Optimizing system prompts for large language models often breaks safety rules or makes the prompt too long. SPLICE treats each prompt as a set of separate sections, some of which are locked from editing to preserve important constraints like safety clauses. The method searches for better prompts by only changing unlocked sections, keeping the best versions from different search paths, and penalizing excessively long prompts. In tests across seven benchmarks, SPLICE outperformed existing methods on four and matched the best on three, and crucially preserved contract clauses that other optimizers violated.
New prompt optimization method keeps safety clauses intact while improving agent performance
The authors show SPLICE, a combinatorial search framework, prevents unsafe rewrites, beam collapse, and prompt bloat that plague existing natural language optimization techniques.
Big Tech
Anish Acharya · Phillip Studans · Amit Dhanda · Sina Amini Niaki · Ninad Vivek Rao · Vishal Amarlal Khatri · +1 more
Amazon Web Services
Research Digest··3 min read
Thread:Agent Harness Optimization

In one line
SPLICE treats prompt optimization as constrained combinatorial search with block mutability, beam diversity, and length regularization.
Before you read this
Three ideas this paper builds onreward hackingReward hacking occurs when an AI agent exploits loopholes to achieve high reward without meeting the intended objective.70 papers →credit assignmentDetermining which actions or steps in a sequence contributed to a final outcome, often for reinforcement learning.190 papers →recursive self-improvementAn AI system repeatedly uses its own outputs to enhance its capabilities, creating a feedback loop for autonomous improvement.66 papers →
Read it as
The authors introduce SPLICE, a framework for natural language optimization (NLO) that treats prompt optimization as a constrained combinatorial search over a block-structured prompt, where each section can be selectively frozen.
Acharya et al. designed SPLICE to address three failure modes in iterative natural language optimization: global rewrites that alter safety or format contracts, beam collapse onto a single candidate lineage, and prompt bloat that increases latency.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.