FM Evaluation Pipeline

[biosynthart]

• CLIP (ViT-B/32) and DINOv2 for embedding rendered simulation frames
• Three ASAL search modes:
  • Supervised target — "find parameters matching these ecological prompts" (e.g., "thriving meadow" → "overgrazing" → "rain recovery")
  • Open-endedness — maximize trajectory novelty in FM embedding space over long rollouts
  • Illumination — discover maximally diverse set of ecosystem configurations
• CMA-ES optimization (gradient-free, handles 600-dim search spaces)
• Physical plausibility constraints (square-cube law, thermal homeostasis limits, trophic sanity)

[biosynthart]

Update: CLIP ViT-B/32 evaluator shipped (search/lila_search/evaluator.py). Illumination search (diversity-driven GA) shipped and working. First run completed successfully.

Remaining work:

• DINOv2 embedding support
• CMA-ES target search toward text prompts
• Open-ended search (temporal novelty in CLIP space)

[biosynthart]

⏸️ Shelved — FM evaluation pipeline (target search, open-ended search) deferred until after the distributed engine scalability work. Track A (illumination search) is complete and stable.