OPRO
Optimization by PROmpting
Uses an LLM as an iterative black-box optimizer conditioned on a sorted trajectory of (solution, objective-score) pairs evaluated by an external oracle.
The loop
- Proposes
An optimizer LLM proposes K new candidate solutions (routes, linear regression weights, or instructions) given the top-scoring history.
- Judged by
An exact objective evaluator computes the ground-truth cost (TSP tour length, MSE loss, or training accuracy) for every candidate.
- Steers
The newly scored candidates are inserted into the sorted priority buffer, shifting the optimization window toward better objective values.
Headline result
Discovered prompts that boosted GSM8K accuracy by up to 8 percentage points over human baselines and solved small-scale TSP and linear regression trajectories.
What this does not prove
On large combinatorial instances (>50 TSP nodes), in-context coordinate proposals lag far behind dedicated Concorde or LKH solvers.
How it generalizes the harness
An external objective evaluator scores every proposed candidate and maintains the sorted optimization trajectory buffer.
Classification
| What it proposes | Heuristic |
|---|---|
| Generative model | PaLM 2 / GPT-4 optimizer model |
| Algorithmic authority | Exact objective function evaluator and sorted trajectory buffer |
| Feedback | Score |
| What gets updated | Population of candidates and prompt context |
| When it runs | Offline, at design time |
| Validity | Empirical (measured) |
| Evidence | Experiments reported |
| Lineage | Original system |