Guardado en:
Detalles Bibliográficos
Autores principales: Zhang, Terry Jingchen, Dev, Gopal, Wang, Ning, Obreiter, Max, Pandey, Punya Syon, Samway, Keenan, Jiang, Wenyuan, Huang, Yinya, Schölkopf, Bernhard, Sachan, Mrinmaya, Jin, Zhijing
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:https://arxiv.org/abs/2509.00072
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866918498764062720
author Zhang, Terry Jingchen
Dev, Gopal
Wang, Ning
Obreiter, Max
Pandey, Punya Syon
Samway, Keenan
Jiang, Wenyuan
Huang, Yinya
Schölkopf, Bernhard
Sachan, Mrinmaya
Jin, Zhijing
author_facet Zhang, Terry Jingchen
Dev, Gopal
Wang, Ning
Obreiter, Max
Pandey, Punya Syon
Samway, Keenan
Jiang, Wenyuan
Huang, Yinya
Schölkopf, Bernhard
Sachan, Mrinmaya
Jin, Zhijing
contents Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.
format Preprint
id arxiv_https___arxiv_org_abs_2509_00072
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Test of Time: Rethinking Temporal Signal of Benchmark Contamination
Zhang, Terry Jingchen
Dev, Gopal
Wang, Ning
Obreiter, Max
Pandey, Punya Syon
Samway, Keenan
Jiang, Wenyuan
Huang, Yinya
Schölkopf, Bernhard
Sachan, Mrinmaya
Jin, Zhijing
Artificial Intelligence
Post-cutoff performance decay of LLMs has been widely interpreted as a temporal signal for benchmark contamination, where public information released before the training cutoff may have been included into training corpora and inflated model performance by memorization. We critically examine this view and demonstrate that this temporal signal is highly sensitive to how benchmark questions are constructed, even if the underlying source material remains invariant. Specifically, we show that LLM-transformed questions can produce remarkably different temporal patterns compared to fill-in-the-blank (cloze) questions directly retrieved from the very same documents. We validate this effect on prior benchmarks that report clear post-cutoff decay (LiveCodeBench), and show that a simple LLM-driven transformation of the same problems can effectively remove the temporal pattern. We further provide a mechanistic understanding of this phenomenon using influence function analysis. Overall, our results suggest that post-cutoff performance decay is a sensitive contamination signal, motivating more robust contamination probes for reliable LLM evaluation.
title Test of Time: Rethinking Temporal Signal of Benchmark Contamination
topic Artificial Intelligence
url https://arxiv.org/abs/2509.00072