SPLICE: Structured Prompt Local Iterative Combinatorial Evolution

Impact 7 arXiv cs.AI

Summary

  • Introduces SPLICE, a unified natural language optimization framework framed as constrained combinatorial search over block-structured, selectively-mutable states.
  • Addresses recurrent natural language optimization failure modes like unsafe global rewrites, beam collapse, and unchecked bloat using formal guarantees and mechanisms.
  • Features per-block mutability to preserve contracts, elitist beam search with diversity slots, and a ratio length regularizer.
  • Across seven public benchmarks, SPLICE wins on four and matches the strongest prior method on three while preserving essential contract clauses.

Why it matters

SPLICE provides a principled method for natural language optimization that avoids safety drift and prompt bloat, improving prompt engineering reliability for production AI systems.

Who should care