Capacity-Dependent Effects of Data Selection for Reasoning

• URL: http://arxiv.org/abs/2608.13721v1
• Date: Thu, 13 Aug 2026 19:31:46 GMT
• Title: Capacity-Dependent Effects of Data Selection for Reasoning
• Abstract summary: We show that the value of likelihood-based data selection depends on model capacity and training duration.<n>High-likelihood data provides faster and more stable early improvements.<n>Low-likelihood data becomes increasingly beneficial for larger models when training is allowed to continue longer.
• Score: 17.988271350238705
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: In reasoning supervised fine-tuning, candidate responses for the same instruction can differ substantially in how well they match the student's current distribution. Recent likelihood-based response selection methods suggest that responses closer to the student distribution provide more effective supervision, motivating the hypothesis that high-likelihood responses may generally be preferable for fine-tuning. In this paper, we revisit this intuition and show that the value of likelihood-based data selection depends critically on model capacity and training duration. Through controlled experiments on mathematical reasoning, using students ranging from 1.5B to 8B parameters and supervision generated by stronger teacher models, we observe a clear \emph{capacity-dependent} ``{\color{SMALLCOLOR}\textbf{Fast-Fit}} / {\color{LARGECOLOR}\textbf{Slow-Gain}}'' pattern. High-likelihood data provides faster and more stable early improvements, especially for smaller models, but low-likelihood data becomes increasingly beneficial for larger models when training is allowed to continue longer. To explain this phenomenon, we analyze learning dynamics, showing that small models often fail to absorb low-likelihood supervision and instead fall into shallow or repetitive behaviors, while larger models are better able to move toward the teacher distribution under such data. We further provide a capacity-constrained theoretical view of distillation that clarifies how data difficulty, data span, and student capacity jointly govern transfer. Overall, our findings show that effective data selection for reasoning should be aware of model capacity and computing budget rather than based on a single universal preference for high-likelihood supervision.

Related papers

• Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal [29.196846051811576]
  We introduce a data-centric post-training pipeline that uses interpretability protocols to develop statistical hypotheses for the latent concepts separating preferred from dispreferred generations.<n>We show that our pipeline diagnoses undesirable signals in existing preference data, mitigates off-target learning, and can also help amplify or shape desired properties such as safeguards and model personality.
  arXiv  Detail & Related papers  (2026-06-10T17:31:16Z)
• SPaRFT: Self-Paced Reinforcement Fine-Tuning for Large Language Models [51.74498855100541]
  Large language models (LLMs) have shown strong reasoning capabilities when fine-tuned with reinforcement learning (RL)<n>We propose textbfSPaRFT, a self-paced learning framework that enables efficient learning based on the capability of the model being trained.
  arXiv  Detail & Related papers  (2025-08-07T03:50:48Z)
• The Delta Learning Hypothesis: Preference Tuning on Weak Data can Yield Strong Gains [50.66245575710432]
  We show that paired preference data consisting of individually weak data points can enable gains beyond the strength of each individual data point.<n>Our work shows that models can learn surprisingly well from paired data that might typically be considered weak.
  arXiv  Detail & Related papers  (2025-07-08T17:14:44Z)
• NaturalThoughts: Selecting and Distilling Reasoning Traces for General Reasoning Tasks [65.70224757972068]
  We select reasoning traces from a strong teacher model based on a large pool of questions from NaturalReasoning.<n>We find that simply scaling up data size with random sampling is a strong baseline with steady performance gains.<n>We find that selecting difficult examples that require more diverse reasoning strategies is more sample-efficient to transfer the teacher model's reasoning skills.
  arXiv  Detail & Related papers  (2025-07-02T17:30:24Z)
• Improving Knowledge Distillation Under Unknown Covariate Shift Through Confidence-Guided Data Augmentation [29.552309706623138]
  knowledge distillation has become an established tool for transferring knowledge from foundation models to small student networks.<n>This work addresses the common practical issue of covariate shift in knowledge distillation, where spurious features appear during training but not at test time.<n>We introduce a novel diffusion-based data augmentation strategy that generates images by maximizing the disagreement between the teacher and the student.<n>Experiments demonstrate that our approach significantly improves worst group and mean group accuracy on CelebA and SpuCo Birds as well as the spurious mAUC on spurious ImageNet.
  arXiv  Detail & Related papers  (2025-06-02T22:15:59Z)
• Data Attribution for Diffusion Models: Timestep-induced Bias in Influence Estimation [53.27596811146316]
  Diffusion models operate over a sequence of timesteps instead of instantaneous input-output relationships in previous contexts. We present Diffusion-TracIn that incorporates this temporal dynamics and observe that samples' loss gradient norms are highly dependent on timestep. We introduce Diffusion-ReTrac as a re-normalized adaptation that enables the retrieval of training samples more targeted to the test sample of interest.
  arXiv  Detail & Related papers  (2024-01-17T07:58:18Z)
• Stubborn Lexical Bias in Data and Models [50.79738900885665]
  We use a new statistical method to examine whether spurious patterns in data appear in models trained on the data. We apply an optimization approach to *reweight* the training data, reducing thousands of spurious correlations. Surprisingly, though this method can successfully reduce lexical biases in the training data, we still find strong evidence of corresponding bias in the trained models.
  arXiv  Detail & Related papers  (2023-06-03T20:12:27Z)
• Characterizing the robustness of Bayesian adaptive experimental designs to active learning bias [3.1351527202068445]
  We show that active learning bias can afflict Bayesian adaptive experimental design, depending on model misspecification. We develop an information-theoretic measure of misspecification, and show that worse misspecification implies more severe active learning bias.
  arXiv  Detail & Related papers  (2022-05-27T01:23:11Z)
• Understanding Robustness in Teacher-Student Setting: A New Perspective [42.746182547068265]
  Adrial examples are machine learning models where bounded adversarial perturbation could mislead the models to make arbitrarily incorrect predictions. Extensive studies try to explain the existence of adversarial examples and provide ways to improve model robustness. Our studies could shed light on the future exploration about adversarial examples, and enhancing model robustness via principled data augmentation.
  arXiv  Detail & Related papers  (2021-02-25T20:54:24Z)
• MixKD: Towards Efficient Distillation of Large-scale Language Models [129.73786264834894]
  We propose MixKD, a data-agnostic distillation framework, to endow the resulting model with stronger generalization ability. We prove from a theoretical perspective that under reasonable conditions MixKD gives rise to a smaller gap between the error and the empirical error. Experiments under a limited-data setting and ablation studies further demonstrate the advantages of the proposed approach.
  arXiv  Detail & Related papers  (2020-11-01T18:47:51Z)

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.