This work argues that an LLM judge running in a production system is better understood as having a lifecycle: it must be built, trained, deployed, and continuously maintained as the surrounding data evolves, and each phase poses distinct technical and operational challenges.
Abstract
LLM-as-a-Judge, which leverages a large language model to evaluate natural language generated by another AI application or model, has become a standard, scalable approach for accelerating and extending costly human evaluation. However, most work treats a judge as a static artifact, evaluating it once at construction or against a fixed benchmark. In contrast, we argue that an LLM judge running in a production system is better understood as having a lifecycle: it must be built, trained, deployed, and continuously maintained as the surrounding data evolves, and each phase poses distinct technical and operational challenges. We present such a lifecycle for the LLM judges that evaluate user-facing recommendation explanations at Netflix, where our pipeline generates and the judges assess hundreds of thousands of distinct show-level explanations per week, served across the mobile experience to millions of members. Our framework has four phases: (I) Birth, defining multiple evaluation criteria and building curated benchmark datasets with human labels and rationales; (II) Training, refining the judges'rubrics via Reasoning-Aligned Rubric Tuning (RART), a rubric-tuning procedure that uses a meta-judge over reasoning output as the learning signal; (III) Deployment, in which one judge serves two production roles: quality gating and reflective generation; and (IV) Monitoring, a continuous Human-in-the-Loop alignment process that detects drift and triggers re-tuning behind a human review gate. We report post-launch results from a five-week A/B test over tens of millions of members, in which the judge-aligned explanations shifted member viewing toward novel content (previously unwatched) and increased successful browse-to-play sessions relative to a no-explanation control, with no quality-related takedowns.
As Large Language Models (LLMs) become foundational to next-generation Intelligent Information Systems, the bridge between natural language interfaces and structured database systems remains a critical bottleneck. While Text-to-SQL generation enables cooperative support for complex query formulation, ensuring the reliability of these generated queries at inference time is a central challenge. Conventional methods rely on coarse execution-based signals, which may limit their ability to capture the nuanced semantic alignment required for high-stakes database environments. In this work, we propose the use of Outcome Reward Models (ORMs) as a fine-grained, probabilistic feedback mechanism for test-time verification in Text-to-SQL tasks. We introduce GradeSQL, a framework for training task-specific ORMs that assign scalar utility scores to candidate SQL queries based on their semantic correctness and alignment with database schema. Our approach is evaluated on the BIRD and Spider benchmarks across multiple open-source LLM families. Experimental results demonstrate that ORM-based verification consistently outperforms traditional execution-based heuristics.
M. Tritto, G. Farano, Dario Di Palma et al.· Journal of Intelligence and...· 0 citations
LLM-as-a-judge evaluation has become a dominant paradigm for ranking language models, yet the ecosystem remains fragmented: most benchmarks ship their own code base, hardcode a specific closed-model judge, and support a single evaluation protocol. This fragmentation makes it difficult to study how design choices--the benchmark, the judge model, the prompt, the inference backend--affect the conclusions we draw about model quality. We introduce JudgeArena, an open-source framework that unifies major LLM-judge benchmarks (AlpacaEval, Arena-Hard, MT-Bench, and m-Arena-Hard) under a single interface with swappable judges and comprehensive metadata logging for increased transparency in reporting and reproducibility. It enables systematic studies of judge choices, as any model accessible via vLLM, llama.cpp, or OpenRouter can serve as both candidate and judge. Furthermore, JudgeArena ships with tuned judge configurations for open models that match or outperform closed-model judges, validated on human preference datasets in both English and multilingual settings, reducing the reliance on opaque closed models. Finally, by combining existing human annotations with LLM-judge evaluations of a target model, JudgeArena can simulate LMArena Elo scores with high accuracy offering a practical, open, and low-cost alternative to large-scale human annotation campaigns.
Erlis Lushtaku, Bora Kargi, Ali Elganzory et al.· 0 citations
Large language models routinely produce fluent answers to single-shot prompts, yet deploying them as reliable components of a domain decision system is substantially harder. Closing this gap is the work of harness engineering: the deliberate composition of deterministic scaffolding (symbolic filters, retrieval, schema-typed I/O, LLM-as-judge loops, HITL gates, persistent state, audit trails) around an LLM core. We present a case study in academic supervision, a domain combining high-stakes recommendation, longitudinal accountability, and structured operational workflows. We compare a baseline Academic Supervision Assistant (ASA), a GPT-5 chatbot with no scaffolding, against a multi-module system, Academic Supervision System (ASuS) that wraps the much smaller GPT-4o-mini in a LangGraph harness with symbolic semantic retrieval, schema-validated outputs, LLM-as-judge with bounded retry, HITL gates, deterministic weighted risk scoring with LLM narration, and a per-node SQLite audit trail. The evaluation rubric is retargeted at six harness-mechanism dimensions (grounding, explainability, consistency, process integrity, cognitive load, constraint adherence). A blind ten-rater hybrid evaluation, supplemented by a 2 x 2 model-harness ablation, finds that ASuS, despite using a much smaller base model, outscores ASA on every dimension. Across ten raters the pooled mean for ASuS is 4.08 versus 1.23 for ASA, and 8 of 10 raters reject the null at alpha = 0.05 on a paired Wilcoxon test; full numbers are in Sections 6.4 and 6.7. The ablation confirms that the structural contributions of the harness are largely model-invariant. We extract seven recurring harness-engineering patterns and argue that where reliability, traceability, and institutional consistency matter more than open-ended fluency, harness engineering challenges the prevailing'bigger model is better'intuition.
Large language models (LLMs) increasingly assess generated content, giving rise to the LLM-as-a-Judge paradigm. These systems now score outputs, filter content, and gate iterative refinement in production pipelines, where each judgment is often assumed to be independent of earlier evaluations. We test this assumption using three prompt conditions: no metadata, revision framing, and anchored metadata containing revision, attempt, and prior-score fields. We show that prior scores, even when included only as context metadata, anchor judgments and systematically shift ratings toward their values. Across 192,000 attempted evaluations (185,271 successful), seven out of the eight evaluated models have 95% task-stratified bootstrap intervals below zero for the total anchored-metadata effect on 20 fixed texts. Cohen's $d$, a standardized measure of the difference between score distributions, reaches an absolute value of 0.71. Token-level analysis of selected model-task probes suggests a threshold-like response pattern: introducing anchored metadata produces a marked redistribution of output-score probabilities, while changing the anchor value within the tested below-threshold range produces comparatively little additional variation. On categorical industry data with human-labeled ground truth, anchored metadata blocks 48% of error corrections and flips 10.18% of correct judgments toward an assigned wrong label, demonstrating the bias extends beyond numerical scoring to categorical decisions. Neither Chain-of-Thought nor a metadata-disregard warning reduces the total effect, although the warning improves the paired accuracy effect relative to baseline in the industry experiment. Reliable LLM evaluation demands careful context engineering rather than an assumption of impartiality. Effective mitigation must be validated for the intended model and task or domain.
A. Kapetanović, Kemal Altwlkany, Andro Merćep et al.· 0 citations
The rapid growth in scale and complexity of modern software systems has intensified the need for intelligent and reliable IT operations. While Artificial Intelligence for IT Operations (AIOps) addresses some challenges, existing solutions predominantly rely on isolated, task-specific models that struggle with interpreting multimodal data, incur high maintenance costs, and lack sufficient transparency. Operations Large Language Models (OpsLLMs) offer unified, knowledge-rich reasoning capabilities, yet their evaluation faces significant barriers, including the absence of Ops-centric evaluation taxonomies, limited availability of public datasets, simplistic question-generation methods, and inadequate quality standards for comprehensive operations tasks. We present Eagle, a comprehensive benchmarking framework tailored for evaluating OpsLLMs. Deployed inside Huawei, Eagle ingests enterprise product documentation and synthesizes 4,845 domain-grounded QA pairs across logs, metrics, traces, and configurations, which are paired with a standardized model evaluation system. This deployment supported multiple evaluations of OpsLLM and culminated in an internal horizontal benchmarking report that informed model selection and rollout decisions. Methodologically, Eagle (i) defines an operations-centric taxonomy aligning core LLM abilities with end-to-end operations tasks; (ii) implements an automated question-generation pipeline with multi-granular quality controls validated by human annotation; and (iii) provides reproducible evaluation suites and metrics for scenario-driven reasoning. In offline studies, Eagle-generated test suites improve expert-rated rubric scores by 22%–49% over state-of-the-art baselines, enabling more precise assessments of anomaly detection, fault diagnosis, and root-cause analysis abilities in OpsLLMs. To foster community adoption and reproducibility, we open-source the framework1 and a sanitized dataset2. By bridging general LLM evaluation and operations practice, Eagle delivers a deployable foundation for advancing large-model applications in AIOps.
Yuhe Liu, Changhua Pei, Hang Wang et al.· SIGSOFT FSE Companion· 0 citations
Large Language Models (LLMs) are increasingly used as
judges
to evaluate, rank, and critique AI-generated text and code. This survey provides a comprehensive overview of recent advances (2020–early 2026) in LLM-based evaluation, covering techniques, applications, and challenges across domains. We make three main contributions: (1) a unified taxonomy of LLM judging tasks spanning text (summarization, dialogue, factuality, safety) and code (correctness checking, code review, security analysis); (2) a systematic review of prompting strategies (zero/few-shot, rubric-based, pairwise comparison, chain-of-thought) and advanced pipelines (ensemble judges, multi-agent debate, tool-augmented verification); and (3) an analysis of LLM judge quality, documenting systematic biases (length, position, self-preference) and their mitigations. We review practical applications including benchmark evaluation (MT-Bench, Chatbot Arena), data filtering, and reward modeling for RLHF/RLAIF. Key challenges discussed include calibration, fairness, reproducibility, and adversarial robustness. We conclude with future directions emphasizing standardized protocols, uncertainty estimation, and human–AI collaboration. LLM-based judging shows promise for scalable evaluation, but careful design and rigorous validation are essential to ensure these AI judges meet human standards of accuracy and fairness.
M. Nadăş· Artificial Intelligence Revi...· 0 citations
