Pith. sign in

REVIEW 9 cited by

On Robustness and Reliability of Benchmark-Based Evaluation of LLMs

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2509.04013 v1 pith:ZBGWQIP3 submitted 2025-09-04 cs.CL cs.AI

On Robustness and Reliability of Benchmark-Based Evaluation of LLMs

classification cs.CL cs.AI
keywords effectivenessllmsacrossbenchmark-basedbenchmarksevaluationquestionsrobustness
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

Large Language Models (LLMs) effectiveness is usually evaluated by means of benchmarks such as MMLU, ARC-C, or HellaSwag, where questions are presented in their original wording, thus in a fixed, standardized format. However, real-world applications involve linguistic variability, requiring models to maintain their effectiveness across diverse rewordings of the same question or query. In this study, we systematically assess the robustness of LLMs to paraphrased benchmark questions and investigate whether benchmark-based evaluations provide a reliable measure of model capabilities. We systematically generate various paraphrases of all the questions across six different common benchmarks, and measure the resulting variations in effectiveness of 34 state-of-the-art LLMs, of different size and effectiveness. Our findings reveal that while LLM rankings remain relatively stable across paraphrased inputs, absolute effectiveness scores change, and decline significantly. This suggests that LLMs struggle with linguistic variability, raising concerns about their generalization abilities and evaluation methodologies. Furthermore, the observed performance drop challenges the reliability of benchmark-based evaluations, indicating that high benchmark scores may not fully capture a model's robustness to real-world input variations. We discuss the implications of these findings for LLM evaluation methodologies, emphasizing the need for robustness-aware benchmarks that better reflect practical deployment scenarios.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 9 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. PRISM: A Benchmark for Programmatic Spatial-Temporal Reasoning

    cs.AI 2026-05 unverdicted novelty 7.0

    PRISM benchmark of over 10k pairs shows LLMs have a 41% average drop from code execution success to spatial correctness in programmatic video generation.

  2. Same Question, Different Answers: Evaluating LLM Reliability Beyond Accuracy

    cs.AI 2026-05 conditional novelty 6.0

    Models flip between correct and incorrect answers on over 23% of questions under meaning-preserving paraphrases, so single-prompt accuracy overstates reliable knowledge.

  3. GSM-SEM: Benchmark and Framework for Generating Semantically Variant Augmentations

    cs.CL 2026-05 unverdicted novelty 6.0

    GSM-SEM is a reusable framework for creating semantically variant augmentations of math benchmarks like GSM8K that alter facts but preserve answers and difficulty, with evaluations showing LLM performance drops of up ...

  4. GSM-SEM: Benchmark and Framework for Generating Semantically Variant Augmentations

    cs.CL 2026-05 unverdicted novelty 6.0

    GSM-SEM generates reusable, stochastic semantic variants of math reasoning benchmarks that alter underlying facts but preserve answers, producing larger LLM performance drops than prior surface-level variants.

  5. StarDrinks: An English and Korean Test Set for SLU Evaluation in a Drink Ordering Scenario

    cs.CL 2026-04 unverdicted novelty 6.0

    StarDrinks is a new English and Korean test set supporting speech-to-slots SLU, transcription-to-slots NLU, and ASR evaluation in a drink ordering scenario.

  6. Safe for Whom? Rethinking How We Evaluate the Safety of LLMs for Real Users

    cs.AI 2025-12 unverdicted novelty 6.0

    LLM safety evaluations for personal advice must test responses against diverse user vulnerability profiles, since context-blind ratings overestimate safety and realistic prompt context does not fix the problem.

  7. Learning to Act under Noise: Enhancing Agent Robustness via Noisy Environments

    cs.AI 2026-05 unverdicted novelty 5.0

    NoisyAgent trains LLM agents with controlled user and tool noise to improve robustness in stochastic environments while also boosting clean-benchmark performance.

  8. MAVEN: Improving Generalization in Agentic Tool Calling

    cs.AI 2026-05 unverdicted novelty 4.0

    MAVEN is a modular verification scaffold that lifts an open 120b model's tool-calling accuracy from 48% to 71% on MAVEN-Bench without retraining.

  9. StarDrinks: An English and Korean Test Set for SLU Evaluation in a Drink Ordering Scenario

    cs.CL 2026-04 conditional novelty 4.0

    StarDrinks provides English and Korean speech, transcripts, and slot labels for drink-ordering to benchmark speech-to-slots SLU, NLU, and ASR under realistic variability.