ARIA AI Scientist Call - Recursive Safeguarding Ltd
This repository demonstrates an AI Scientist system for the ARIA AI Scientist Call (14 November 2025). It is a fork of Stanford's BoxingGym benchmark, extended with:
- LiteLLM multi-provider support for open-weights models
SingleCellMicrofluidicsenvironment (Poisson + stretch modes) for single-cell encapsulation optimization
The system implements the end-to-end scientific loop:
- an LLM agent proposes experiments
- observations from the computational environment are used to update the model's beliefs
- Expected Information Gain (EIG) is used as a metric
- Box's Loop discovers a PyMC statistical model
LiteLLM multi-provider support lets you use any configured model. The demo auto-detects which API keys you have and picks an appropriate model, or you can specify explicitly via environment variables.
Configured providers in conf/llms/:
- DeepSeek (chat, reasoner) - China/open, $0.14-0.55/M tokens
- MiniMax M2 - China/open, $0.30/$1.20/M tokens (free trial)
- GLM-4.6 - Zhipu AI/open
- Qwen3 - Alibaba/open (MLX local inference)
- GPT-5.1 - OpenAI (for comparison)
All configs use environment variables for API keys (no hardcoded secrets).
A new SingleCellMicrofluidics environment for single-cell encapsulation optimization:
Base problem: Poisson regime
- Control variables: cell concentration, flow rate, surfactant
- Goal: Maximize single-cell capture rate (~37% theoretical max)
- Evaluation: EIG regret, parameter recovery
Extended problem (stretch): More physics-informed mode
- Additional physics: flow oscillations (vortex effects), cell crowding, correlated clustering
- Demonstrates agent behavior on harder optimization landscape
One-command execution: bash scripts/run_aria_demo.sh