Prompt optimization stops rewriting whole strings: Splice freezes contract clauses, winning four of seven benchmarks and matching on three
Synopsis
Splice recasts natural-language prompt optimization as constrained combinatorial search over an edit graph of block-structured states: safety clauses, tool schemas and output formats are tagged immutable so no edit can reach them, and length growth is priced by a ratio penalty. Across seven public benchmarks under five reproduced protocols, Splice wins on four and matches the strongest prior method on three; in a contract stress test, unconstrained optimizers trade away the clauses that cost reward, and frozen text does not mean frozen behavior.
Interpretation
Prompt optimization moves from rewriting one editable string to constrained search over block-structured states, where each block is tagged mutable or immutable and immutable text is preserved verbatim throughout the run. Earlier iterative natural-language optimization treated the prompt as a single editable string, producing three recurrent failures: global rewrites that break safety clauses, candidate populations collapsing onto one lineage, and unchecked text bloat that nothing prices. In a seven-clause contract stress test on SST-5 with gemma-3-27b-it, Splice with the contract frozen modified zero candidates, while the all-mutable, monolithic-editable and do-not-edit-instruction settings all altered contract clauses.
The search loop is driven by three pieces: a critique agent returns positive and negative textual gradients per block, a proposer samples edits only on mutable blocks, and a selector keeps parents and reserves diversity slots. Section-local critique lets an update land on a specific block instead of giving the whole prompt one vague direction, and retaining parents keeps the current leader's validation score from falling. Reproducing five published protocols without modification across seven public benchmarks, Splice wins on SST-5, AIME-2025, GSM8K and BBH and matches the strongest prior method on ARC-Challenge, GPQA-Diamond and PUPA, with its largest lead over GEPA on AIME-2025.
The cost of freezing contract clauses is written as a bound: the loss from freezing is at most the frozen blocks' best single-block gains, inflated by a factor tied to the submodularity ratio. Earlier work neither treated which blocks may not change as a property of the search space nor said what that restriction costs. The contract stress test measures one end of that cost: a clause whose compliance forfeits reward lets editable settings lead by roughly ten points on the proxy metric while trailing by roughly ten points on the contract-aware objective.
Perspective
The demonstrated scope is prompt optimization on seven public benchmarks plus a seven-clause contract stress test built on SST-5 with gemma-3-27b-it; the theoretical results hold under explicit assumptions including cached scoring, bounded rewards, i.i.d. batches and a proposer-coverage condition.
The guarantees are conditional: proposal coverage gates only the convergence rate, and when it fails the search stalls rather than regresses; terminal-quality results rest on a submodularity ratio for which there is no per-benchmark estimator. Freezing preserves a contract's text, not its effect: the unoptimized seed prompt complies on more triggered examples than the frozen-contract optimized result, so frozen blocks still need a behavioral audit.
