Alien Biology is a framework for measuring complex, agentic AI reasoning and learning. It constructs procedurally generated biological universes — complete with novel chemistry, multi-compartment organisms, diseases, and diagnostic challenges — that are:
- 1. Untainted — Entirely synthetic "alien" biology avoids training-set contamination. Agents cannot memorize answers; they must reason from scratch.
- 2. Controllable — Parametric generation enables fine-grained difficulty scaling across multiple dimensions (chemistry complexity, information disclosure, diagnostic ambiguity).
- 3. Real-world structured — Multi-level systems (atoms → molecules → reactions → organisms → diseases) mirror the hierarchical complexity of real biological reasoning.
- 4. Generative — Unlimited unique test instances prevent overfitting and enable smooth performance curves rather than binary pass/fail.
Framework Capabilities
- Simulation — Mass-action kinetics with configurable reaction networks, multi-compartment organisms with transport flows, and equilibrium/stability analysis
- Disease & Diagnosis — Perturbation generation, symptom detection, baseline comparison, and multi-candidate diagnostic tasks
- Skinning — Replace all molecule/reaction names with opaque alien terminology at configurable detail levels, forcing agents to reason without training-data shortcuts
- Agent Evaluation — Test suite framework with difficulty scaling, oracle/random/zero baselines, and comparative scoring
- JAX Acceleration — JIT-compiled simulator via XLA for GPU-accelerated large-scale simulations
Example Biological Processes — Executable examples of terrestrial biological processes:
© 2025 Dan Oblinger