task development subset)	MemGUI-Bench	memgui_8b_on_mg_26062901	4	13.1
FARPO rho=9 = MemGUI-8B-RL (128 tasks: -check1 is the session reported in the paper, 26072401 the earlier session used for the case studies; 26062901 = 40-task development subset)	MobileWorld	memgui_on_mw_8b_26062801	1	2.4
natural corpus, all folds (sweep)	MemGUI-Bench	memgui-8b-26062202	6	29.9
natural corpus, all folds (sweep)	MobileWorld	memgui_8b_26062302	1	2.1
span folds only (sweep)	MemGUI-Bench	memgui-8b-26062201	4	13.0
span folds only (sweep)	MobileWorld	memgui_8b_26062301	1	2.4
rho=4 (sweep)	MemGUI-Bench	memgui_8b_on_mg_26070401	4	13.9
rho=4 (sweep)	MobileWorld	memgui_on_mw_8b_26070301	1	2.2
rho=3 (sweep)	MemGUI-Bench	memgui_8b_on_mg_26070602	4	10.8
rho=3 (sweep)	MobileWorld	memgui_on_mw_8b_26070602	1	2.5
rho=2 (sweep)	MemGUI-Bench	memgui_8b_on_mg_26070702	4	13.4
rho=2 (sweep)	MobileWorld	memgui_on_mw_8b_26070603	1	2.4
rho=1/4, reverts to the natural ratio (sweep)	MemGUI-Bench	memgui_8b_on_mg_26070701	4	14.2
rho=1/4, reverts to the natural ratio (sweep)	MobileWorld	memgui_on_mw_8b_26070601	1	2.8
direct RL from Qwen3-VL-8B-Instruct (rho=9)	MemGUI-Bench	memgui_8b_on_mg_26090601	9	86.9
direct RL from Qwen3-VL-8B-Instruct (rho=9)	MobileWorld	memgui_on_mw_8b_26090601	1	2.2
ablation step 1: GRPO recipe (scalar estimator, natural sampling; -check1 is the session reported in the paper, 26090701 the earlier session used for the case studies)	MemGUI-Bench	memgui_8b_on_mg_26090701-check1	9	77.6
ablation step 1: GRPO recipe (scalar estimator, natural sampling; -check1 is the session reported in the paper, 26090701 the earlier session used for the case studies)	MemGUI-Bench	memgui_8b_on_mg_26090701	36	96.5
ablation step 1: GRPO recipe (scalar estimator, natural sampling; -check1 is the session reported in the paper, 26090701 the earlier session used for the case studies)	MobileWorld	memgui_on_mw_8b_26090701	1	2.9
ablation step 2: + reward decoupling (decoupled estimator, natural sampling; session -check2 is the one reported in the paper)	MemGUI-Bench	memgui_8b_on_mg_26090901-check2	10	66.0
ablation step 2: + reward decoupling (decoupled estimator, natural sampling; session -check2 is the one reported in the paper)	MobileWorld	memgui_on_mw_8b_26090901-check1	1	2.6

manifest.json maps every archive (compact and full) to the row of the paper it supports. Serving-endpoint identifiers, judge model ids, API keys and local paths are redacted in all text files; screenshots are unmodified.