Research map / Large language models / Language, safety and interpretability

Brittlebench: Quantifying LLM robustness via prompt sensitivity

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Statistical frameworks for LLM evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces Brittlebench, a variance-decomposition meta-evaluation that applies 15 semantics-preserving prompt perturbations to six MCQ benchmarks and separates item-difficulty variance from perturbation-induced variance across open-weight and commercial LLMs.

Problem

Static benchmark scores may overstate model capability because they ignore sensitivity to semantics-preserving prompt variations (typos, spacing, padding, paraphrasing), and no principled method quantifies how much performance variance comes from prompt formulation.

Approach

A variance decomposition (law of total variance, equivalent to a nested random-effects ANOVA) splits correctness variance over data items and perturbations into item-difficulty and perturbation-driven components, yielding model- and benchmark-level brittleness scores over a taxonomy of perturbations.

Key results

• Semantics-preserving perturbations reduce average accuracy by up to ~12% (Qwen3-4B few-shot prompt padding with new lines drops 12.83%), and Claude 4.5 Opus drops ~2% on MMLU under character typos.
• Perturbation-driven variance accounts for roughly half of total variance in most open-weight models and more than ~25% for commercial models; MMLU and GPQA are dominated by perturbation variance.
• Model rankings change under a single perturbation in 63% of cases; CoT reduces the perturbation-induced accuracy drop only from 2.79% to 2.38% for Claude 4.5 Opus.

Surprising findings

• Paraphrasing perturbations are benign and occasionally improve open-weight model accuracy, possibly due to self-bias or standardization of language.
• Chain-of-thought improves accuracy but only weakly mitigates brittleness (17% relative reduction in drop).

Limitations

• GPT-5 evaluated in minimal reasoning mode, so results are a lower bound on its performance.
• Semantic preservation is defined behaviorally and partly validated with embedding similarity; some perturbations may alter meaning and paraphrase quality had modest LLM-judge agreement (Spearman 0.447).
• Evaluation limited to MCQ benchmarks and a small set of models; mixed-effects CoT effect size is small.

Experimental setup

Datasets: MMLU ARC TruthfulQA MathQA LogiQA GPQA

Benchmarks: MMLU ARC TruthfulQA MathQA LogiQA GPQA

Metrics: accuracy accuracy drop variance decomposition (brittleness score) Spearman rank correlation McNemar test cosine similarity

Baselines: Claude 4.5 Opus 0-shot unperturbed GPT-5 0-shot unperturbed

Builds on: lm-evaluation-harness PromptRobust EmotionPrompt

Closest papers by meaning

• On the Worst Prompt Performance of Large Language Models (NeurIPS 2024 · similarity 0.92)
• Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting (ICLR 2024 · similarity 0.91)
• Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.90)
• Mapping from Meaning: Addressing the Miscalibration of Prompt-Sensitive Language Models (AAAI 2025 · similarity 0.90)
• An Empirical Analysis of Uncertainty in Large Language Model Evaluations (ICLR 2025 · similarity 0.90)
• PertEval: Unveiling Real Knowledge Capacity of LLMs with Knowledge-Invariant Perturbations (NeurIPS 2024 · similarity 0.89)