Prompt optimisation

The loop

1. 1EvaluateThe current prompt plays one episode on each of three training floors (the simple floor, gen01, gen02) at temperature 0; fitness is the mean goal score.
2. 2Build the requestScope, a description of the observation, the rules a prompt must follow, the current prompt, one trace per episode and up to two worse rewrites.
3. 3ReflectThe teacher writes a short diagnosis and one complete revised prompt.
4. 4ValidateA fixed check rejects coordinates, monster names, routes, lists of allowed fights and numbers of 100 or more; one retry with the reasons.
5. 5SelectThe candidate plays the same floors; it becomes the parent if it scores at least as well as the best so far.

This is a (1+1) evolutionary loop. The Pareto loop of the second study keeps every valid candidate, draws parents from those that are best on at least one floor, and merges the two best once (Section 5.7, Table 5.2).

Study 1: does evolution add anything once the observation is annotated?

Two weaker models, three prompts (rules-only, evolved from it, hand-written single-floor prompt), each under the full and the bare observation, evaluated on thirteen held-out generated floors with one episode each. Hypotheses were stated before the runs.

What the evolved prompts contain

How often the candidates of both studies state each rule of the hand-written prompts, all written from the rules-only prompt by a teacher given no example rules (Table 6.7, counted by keyword patterns).

All runs