There Is No Neutral Harness: Modern LLM Leaderboards Are Manufactured by Config-Fragile Items
06:00 · August 25, 2026 · arXiv cs.AI RSS
Multiple-choice benchmarks fix the questions and the correct answers, but not the harness: the order of the options, the wording of the prompt, and whether a language model's answer is read from generated text or from per-option likelihoods. Work on this harness sensitivity reports it as aggregate score variance, leaving unexamined which items the variance falls on and whether they are the items that separate one model from the next. We treat the evaluation harness of large language models (LLMs) as an independent variable and resolve its effect to single items. We introduce the \textit{fragility grid}: 12 open-weight instruction-tuned LLMs from 4 families answer the same 3{,}679 items from 4 benchmarks (ARC, HellaSwag, MMLU, TruthfulQA) under 26 equally defensible harness configurations, recording one correctness bit for every model, item, and configuration. The comparison is matched, since the items, the weights, and the greedy decoding stay fixed while only the harness varies. Under the grid a model's score is a band rather than a point: gemma4-31b scores between 31 and 89 percent depending only on the harness. Three results follow. On the items that two adjacent models both answer stably the pair is tied, and config-fragile items carry 95.7 percent of a pair's gap on average. Four of the 12 models reach rank one under some configuration, so the harness selects the winner. Item discrimination, the property that benchmark-compression methods maximize, correlates with fragility at 0.28 (95 percent CI 0.25 to 0.30), so compression keeps the fragile items rather than removing them. The scoring choice, not the option order that protocols usually fix, is the load-bearing axis. We release the per-item records and the analysis script, from which every number regenerates on a CPU in seconds, and we position the fragility grid as a check a leaderboard can run before it reports an order.
Summary
Modern LLM leaderboards rest on multiple-choice benchmarks whose questions and answers are fixed, yet whose evaluation harness remains under-specified. The harness encompasses option ordering, prompt phrasing, and the method used to extract an answer—whether from generated text or from per-option likelihoods. Each of these choices is defensible and appears in widely used evaluation code, but they are rarely treated as controlled variables.
A fragility grid constructed for the study holds items, model weights, and greedy decoding constant while varying only the harness across 26 configurations. Twelve open-weight instruction-tuned models from four families answered the same 3,679 items drawn from ARC, HellaSwag, MMLU, and TruthfulQA. Under this matched design a model’s accuracy becomes a band rather than a point; one model ranged from 31 percent to 89 percent solely by harness choice.
When adjacent models are compared on items both answer stably across all configurations, the pair is effectively tied. Five of eleven pairs match exactly, four differ on a single item, and the remainder net to zero. Every observed gap on the full set is therefore carried by config-fragile items, which account for 95.7 percent of the difference on average. Four of the twelve models reach first place under at least one configuration, demonstrating that the harness itself selects the reported winner.
Item discrimination, the quantity maximized by common benchmark-compression techniques, correlates only modestly with fragility (0.28, 95 percent CI 0.25–0.30). The most discriminative items prove fragile for 96 percent of models, so compression tends to retain rather than eliminate the sources of rank instability. Among the three harness axes, the choice between generation-based and likelihood-based scoring produces the largest swings in credited answers; option order and prompt wording exert smaller effects once scoring is fixed.
The authors release per-item correctness records and an analysis script that regenerates every reported figure on a CPU in seconds. They position the fragility grid as a lightweight pre-publication check that any leaderboard can apply before asserting an ordering.
Why it matters
For Dutch AI researchers and enterprises focused on transparent and robust AI, this paper exposes critical flaws in standard LLM evaluation methods. Understanding harness fragility is essential for accurately assessing model capabilities and ensuring compliance with EU standards for reliable AI deployment.
