Challenge: a large number of large language models (LLMs) are being evaluated for their performance, but their reliability is threatened by test set contamination.
Approach: They propose a framework that decontaminates large language models by applying small perturbations to the input embedding space.
Outcome: The proposed framework achieves strong decontamination effectiveness while incurring minimal degradation in benign utility.

Similar Papers

Inference-Time Decontamination: Reusing Leaked Benchmarks for Large Language Model Evaluation (2024.findings-emnlp)

Copied to clipboard

Challenge: Current LLMs are achieving better performance on various benchmarks, but their performance in practical applications does not always match their benchmark results.
Approach: They propose to detect and rewrite leaked benchmarks without altering their difficulties by using Inference-Time Decontamination (ITD) to mitigate performance inflation caused by memorizing leaked samples.
Outcome: The proposed method reduces inflated accuracy by 22.9% on GSM8K and 19.0% on MMLU.
Investigating Data Contamination in Modern Benchmarks for Large Language Models (2024.naacl-long)

Copied to clipboard

Challenge: Existing evaluation benchmarks for large language models are inflated and inconsistent with actual performance.
Approach: They propose a retrieval-based system to explore potential overlaps between benchmarks and pretraining corpora and a protocol to investigate testset slot guessing.
Outcome: The proposed method exploits overlaps between evaluation benchmarks and pretraining corpora and masks a wrong answer in a multiple choice question and prompts the model to fill in the gap.
Generalization or Memorization: Data Contamination and Trustworthy Evaluation for Large Language Models (2024.findings-acl)

Copied to clipboard

Challenge: Considering the vast size and wide-ranging sources of LLMs’ training data, it could explicitly or implicitly include test data.
Approach: They propose a Contamination Detection via output Distribution (CDD) which detects data contamination only by identifying the peakedness of LLM's output distribution.
Outcome: The proposed method improves performance by 21.8%-30.2% on humanEval and TED: trustworthy evaluation via output distribution.
NLP Evaluation in trouble: On the Need to Measure LLM Data Contamination for each Benchmark (2023.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for evaluating large language models using annotated benchmarks are in trouble . data contamination can cause wrong scientific conclusions being published .
Approach: They argue that the evaluation of NLP tasks using annotated benchmarks is in trouble . they define different levels of data contamination and propose a community effort .
Outcome: The proposed measures should detect when data from a benchmark was exposed to a model and flag papers with conclusions compromised by data contamination.
DCR: Quantifying Data Contamination in LLMs Evaluation (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) memorize evaluation data during training, inflating performance metrics and undermining genuine generalization assessment.
Approach: They propose a framework to detect and quantify benchmark data contamination (BDC) by synthesizing contamination scores via a fuzzy inference system.
Outcome: The proposed framework detects and quantifies BDC risk across semantic, informational, data, and label levels.
Benchmarking Large Language Models Under Data Contamination: A Survey from Static to Dynamic Evaluation (2025.emnlp-main)

Copied to clipboard

Challenge: In the era of evaluating large language models, data contamination is an increasingly prominent concern . static benchmarking has been used for evaluation, but there are limitations of *dynamic* benchmarks .
Approach: They propose a series of optimal design principles for *dynamic* benchmarking and analyze the limitations of existing *static* benchmarks.
Outcome: The proposed benchmarks highlight a critical gap in the evaluation of LLMs.
Quantifying Contamination in Evaluating Code Generation Capabilities of Language Models (2024.acl-long)

Copied to clipboard

Challenge: Recent studies have shown that large language models are contaminated with data from pretraining and finetuning tasks.
Approach: They perform extensive analysis on the factors that affect model memorization and generalization, such as model size, problem difficulty, and question length.
Outcome: The results show that models perform better on the subset of the benchmarks where similar solutions are seen during training.
AntiLeakBench: Preventing Data Contamination by Automatically Constructing Benchmarks with Updated Real-World Knowledge (2025.acl-long)

Copied to clipboard

Challenge: Existing studies solve this challenge by updating benchmarks with newly collected data, but they fail to guarantee contamination-free evaluation as the newly collected knowledge may contain pre-existing knowledge.
Approach: They propose an automated anti-leakage benchmarking framework that builds and updates benchmarks without human labor instead of using newly collected data.
Outcome: The proposed framework significantly reduces the cost of benchmark maintenance to accommodate emerging LLMs.
Data Contamination Can Cross Language Barriers (2024.emnlp-main)

Copied to clipboard

Challenge: Existing methods to detect contamination of public benchmarks are too superficial to reflect deeper forms of contamination.
Approach: They propose generalization-based approaches to unmask a cross-lingual form of contamination that inflates LLMs’ performance while evading current detection methods.
Outcome: The proposed model outperforms existing detection methods while avoiding contamination of public benchmarks in the pre-training data.
Forget What You Know about LLMs Evaluations - LLMs are Like a Chameleon (2025.emnlp-main)

Copied to clipboard

Challenge: Large language models (LLMs) excel on public benchmarks, but high scores may mask overreliance on dataset-specific surface cues rather than true language understanding.
Approach: They propose a meta-evaluation framework that systematically rephrases benchmark inputs to detect overfitting.
Outcome: The proposed framework detects performance degradation indicative of superficial pattern reliance on dataset-specific cues and distortion levels.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations