TBNAI
PromptOps: Iterative Prompt Optimization for LLM-as-a-Judge Systems
Publication: cs.AI updates on arXiv.org
Author: Sven Kleinkop, Tam\'as Janusko, Ricardo Bochnia, Gunnar Hempel, Steffen Tomschke, J\"urgen Anke, Maik Thiele
Source published: 06/11/2026, 05:00:00 UTC
Briefing created: 06/11/2026, 05:54:10 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 06/11/2026, 05:03:25 UTC.
Briefings
This article introduces PromptOps, a closed-loop framework designed to maintain Large Language Models (LLMs) acting as automated judges for AI systems. The study addresses 'judge drift,' where LLM agreement with human evaluation policies degrades over time due to changing user interactions and score distributions. The authors instantiate PromptOps with five optimization methods: GEPA, SIMBA, Rubric Synthesis, and exemplar selection using simulated annealing or a Tree-structured Parzen Estimator. Experiments show that iterative optimization can recover performance after controlled shifts in reference score distributions, establishing it as an inherently iterative maintenance problem.
Key points
- PromptOps is a closed-loop framework that iteratively recalibrates LLM-as-a-Judge systems to counteract judge drift.
- Five optimization methods were tested: GEPA, SIMBA, Rubric Synthesis, and exemplar selection using simulated annealing or a Tree-structured Parzen Estimator.
Context
LLMs are increasingly used as automated judges for evaluating chatbots and AI systems. However, their agreement with intended human evaluation policies often degrades as user interactions and score distributions change. Existing methods are typically one-shot procedures on static development sets, offering limited guidance for maintaining LLM judges over time.
Limitations
The study's findings are limited to the specific experimental settings used. Optimization algorithm choice and update timing do not generalize universally across different settings. The framework's effectiveness depends on the specific context of the deployment, and the study does not claim universal applicability of the methods.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 06/11/2026, 05:54:10 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeig6salh5gco6x3txnx7nvl3wlvrsm5yyiqfdkypizwh5wmynsh5le