Mind Evolution
Island-model evolutionary search in natural language paired with programmatic constraint evaluators to solve multi-constraint planning tasks.
The loop
- Proposes
An LLM generates, recombines, and refines candidate travel plans, schedules, and cipher solutions across parallel evolutionary islands.
- Judged by
A programmatic evaluator scores each plan and returns textual feedback on violated constraints.
- Steers
Boltzmann tournament selection, island migration, and elitism propagate high-scoring plans into the next generation.
Headline result
Solved >95% of TravelPlanner and Natural Plan multi-constraint benchmarks with Gemini 1.5 Flash, surpassing Best-of-N and sequential Reflexion by wide margins.
What this does not prove
Requires a programmatic constraint evaluator for the target planning domain to score candidate solutions accurately.
How it generalizes the harness
- The algorithm is the solver
- Inverted control
- A domain authority judges
- Search over a population
- Recursive meta-evolution
- Intra-algorithmic surgery (inside the while-loop)
A multi-island genetic algorithm with programmatic constraint evaluators drives selection, crossover, and migration across generations.
Classification
| What it proposes | Mathematical construction |
|---|---|
| Generative model | Gemini 1.5 Flash / Pro evolutionary operators |
| Algorithmic authority | Programmatic constraint checker and island genetic algorithm |
| Feedback | Verifier errors |
| What gets updated | Population of candidates and prompt context |
| When it runs | At inference time |
| Validity | Formal (machine-checked) |
| Evidence | Experiments reported |
| Lineage | Original system |