AI
SPLICE: Structured Prompt Local Iterative Combinatorial Evolution
Publication: cs.AI updates on arXiv.org
Author: Anish Acharya, Phillip Studans, Amit Dhanda, Sina Amini Niaki, Ninad Vivek Rao, Vishal Amarlal Khatri, Brian Verkhovsky
Source published: 17/11/2026, 05:00:00 UTC
Briefing created: 17/11/2026, 08:34:51 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 17/11/2026, 08:33:24 UTC.
Briefings
SPLICE is a new framework for natural language optimization that uses constrained combinatorial search over an edit graph to improve agent policies. It addresses three failure modes of iterative methods by enforcing per-block mutability, using elitist beam search with diversity slots, and applying a ratio length regularizer. The method was tested on seven public benchmarks under five protocols, winning on four and matching the strongest prior method on three. On a production-style contract stress test, SPLICE successfully preserved contract clauses that other unconstrained optimizers discarded, raising the proxy metric. The study was published on arXiv in November 2026.
Key points
- SPLICE uses a unified NLO framework that treats optimization as constrained combinatorial search over an edit graph of block-structured states.
- The framework maps three failure modes to specific guarantees: per-block mutability preserves contract text verbatim with a price-of-freezing bound, elitist beam search with diversity slots prevents leader regression under noisy re-evaluation, and a ratio length regularizer bounds accuracy-length bias.
Context
Natural language optimization (NLO) improves agents by having models critique and rewrite their natural language policy parameters, including system prompts, tool instructions, and reasoning directives. Most iterative NLO methods fail due to unsafe global rewrites, beam collapse, or unchecked bloat, whereas SPLICE introduces formal guarantees for each failure mode.
Limitations
The study was conducted on seven public benchmarks under five reproduced protocols. Results on a production-style contract stress test were limited to the specific ablation showing gains from critique and elitist search, with no further details provided on the full scope of contract preservation or other stress test outcomes.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 17/11/2026, 08:34:51 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeib6yj3m64dovhjls3lfevo3fw7keffmrjccrnnbknfmvtuhcqpaoy