Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.30512

Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals

Louise Yorke
View original source ↗Read source document ↗Export metadata ↓

Abstract

LLM judges are used to determine which of two answers is better. Trusting these verdicts requires knowing how repeatable they are and how much they depend on the judge's instructions. We show that similar aggregate scores can conceal substantial changes in individual verdicts, both when the same setup is repeated and when the instructions change. Using 500 Arena-Hard v2.0 prompts with two models' answers held fixed, we run GPT-4.1 as the judge under all eight combinations of criteria, output format and task framing, and repeat every setup unchanged at temperature zero on a pinned model version. Rerunning the default setup moves the aggregate score by about one point but changes the win/tie/loss outcome on 21.2% of prompts. Instruction changes introduce further sensitivity: in one setup, which asks for the verdict without reasoning and also drops several other instructions, outcomes change on 28.2 percentage points more prompts than the two setups' unchanged reruns do on average, and the aggregate score falls by 4.09 points. In another, where the criteria prioritise accuracy, outcomes change on 6.7 percentage points more prompts by the same comparison, while the aggregate score barely moves. A stable aggregate score therefore does not establish that individual verdicts are repeatable. This distinction matters when those verdicts are used as preference labels or to assess individual responses. Evaluations that rely on individual verdicts should repeat each setup they use and report both aggregate scores and prompt-level agreement.

Explore related subjects

cs.CLcs.AIcs.CY

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Louise Yorke. 2026-10-21. Behind Aggregate Scores: What Systematically Varying Arena-Hard's LLM Judge Reveals. https://arxiv.org/abs/2610.30512

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Kongzi: A Historical Large Language Model with Fact Enhancement

Large Language Models (LLMs) can exhibit factual inaccuracies during long-chain reasoning or under misleading prompts, limiting their reliability in historical analysis. We propose Kongzi, a language model for historical understanding and reasoning that combines continued pre-training on historical corpora, supervised fine-tuning, and fact-enhanced reinforcement learning. Its factuality reward compares entities and relations extracted from generated responses with historical references through knowledge-graph similarity. We evaluate 0.5B and 7B models on historical question answering and counterfactual reasoning, using reported ratings from three LLM judges, and on CMMLU-Chinese-History. Kongzi-7B receives the highest historical-accuracy ratings among the compared models under all three judges. Reinforcement learning improves the mean answer scores of both model sizes and their CMMLU-Chinese-History accuracy, although Kongzi-7B remains below the strongest reported baselines on that benchmark. These descriptive results support further study of factuality rewards for historical language models.

cs.CL↗
arXiv2026

ShIOEnv: A Command Evaluation Environment for Grammar-Constrained Synthesis and Execution Behavior Modeling

We introduce ShIOEnv, a Gymnasium-compatible Bash environment for synthesizing execution-annotated shell input traces. Datasets containing such traces are useful for training models to perform execution-free shell simulation, which in turn is useful for deploying honeypots for cyber deception. Existing datasets, such as NL2CMD and other Bash command corpora, do not include complete execution artifacts (stdout, stderr, exit status, and changes to the system's state) associated with a command's execution. ShIOEnv synthesizes shell inputs, executes them in a controlled system context, and records stdout, stderr, exit status, and structured filesystem and environment patches. Unconstrained (random) generation of shell inputs (command and arguments) results in multiple instances of arguments that are mostly redundant or result in execution errors. We introduce grammar-constrained synthesis to generate syntactically valid arguments with high input irreducibility, which we estimate through self-ablative contrasts of execution artifacts. We prepare 1.6M samples using grammar-constrained and unconstrained synthesis and collect their execution artifacts for 86 Linux/Bash utilities. Using this dataset, we train small language models (CodeT5) and evaluate their accuracy in predicting the actual execution behavior of attacker-sourced Bash commands. These models outperform commonly used rule-based simulation (Cowrie) and prompted large language models by up to 14.2 pp in output edit similarity (O-ES). We enhance the NL2CMD dataset by constructing 500K samples executable in a grounded system context along with their complete execution artifacts. We empirically demonstrate that models trained on grammar-constrained samples outperform those trained on NL2CMD samples by 22.5 pp in O-ES. ShIOEnv, command grammars, and 2.1M samples used in our primary evaluations are released for use by other researchers.

cs.CL↗
arXiv2026

How Well Does Mathematical Reasoning Transfer Across Languages in English-Centric LLMs?

Recent advances in Reinforcement Post-Training (RPT) have substantially improved the mathematical reasoning capabilities of large language models (LLMs), but it remains unclear how well such gains transfer across languages. In this work, we study cross-lingual mathematical reasoning transfer in English-centric LLMs under controlled multilingual evaluation. We first evaluate a range of English-centric reasoning models on multilingual mathematical reasoning benchmarks and find that transfer is highly variable across initial model choice, target language, and training paradigm. Through controlled comparative experiments, we further find that stronger English-centric initialization does not necessarily lead to stronger relative transfer efficiency, even when it yields higher multilingual reasoning accuracy. Finally, we show that the largest improvement comes from moving beyond English-only supervision: adding a single parallel language yields a substantial gain, while further gains from additional languages are smaller but consistent. Overall, our results highlight the need to evaluate multilingual mathematical reasoning directly, using both absolute performance and relative transfer-efficiency metrics.

cs.CL↗