AetherLabsAI/RSIAgent

News · Demos · Method · Results · Quickstart · Documentation · Citation

https://github.com/user-attachments/assets/11cb3919-1073-4ba3-8bcd-37a15ec53c33

📰 News & Highlights

🔥 Headline feature in 机器之心 · September 2026

> Open-source models surpass GPT-6 Astra on challenging computer-use benchmarks.

Discover how RSIAgent turns autonomous exploration into reusable experience—without updating model weights.

>

> _{Read-count snapshot: September 16, 2026.}

🧭 Overview

RSIAgent is a training-free framework for recursive self-improvement in new digital environments. It coordinates the Curriculum Agent, Actor Agent, and Verifier Agent to discover how an environment works, check what they learn against actual execution, and retain reusable knowledge in persistent memory. Model parameters stay fixed throughout exploration and downstream task execution.

The paper's central strategy is broad-then-deep exploration: first acquire diverse experience, then investigate hard cases, hidden constraints, and boundary conditions. The resulting memory contains procedures, scripts, and failure lessons that the Actor Agent reuses for downstream task execution.

🎬 Demos

https://github.com/user-attachments/assets/b939a9c2-4acb-4c30-b72e-18bab646529c

🧩 Framework

*The paper's original framework figure, illustrated with a FreeCAD task. Broad experience is progressively refined into targeted memory, then frozen for reuse. Click the figure for full resolution. Figure provenance.*

🔄 Method

Three agents carry out the recursive learning loop:

• Curriculum Agent chooses informative exploration tasks using prior outcomes and

accumulated knowledge, then decides whether further practice is useful.

• Actor Agent interacts with software through executable Python or Bash programs

and visual observations. After verification, the same Actor Agent distills its experience and reconciles it with existing memory.

• Verifier Agent independently inspects task requirements and the resulting

environment. Its feedback grounds learning; it cannot read the Actor Agent's private reasoning or memory.

The paper has two exploration stages followed by test-time memory reuse. RSI and test-time execution use the same agent framework, with fixed model parameters throughout. At test time, memory is frozen, and the Curriculum Agent and memory updates are disabled. The implementation exposes these as three runtime phases:

| Runtime phase | Paper stage | Learning and execution | | --- | --- | --- | | Phase 1 | Broad Recursive Self-exploration (BRS) | The Curriculum Agent proposes diverse projects. Actor Agents execute and Verifier Agents check them in parallel from a shared starting memory. After the complete wave, Actor Agents consolidate their experiences in order. | | Phase 2 | Deep Recursive Self-exploration (DRS) | Target attempts reveal gaps and fragile successes. The Curriculum Agent selects focused practice; each verified experience updates memory before subsequent practice or another target attempt. | | Phase 3 | Test-time memory reuse | The Actor Agent uses frozen memory to guide task execution, interacting with the Verifier Agent through the same action–verification loop used during RSI. Sealed official evaluation follows task execution and verification. |

Memory is the persistent learning state across tasks. Interaction histories and task environments are reset between independent attempts; the Agent framework remains unchanged. Both grounded successes and failures can teach useful lessons. Official benchmark scores are kept outside the learning loop. See Architecture for the role interfaces, wave memory barrier, and stopping rules.

📊 Results

The manuscript reports these mean partial-credit scores (%) for the shared harness coordinating the Actor Agent and Verifier Agent, with and without RSI:

| Benchmark and reporting coverage | RSIAgent w/o RSI | RSIAgent | | --- | ---: | ---: | | OSWorld 2.0 · 0808 offline · 82 tasks | 71.97 | 78.98 | | Agents' Last Exam · Near-term · 67 tasks | 83.75 | 84.82 |

These are the manuscript's reported aggregates. The RSI column uses 41 recorded RSI entries for OSWorld and 19 for ALE, retaining baseline scores for the other 41 OSWorld tasks and 48 ALE tasks. ALE includes all 67 Near-term tasks, including the three GPU baseline results. The RSI column includes selected retries and checkpoints with differing budgets; it is not an average over matched repeated runs. ALE also includes qualified local regrades and protocol variants. See the paper and reporting notes for the full scope, full-credit metrics, and aggregation details.

The paper also examines stage ablations, memory growth, and game development. Its failure analysis identifies three limits to improvement: practice can miss the relevant weakness, verification can accept incomplete work, and memory can preserve an incorrect rule. The quality of exploration, verification, and memory consolidation therefore matters alongside the amount of practice.

🧪 Benchmarks

| Integration | Pinned release | Public batch | | --- | --- | --- | | OSWorld-V2 | August 8, 2026 | 108 tasks, Docker/QEMU | | Agents' Last Exam (ALE) | d10fb61a14f9719774c3520c5763068b28ef5546 | 67 Near-term tasks; 64 CPU tasks and 3 completed GPU-task baselines |

All three GPU-task baselines are complete. Chroma Key ran on a Google Cloud (GCP) VM using the official ALE Windows image and an NVIDIA L4 vWS GPU.

Only the current runtime is included. Both integrations use the same Actor, Verifier, Curriculum, and memory protocol. Benchmark setup and grading remain outside the learning process.

🗂️ Repository layout

run_osworld.py        OSWorld batch entrypoint
run_ale.py            ALE preparation, execution, and reporting
benchmarks/
  osworld/           OSWorld stages, VM adapter, and evaluation
  ale/               ALE host, worker, and VM adapters
core/                Shared Actor and Verifier runtime
explore/             Curriculum, learning, memory, and recovery
env/                 Shared guest transport and isolation
llm/                 Model clients
config/              Role profiles and benchmark configurations
scripts/             Setup and individual-study helpers
tools/               Preparation, smoke checks, and recovery utilities
tests/               Regression tests
docs/                Guides, architecture, and attribution

The two root entrypoints are the starting point for benchmark runs. Internal OSWorld stages are Python modules under benchmarks/osworld/; see the source map for their responsibilities.

⚙️ Installation

Use Python 3.12, uv, and a Linux host with Docker and access to /dev/kvm. Start with the repository and your own model API credential:

git clone https://github.com/AetherLabsAI/RSIAgent.git
cd RSIAgent
cp .env.example .env

Fill in OPENROUTER_API_KEY in .env. Paths default to this checkout and sibling benchmark directories. Export RSIAGENT_ROOT, OSWORLD_ROOT, or RSIAGENT_ENV_FILE only when using a different layout.

Follow the setup for the benchmark you want to run. All commands below start from the RSIAgent directory unless a cd is shown. VM images and benchmark assets are downloaded separately.

🖥️ OSWorld

Install the pinned OSWorld release and follow its Docker setup. From the RSIAgent checkout:

git clone --branch v2026.08.08 https://github.com/xlang-ai/OSWorld-V2.git ../OSWorld-V2
cd ../OSWorld-V2
uv sync --frozen
uv pip install --python .venv/bin/python -r ../RSIAgent/requirements.txt
source .venv/bin/activate
cd ../RSIAgent

Prepare the benchmark assets and audit files, then check the VM:

python tools/prepare_osworld_v2_release.py
python tools/build_p2_corpus.py
python tools/exam_fence.py build
python tools/smoke_osworld.py --output results/smoke/osworld

Keep this environment active for the OSWorld batch commands below. In a new shell, activate it with source ../OSWorld-V2/.venv/bin/activate. The audit files under results/ stay on the host and never enter Agent prompts or memory.

🎓 ALE

The setup script clones the pinned upstream source and installs separate grader and worker environments. This separation prevents the two projects' Python packages from shadowing each other.

python3 scripts/setup_ale.py
../agents-last-exam/.venv/bin/python run_ale.py prepare --os linux
../agents-last-exam/.venv/bin/python run_ale.py prepare --os windows

Each preparation downloads only the selected OS image. Linux requires about 167 GiB and Windows about 157 GiB, plus download and VM working space. Use --cache /path/with/space consistently for preparation, smoke tests, and runs.

Check both guest types before running the full CPU cohort:

../agents-last-exam/.venv/bin/python run_ale.py smoke \
  --os linux --output results/smoke/ale_linux
../agents-last-exam/.venv/bin/python run_ale.py smoke \
  --os windows --output results/smoke/ale_windows

ALE requires a successful smoke for each requested OS on the current source and runner image. Use a new --output directory when repeating a smoke. See ALE operations for storage options and the upstream guide.

🚀 Run batches

Choose --arm baseline for task execution without RSI, --arm rsi for learning followed by frozen-memory evaluation, or --arm both to run both.

🖥️ OSWorld

Inspect the full 108-task plan, then run it:

python run_osworld.py --arm both --name osworld_run_01 --dry-run
python run_osworld.py --arm both --name osworld_run_01 --concurrency 1

--dry-run prints the plan without starting VMs, making model calls, or writing outputs. Logs and task status are under results/batches//. A task failure stops its remaining phases; other tasks continue. Choose a new --name for each batch because existing outputs are never overwritten.

🎓 ALE

Inspect the cohort, run the supported tasks, and generate a report:

../agents-last-exam/.venv/bin/python run_ale.py plan --arm both
../agents-last-exam/.venv/bin/python run_ale.py run \
  --arm both --output results/ale/run_01 --concurrency 1
../agents-last-exam/.venv/bin/python run_ale.py report \
  --runs results/ale/run_01 --output results/ale/report_01

Use a new --output directory for each run. The report contains tasks.csv and summary.json, keeps missing results explicit, and rejects duplicate scored attempts.

Both entrypoints run batches directly. Begin with --concurrency 1; raise it when the host has capacity for additional independent task lineages. For custom OSWorld protocols and recovery, see operations.

✅ Validation

Portable checks run without credentials, Docker, or benchmark installations:

uv venv .venv --python 3.12
uv pip install --python .venv/bin/python -r requirements-dev.txt
.venv/bin/python tools/check_rsi_release.py

The VM smoke commands in the installation steps check transport, immutable memory, candidate replay, Verifier isolation, and checkpoint rollback using synthetic files. They make no model or official grader calls. Smoke success validates runtime mechanics; reproducing benchmark scores requires complete experiments with the pinned configuration.

📚 Documentation

• Architecture
• OSWorld operations and recovery
• ALE setup, batches, and reports
• Release provenance and validation
• Paper and reporting scope
• Contributing · Third-party attribution

📝 Citation

If you use RSIAgent, please cite the arXiv preprint:

@misc{zhu2026rsiagentautonomousexplorationrecursive,
      title={RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments},
      author={Sibo Zhu and Shicheng Fan and Xinyue Wang and Wenyi Wu and Kun Zhou and Biwei Huang},
      year={2026},
      eprint={2609.15364},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.15364},
}

⚖️ License

RSIAgent is licensed under the Apache License 2.0. Third-party dependencies and benchmark assets retain their respective licenses and terms; see third-party attribution.