Promptbreeder
Self-referential evolutionary algorithm that simultaneously evolves task heuristics and the mutation prompts that generate them under benchmark fitness.
The loop
- Proposes
An LLM applies a population of evolving mutation-prompts to parent task-prompts (plus Lamarckian working-out traces) to produce offspring.
- Judged by
An automated benchmark harness evaluates each candidate on multi-step mathematical, logic, and commonsense training batches.
- Steers
Binary tournament genetic selection updates both the task-prompt genome and the meta-level mutation-prompt genome.
Headline result
Outperformed Chain-of-Thought, Plan-and-Solve, and OPRO across arithmetic, commonsense, and hate-speech classification benchmarks.
What this does not prove
Requires thousands of forward evaluation passes across the evolutionary population before convergence.
How it generalizes the harness
A binary-tournament genetic algorithm controls population evolution over both task heuristics and hyper-mutation operators.
Classification
| What it proposes | Heuristic |
|---|---|
| Generative model | PaLM 2L self-referential evolutionary mutator |
| Algorithmic authority | Tournament genetic algorithm and benchmark fitness evaluator |
| Feedback | Score |
| What gets updated | Population of candidates |
| When it runs | Offline, at design time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Original system |