Pith. sign in

REVIEW 11 cited by

Rethinking Benchmark and Contamination for Language Models with Rephrased Samples

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

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 2311.04850 v2 pith:KCWBYPZB submitted 2023-11-08 cs.CL cs.AI

Shuo Yang , Wei-Lin Chiang , Lianmin Zheng , Joseph E. Gonzalez , Ion Stoica This is my paper · ORCID
classification cs.CLcs.AI
keywords datadecontaminationbenchmarkcontaminationtestbenchmarksmodelspre-training
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Large language models are increasingly trained on all the data ever produced by humans. Many have raised concerns about the trustworthiness of public benchmarks due to potential contamination in pre-training or fine-tuning datasets. While most data decontamination efforts apply string matching (e.g., n-gram overlap) to remove benchmark data, we show that these methods are insufficient, and simple variations of test data (e.g., paraphrasing, translation) can easily bypass these decontamination measures. Furthermore, we demonstrate that if such variation of test data is not eliminated, a 13B model can easily overfit a test benchmark and achieve drastically high performance, on par with GPT-4. We validate such observations in widely used benchmarks such as MMLU, GSK8k, and HumanEval. To address this growing risk, we propose a stronger LLM-based decontamination method and apply it to widely used pre-training and fine-tuning datasets, revealing significant previously unknown test overlap. For example, in pre-training sets such as RedPajama-Data-1T and StarCoder-Data, we identified that 8-18\% of the HumanEval benchmark overlaps. Interestingly, we also find such contamination in synthetic dataset generated by GPT-3.5/4, suggesting a potential risk of unintentional contamination. We urge the community to adopt stronger decontamination approaches when using public benchmarks. Moreover, we call for the community to actively develop fresh one-time exams to evaluate models accurately. Our decontamination tool is publicly available at https://github.com/lm-sys/llm-decontaminator.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 11 Pith papers

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

  1. CapBencher: Give Your LLM Benchmark a Built-in Alarm for Test-Set Overfitting

    cs.LG 2025-05 accept novelty 7.0 of 10

    Capping achievable accuracy with randomized correct answers turns any model that exceeds the cap into a detectable contamination alarm.

  2. EuroExec: Frontier Language Models Fall Short of Expert Judgment on European Executive Decision Tasks

    cs.CL 2026-08 conditional novelty 6.0 of 10

    Frontier LLMs solve at most 56.9% of 413 expert-authored European executive decision tasks, while expert-written reference answers solve 92.4%, showing a large capability gap.

  3. Temporal Leakage in LLM Backtesting: Measurement, Validation, and Adjusted Scores

    cs.LG 2026-08 conditional novelty 6.0 of 10

    The standard pre/post cutoff check cannot separate memorization from recency, and a single external reference is needed to measure and adjust for temporal leakage.

  4. Do Agent Benchmarks Measure Capability? Protocol Validity in the Age of Agentic AI

    cs.AI 2026-07 conditional novelty 6.0 of 10

    A post-hoc audit framework finds that many agent benchmark scores are inflated by protocol exposures (public answers, readable hidden state, generator regularities, feedback, or scoring flaws), with measured inflation...

  5. SwiftEval: Developing a Language-Specific Benchmark for LLM-generated Code Evaluation

    cs.LG 2025-05 conditional novelty 6.0 of 10

    SwiftEval, a 28-problem hand-crafted Swift benchmark, evaluates 44 code LLMs and shows large performance drops on Swift tasks, especially for smaller models.

  6. Correcting Prompt Dependence in LLM Benchmarks: A Bayesian Hierarchical Model with Embedding-Space Clustering

    cs.CR 2025-10 conditional novelty 5.0 of 10

    A Bayesian model that groups similar LLM test prompts into clusters gives better predictive scores than a no-clustering baseline but does not prove that it truly corrects prompt dependence.

  7. OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique

    cs.CL 2025-07 conditional novelty 5.0 of 10

    A 2.5M-example code reasoning dataset with critique traces enables Qwen2.5-based models to surpass prior open-weight distilled models on LiveCodeBench via test-time self-critique selection.

  8. BenchHub: A Unified Benchmark Suite for Holistic and Customizable LLM Evaluation

    cs.LG 2025-05 conditional novelty 5.0 of 10

    BenchHub is an automatically categorized, customizable LLM benchmark suite covering 303K questions across 38 benchmarks in English and Korean.

  9. Unbiased Evaluation of Large Language Models from a Causal Perspective

    cs.AI 2025-02 reject novelty 5.0 of 10

    The paper argues that perturbing benchmark questions with rule-based interventions gives a less contaminated, more interpretable evaluation of LLMs than static benchmarks or agent-generated questions.

  10. Llama-3.1-FoundationAI-SecurityLLM-8B-Instruct Technical Report

    cs.CR 2025-08 conditional novelty 4.0 of 10

    Foundation-Sec-8B-Instruct, an instruction-tuned 8B cybersecurity LLM, is released and claimed to beat Llama 3.1-8B-Instruct on CTIBench-RCM and CTIBench-MCQA while remaining competitive on general instruction-following.

  11. Can We Trust AI Benchmarks? An Interdisciplinary Review of Current Issues in AI Evaluation

    cs.AI 2025-02 conditional novelty 4.0 of 10

    A meta-review of about 110 critical studies finds nine systemic weaknesses in AI benchmarking and concludes that benchmarks are receiving disproportionate trust in AI governance.

Pith tools