Ariaska RL
Multi-agent reinforcement learning research
An experimental research platform by founder Filip Volf: five agents with distinct roles learn through reinforcement learning in isolated, simulated environments, with a language model as an optional mentor that can be switched off.
- Status
- Experimental · core recovery merged 17 Oct 2026 · partial research readiness
- Role
- The founder’s research project, published as source-available under a non-commercial licence. Not a DEUS client product.
- Technology
- Python
- PyTorch
- pytest
- GitHub Actions
- Optional local or cloud LLM mentor
01Research question
How far can a team of specialised learning agents improve through reinforcement learning when a language model is only an optional teacher — and how do you keep that research reproducible?
02Architecture
Five roles, one decision core.
How the learning system is organised, as documented in the repository.
Optional mentor
A local or cloud language model. Its demonstrations become behavioural-cloning signals, and it can be switched off entirely.
Five role policies
Each role learns its own PPO actor-critic policy.
Auxiliary learners
SAC, double DQN and curiosity-driven exploration around the main policies.
Decision core
Weighted arbitration over the candidates the roles propose.
Isolated simulated environment
Research runs in isolated, simulated environments; the CI suite runs offline with network access and tool execution denied.
What was built
- Five role policies, each a PPO actor-critic, with a weighted decision core arbitrating between candidate actions
- Auxiliary learners around the main policies: SAC, double DQN and curiosity-driven exploration (RND)
- An optional mentor — a local or cloud language model — whose demonstrations become behavioural-cloning signals, disabled with a single setting
- Offline CI on GitHub Actions with network and tool execution denied, and checkpoints that restore all five role policies with optimiser and random-number state
03Engineering highlights
Reproducible by default
The final local suite reports 3,318 passing tests (6 skipped, 0 failed); a synthetic evaluation over three seeds reran byte-identically.
Learns without the mentor
With the mentor disabled, all five roles completed two policy updates each with zero language-model requests.
Honest results
On the synthetic benchmark the trained policies reach 96.18% mean accuracy against 48.8–50.5% for baselines (±3.58 pp across seeds) — a result the repository states does not establish general learning ability.
Figures as reported in the repository’s own validation documents.
04Not yet established
What this project does not claim.
Stated plainly, so research is never mistaken for a finished product.
- Live capability and production readiness have not been independently demonstrated
- GPU, live-provider and production behaviour are not verified
- Exact restoration of every experimental controller is still in progress
More work
- Production product
Legal-practice SaaS for law offices in Serbia
Next step
Have a system worth building?
Tell us what you are trying to build and how the work is done today. We will tell you how we would approach it.