Benchmark Shadows: Data Alignment, Parameter Footprints, and Generalization in Large Language Models

• URL: http://arxiv.org/abs/2604.07363v1
• Date: Wed, 01 Apr 2026 13:29:42 GMT
• Title: Benchmark Shadows: Data Alignment, Parameter Footprints, and Generalization in Large Language Models
• Abstract summary: Large language models often achieve strong benchmark gains without corresponding improvements in broader capability.<n>We find that benchmark-aligned data improves narrow evaluation metrics while limiting broader representational development.<n>We introduce parameter-space diagnostics based on spectral and rank analyses, which reveal distinct structural signatures of these regimes.
• Score: 8.94544779275887
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large language models often achieve strong benchmark gains without corresponding improvements in broader capability. We hypothesize that this discrepancy arises from differences in training regimes induced by data distribution. To investigate this, we design controlled data interventions that isolate distributional effects under fixed training settings. We find that benchmark-aligned data improves narrow evaluation metrics while limiting broader representational development, whereas coverage-expanding data leads to more distributed parameter adaptation and better generalization. We further introduce parameter-space diagnostics based on spectral and rank analyses, which reveal distinct structural signatures of these regimes. Similar patterns are observed across diverse open-source model families, including multimodal models as a key case study, suggesting that these effects extend beyond controlled settings. A case study on prompt repetition shows that not all data artifacts induce regime shifts. These results indicate that benchmark performance alone is insufficient to characterize model capability, and highlight the importance of data distribution in shaping learning dynamics.

Related papers

• What properties of reasoning supervision are associated with improved downstream model quality? [3.5354214473997225]
  We propose a suite of quantitative measures and evaluate their predictive power by fine-tuning 8B and 11B models on semantically distinct variants of a Polish reasoning dataset.<n>Our analysis reveals that these intrinsic metrics demonstrate strong and significant correlations with downstream model performance.
  arXiv  Detail & Related papers  (2026-05-13T10:04:38Z)
• Nonparametric Data Attribution for Diffusion Models [57.820618036556084]
  Data attribution for generative models seeks to quantify the influence of individual training examples on model outputs.<n>We propose a nonparametric attribution method that operates entirely on data, measuring influence via patch-level similarity between generated and training images.
  arXiv  Detail & Related papers  (2025-10-16T03:37:16Z)
• Sequential Data Augmentation for Generative Recommendation [54.765568804267645]
  Generative recommendation plays a crucial role in personalized systems, predicting users' future interactions from their historical behavior sequences.<n>Data augmentation, the process of constructing training data from user interaction histories, is a critical yet underexplored factor in training these models.<n>We propose GenPAS, a principled framework that models augmentation as a sampling process and enables flexible control of the resulting training distribution.<n>Our experiments on benchmark and industrial datasets demonstrate that GenPAS yields superior accuracy, data efficiency, and parameter efficiency compared to existing strategies.
  arXiv  Detail & Related papers  (2025-09-17T02:53:25Z)
• Prior Distribution and Model Confidence [0.0]
  We propose a framework to understand the confidence of model predictions on unseen data without the need for retraining.<n>Our approach filters out low-confidence predictions based on their distance from the training distribution in the embedding space.<n>The proposed method is model-agnostic and generalizable, with potential applications beyond computer vision.
  arXiv  Detail & Related papers  (2025-09-05T20:17:26Z)
• Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning [77.120955854093]
  We show that data diversity can be a strong predictor of generalization in language models.<n>We introduce G-Vendi, a metric that quantifies diversity via the entropy of model-induced gradients.<n>We present Prismatic Synthesis, a framework for generating diverse synthetic data.
  arXiv  Detail & Related papers  (2025-05-26T16:05:10Z)
• Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining [74.83412846804977]
  Reinforcement learning (RL)-based fine-tuning has become a crucial step in post-training language models.<n>We present a systematic end-to-end study of RL fine-tuning for mathematical reasoning by training models entirely from scratch.
  arXiv  Detail & Related papers  (2025-04-10T17:15:53Z)
• Influence Functions for Scalable Data Attribution in Diffusion Models [52.92223039302037]
  Diffusion models have led to significant advancements in generative modelling.<n>Yet their widespread adoption poses challenges regarding data attribution and interpretability.<n>We develop an influence functions framework to address these challenges.
  arXiv  Detail & Related papers  (2024-10-17T17:59:02Z)
• The Data Addition Dilemma [4.869513274920574]
  In many machine learning for healthcare tasks, standard datasets are constructed by amassing data across many, often fundamentally dissimilar, sources. But when does adding more data help, and when does it hinder progress on desired model outcomes in real-world settings? We identify this situation as the textitData Addition Dilemma, demonstrating that adding training data in this multi-source scaling context can at times result in reduced overall accuracy, uncertain fairness outcomes, and reduced worst-subgroup performance.
  arXiv  Detail & Related papers  (2024-08-08T01:42:31Z)
• Quantifying Distribution Shifts and Uncertainties for Enhanced Model Robustness in Machine Learning Applications [0.0]
  This study explores model adaptation and generalization by utilizing synthetic data. We employ quantitative measures such as Kullback-Leibler divergence, Jensen-Shannon distance, and Mahalanobis distance to assess data similarity. Our findings suggest that utilizing statistical measures, such as the Mahalanobis distance, to determine whether model predictions fall within the low-error "interpolation regime" or the high-error "extrapolation regime" provides a complementary method for assessing distribution shift and model uncertainty.
  arXiv  Detail & Related papers  (2024-05-03T10:05:31Z)
• On the Benefits of Invariance in Neural Networks [56.362579457990094]
  We show that training with data augmentation leads to better estimates of risk and thereof gradients, and we provide a PAC-Bayes generalization bound for models trained with data augmentation. We also show that compared to data augmentation, feature averaging reduces generalization error when used with convex losses, and tightens PAC-Bayes bounds.
  arXiv  Detail & Related papers  (2020-05-01T02:08:58Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.