Same Input, Different Scores: A Multi Model Study on the Inconsistency of LLM Judge

• URL: http://arxiv.org/abs/2603.04417v1
• Date: Wed, 04 Feb 2026 16:03:08 GMT
• Title: Same Input, Different Scores: A Multi Model Study on the Inconsistency of LLM Judge
• Abstract summary: This study systematically evaluates scoring stability across five commonly used models, GPT-4o, GPT-4o-mini, Gemini-2.5-Flash, Claude-Haiku-4.5, and Claude-Sonnet-4.5.<n>We address three questions: how stable a model's scores are across repeated runs, how differently models score identical inputs, and how temperature affects scoring consistency.
• Score: 0.0
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Large language models are increasingly used as automated evaluators in research and enterprise settings, a practice known as LLM-as-a-judge. While prior work has examined accuracy, bias, and alignment with human preferences, far less attention has been given to how consistently LLMs assign numerical scores, an important concern for many production workflows. This study systematically evaluates scoring stability across five commonly used models, GPT-4o, GPT-4o-mini, Gemini-2.5-Flash, Claude-Haiku-4.5, and Claude-Sonnet-4.5, two temperature settings, and real enterprise question-answer pairs drawn from a retrieval-augmented generation (RAG) system. We address three questions: how stable a model's scores are across repeated runs, how differently models score identical inputs, and how temperature affects scoring consistency. Temperature controls the determinism of an LLM's output. Despite expectations of stability at temperature=0, we observe substantial variability across models, with completeness scoring showing the largest fluctuations. Cross-model comparisons reveal systematic differences in strictness and interpretive style, leading to divergent ratings for the same answers. Lower temperatures improve stability for some models, notably GPT-4o and Gemini, but have limited or inconsistent effects for Anthropic models. These findings have important implications for enterprise pipelines that rely on LLM-generated scores for routing, triage, gating, or quality control. Identical inputs can receive different scores depending on model, family, or temperature, raising concerns around fairness, reproducibility, and operational reliability. Our results highlight the need for monitoring, robust parsing, and hybrid human-LLM evaluation strategies to ensure dependable use of LLM-as-a-judge in production environments.

Related papers

• Chain-of-Experience for Continual LLM Improvement [57.11645184182825]
  We study how large language models learn from iterative experience at test time.<n>We instantiate Chain-of-Experience (CoE) with diverse feedback mechanisms.<n>Our study shows that leveraging iterative experience consistently outperforms feedback-free baselines.
  arXiv  Detail & Related papers  (2026-08-18T17:22:54Z)
• DualEval: Joint Model-Item Calibration for Unified LLM Evaluation [67.99119926012686]
  DualEval is a latent model-item calibration framework that represents models and evaluation items in a shared space.<n>We apply DualEval across four domains: coding, math, miscellaneous domain-knowledge tasks, and generic everyday user queries.
  arXiv  Detail & Related papers  (2026-06-24T22:40:46Z)
• How Independent are Large Language Models? A Statistical Framework for Auditing Behavioral Entanglement and Reweighting Verifier Ensembles [46.63622714488747]
  Shared pretraining data, distillation, and alignment pipelines can induce hidden behavioral dependencies, latent entanglement.<n>In practice, this manifests as correlated reasoning patterns and synchronized failures.<n>We develop a statistical framework for auditing behavioral entanglement among black-box language models.
  arXiv  Detail & Related papers  (2026-04-08T23:32:06Z)
• Does Tone Change the Answer? Evaluating Prompt Politeness Effects on Modern LLMs: GPT, Gemini, LLaMA [0.6263481844384227]
  This work proposes a systematic evaluation framework to examine how interaction tone affects model accuracy.<n>We apply this framework to three recently released and widely available large language models: GPT-4o mini (OpenAI), Gemini 2.0 Flash (Google DeepMind), and Llama 4 Scout (Meta)<n>Our results show that tone sensitivity is both model-dependent and domain-specific. Neutral or Very Friendly prompts generally yield higher accuracy than Very Rude prompts, but statistically significant effects appear only in a subset of Humanities tasks.
  arXiv  Detail & Related papers  (2025-12-14T19:25:20Z)
• Evaluating from Benign to Dynamic Adversarial: A Squid Game for Large Language Models [57.33350664910483]
  We introduce Squid Game, a dynamic and adversarial evaluation environment with resource-constrained and asymmetric information settings.<n>We evaluate over 50 LLMs on Squid Game, presenting the largest behavioral evaluation study of general LLMs on dynamic adversarial scenarios.
  arXiv  Detail & Related papers  (2025-11-12T06:06:29Z)
• The Chameleon Nature of LLMs: Quantifying Multi-Turn Stance Instability in Search-Enabled Language Models [1.4323566945483497]
  We present the first systematic investigation of "chameleon behavior" in Large Language Models.<n>We expose fundamental flaws in state-of-the-art systems.<n>Our analysis uncovers the mechanism: strong correlations between source re-use rate and confidence are statistically significant.
  arXiv  Detail & Related papers  (2025-10-19T04:51:14Z)
• Analyzing the Instability of Large Language Models in Automated Bug Injection and Correction [0.0]
  Large Language Models (LLMs) are used in software engineering tasks.<n>When executed at different times with the same input, they can generate radically different code.<n>This study examines how unstable an LLM is when it comes to fixing code bugs.
  arXiv  Detail & Related papers  (2025-09-08T08:23:49Z)
• Play Favorites: A Statistical Method to Measure Self-Bias in LLM-as-a-Judge [17.40713507922006]
  Large language models (LLMs) can serve as judges that offer rapid and reliable assessments of other outputs.<n>LLMs may systematically assign overly favorable ratings to their own outputs, a phenomenon known as self-bias.<n>We present a statistical framework that explicitly formalizes assumptions under which self-bias can be identified and estimated.
  arXiv  Detail & Related papers  (2025-08-08T21:22:12Z)
• DeepCritic: Deliberate Critique with Large Language Models [77.5516314477878]
  We focus on studying and enhancing the math critique ability of Large Language Models (LLMs)<n>Our developed critique model built on Qwen2.5-7B-Instruct significantly outperforms existing LLM critics on various error identification benchmarks.
  arXiv  Detail & Related papers  (2025-05-01T17:03:17Z)
• Influences on LLM Calibration: A Study of Response Agreement, Loss Functions, and Prompt Styles [4.477423478591491]
  Calib-n is a novel framework that trains an auxiliary model for confidence estimation.<n>We find that few-shot prompts are the most effective for auxiliary model-based methods.
  arXiv  Detail & Related papers  (2025-01-07T18:48:42Z)
• VHELM: A Holistic Evaluation of Vision Language Models [75.88987277686914]
  We present the Holistic Evaluation of Vision Language Models (VHELM) VHELM aggregates various datasets to cover one or more of the 9 aspects: visual perception, knowledge, reasoning, bias, fairness, multilinguality, robustness, toxicity, and safety. Our framework is designed to be lightweight and automatic so that evaluation runs are cheap and fast.
  arXiv  Detail & Related papers  (2024-10-09T17:46:34Z)
• Calibrating Large Language Models with Sample Consistency [76.23956851098598]
  We explore the potential of deriving confidence from the distribution of multiple randomly sampled model generations, via three measures of consistency. Results show that consistency-based calibration methods outperform existing post-hoc approaches. We offer practical guidance on choosing suitable consistency metrics for calibration, tailored to the characteristics of various LMs.
  arXiv  Detail & Related papers  (2024-02-21T16:15:20Z)
• Are You Sure? Challenging LLMs Leads to Performance Drops in The FlipFlop Experiment [82.60594940370919]
  We propose the FlipFlop experiment to study the multi-turn behavior of Large Language Models (LLMs) We show that models flip their answers on average 46% of the time and that all models see a deterioration of accuracy between their first and final prediction, with an average drop of 17% (the FlipFlop effect) We conduct finetuning experiments on an open-source LLM and find that finetuning on synthetically created data can mitigate - reducing performance deterioration by 60% - but not resolve sycophantic behavior entirely.
  arXiv  Detail & Related papers  (2023-11-14T23:40:22Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.