Skip to content
TheChatArchiveThe Chat Archive
gpt-4o-search-preview· OpenAIHugging Face LMArena Search Arena 24K

Could you find notable papers or properly structured articles that complain against LLMs…

From a Chatbot Arena battleThe same person sent this prompt to two anonymous models at once. Compare it with gpt-4o-search-preview-high’s answer.
  1. User

    Could you find notable papers or properly structured articles that complain against LLMs benchmarks because one can do well on them through overfitting or contamination?

  2. Assistant

    Concerns have been raised about the reliability of Large Language Model (LLM) benchmarks due to issues like overfitting and data contamination. Notable discussions include:

    1. "NLP Evaluation in Trouble: On the Need to Measure LLM Data Contamination for Each Benchmark" by Sainz et al. (2023): This paper highlights the problem of LLMs being trained on test data from benchmarks, leading to inflated performance metrics. The authors advocate for community efforts to detect and mitigate data contamination in evaluations[1].

    2. "Investigating Data Contamination in Modern Benchmarks for Large Language Models" by Deng et al. (2024): This study introduces methods to detect overlaps between training data and evaluation benchmarks, revealing that some commercial LLMs can predict missing options in test sets, indicating potential contamination[2].

    3. "Benchmark Data Contamination of Large Language Models: A Survey" by Xu et al. (2024): This survey examines how LLMs inadvertently incorporate benchmark data during training, leading to unreliable performance evaluations, and explores alternative assessment methods to address this issue[3].

    4. "The Vulnerability of Language Model Benchmarks: Do They Accurately Reflect True LLM Performance?" by Banerjee et al. (2024): This paper discusses how LLMs can exploit benchmark weaknesses through overfitting and data contamination, resulting in a false perception of progress in language understanding capabilities[4].

    These works collectively emphasize the need for more robust and contamination-free evaluation methodologies to accurately assess the true capabilities of LLMs.

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.