The authors replayed 600 logged runs spanning five optimization tasks and three evolutionary search methods.
Token accounting shrinks evolutionary search’s apparent advantage over sampling
Across five program-optimization tasks, matching token budgets rather than proposal counts erased most of evolution’s reported lead.
Big Tech
Tal Oved · Roi Pony · Oshri Naparstek · Udi Barzelay
IBM Research
Research Digest··3 min read
Oved and colleagues reanalyzed 600 logged runs from three LLM evolutionary search strategies, measuring search cost in tokens and auditing the programs selected by their scorers.
§