Training Data Selection

Offline preference optimization aligns reasoning models from fixed chosen--rejected pairs, yet standard methods apply gradient updates from every pair regardless of its training value under the current policy. We argue that this uniform treatment is wasteful and potentially harmful. From the perspective of gradient utility, we show that a pair's contribution depends jointly on informativeness and stability. Pair utility drifts as the policy evolves, high-gradient samples can coincide with high-curvature regions, leading to noisy and destabilizing updates, and the most effective supervision comes from stable confident errors where the model is reliably wrong yet curvature remains low. These findings motivate SAGE (Stability-Aware Gradient Efficiency), which maintains difficulty-stratified candidate pools refreshed during training and selects pairs within each pool by a forward-pass signal-to-curvature score. Only pairs with high current utility receive gradient computation; the rest are excluded from backpropagation. On mathematical reasoning benchmarks across multiple model scales, SAGE outperforms full-data and size-matched baselines while producing substantially smoother optimization trajectories.

Maintaining or improving the performance of Deep Neural Networks (DNNs) through fine-tuning requires labeling newly collected inputs, a process that is often costly and time-consuming. To alleviate this problem, input selection approaches have been developed in recent years to identify small, yet highly informative subsets for labeling. Diversity-based selection is one of the most effective approaches for this purpose. However, they are often computationally intensive and lack scalability for large input sets, limiting their practical applicability. To address this challenge, we introduce Concept-Based Diversity (CBD), a novel and highly efficient diversity metric for image inputs that leverages Vision-Language Models (VLMs). Our results show that CBD exhibits a strong correlation with Geometric Diversity (GD), an established diversity metric, while requiring only a fraction of its computation time. Building on this finding, we propose a hybrid input selection approach that combines CBD with Margin, a simple uncertainty metric. We conduct a comprehensive evaluation across a diverse set of DNN models, input sets, selection budgets, and six most effective state-of-the-art selection baselines. The results demonstrate that the CBD-based selection consistently outperforms all baselines at guiding input selection to improve the DNN model. Furthermore, the CBD-based selection approach remains highly efficient, requiring selection times close to those of simple uncertainty-based methods such as Margin, even on larger input sets like ImageNet. These results confirm not only the effectiveness and computational advantage of the CBD-based approach, particularly compared to hybrid baselines, but also its scalability in repetitive and extensive input selection scenarios.

Selecting compact training sets from heterogeneous text sources requires balancing target relevance, label reliability, and coverage. For small classifiers, discarding unfamiliar but useful texts can be as harmful as retaining annotation errors. Gradient-based selection estimates immediate target utility, whereas confidence and coverage scores capture complementary properties. Combining them with fixed weights leaves the utility sacrificed for this evidence uncontrolled across candidate pools. We propose Data Value Contribution (\DVC), which gates lexical, hidden-layer, and prediction-level evidence by neighborhood label support and calibrates its influence under a local utility budget; exact MLP gradient contractions and a monotone scoring path make the budget directly verifiable. At 20% source retention, DVC achieves 75.4 mean macro F1 across eight classification and domain-transfer tasks, exceeding the strongest baseline under shared final training by 0.7 points and random selection by 4.3; an equal-budget tuning study confirms a 0.6-point gain over calibrated quality-only selection.

Large-scale models are pretrained on massive web-crawled datasets containing documents of mixed quality, making data filtering essential. A popular method is Classifier-based Quality Filtering (CQF), which trains a binary classifier to distinguish between pretraining data and a small, high-quality set. It assigns each pretraining document a quality score defined as the classifier's score and retains only the top-scoring ones. We provide an in-depth analysis of CQF. We show that while CQF improves downstream task performance, it does not necessarily enhance language modeling on the high-quality set. Importantly, we find that training on CQF-selected data can outperform training directly on the high-quality set, even when the latter is sufficiently large. This finding alone is particularly striking, given the substantial effort and cost recently devoted to augmenting high-quality data. We explain this paradox by the fact that CQF implicitly filters the high-quality dataset as well as the low-quality one. Finally, we introduce an optimization-driven notion of data quality and demonstrate that it can be reliably estimated using small-scale proxy experiments. Altogether, our results both elucidate the mechanisms behind CQF and deepen our understanding of data selection methods widely used in practice.

Given the prohibitive cost of pre-training large language models, it is essential to leverage smaller proxy models to optimize datasets before scaling up. However, this approach becomes challenging for reasoning capabilities, which exhibit emergent behavior that only appear reliably at larger model sizes, often exceeding 7B parameters. To address this, we introduce rBridge, showing that small proxies ($\leq$1B) can effectively predict large-model reasoning by aligning more closely with (1) the pre-training objective and (2) the target task. rBridge achieves this by weighting negative log-likelihood with task alignment, using reasoning traces from frontier models as gold labels. In our experiments, rBridge (i) reduces dataset ranking costs by over 100x relative to the best baseline, (ii) achieves the strongest correlation across six reasoning benchmarks at 1B to 32B scale, and (iii) zero-shot transfers predictive relationships across pre-training datasets at 1B to 7B scale. These findings indicate that rBridge offers a practical path for exploring reasoning-oriented pre-training at lower cost.

Visual grounding aims to localize the image regions based on a textual query. Given the difficulty of large-scale data curation, we investigate how to effectively learn visual grounding under data-scarce settings in this paper. To address the data scarcity, we propose a novel framework, POBF (Paint Outside the Box and Filter). POBF synthesizes images by inpainting outside the box, tackling a label misalignment issue encountered in previous works. Furthermore, POBF leverages an innovative filtering scheme to select the most effective training data. This scheme combines a hardness score and an overfitting score, balanced by a penalty term. Extensive experiments across four benchmark datasets demonstrate that POBF consistently improves performance, achieving an average gain of 5.83% over the real-data-only method and outperforming leading baselines by 2.29%-3.85% in accuracy. Additionally, we validate the robustness and generalizability of POBF across various generative models, training data sizes, and model architectures.

As AI-generated content increasingly populates the web, generative AI models are at growing risk of being trained on their own outputs, a process known as AI autophagy. This feedback loop has been shown to induce model collapse, typically characterized by a loss of diversity in generated content. However, existing work offers a limited understanding of this phenomenon and relies on mitigation strategies that assume access to human-authored data. In this paper, we conduct extensive simulations across multiple datasets and LLMs to address key gaps in the study of model collapse. First, we introduce model-intrinsic measures based on next-token probability distributions, showing that model collapse corresponds to an increasing concentration of probability mass on a small set of tokens. Second, we demonstrate that model collapse is also associated with a loss of common sense, as measured by a decline in commonsense inference accuracy. Third, we identify perplexity (a measure of model "surprise") as a key driver of collapse: fine-tuning on the least "surprising" documents leads to more severe degeneration. Building on this insight, we propose a perplexity-based filtering strategy that prioritizes high-surprise documents during fine-tuning. Unlike existing approaches, our method does not require distinguishing between human-authored and AI-generated content. Across datasets and LLM families, this strategy consistently mitigates model collapse, achieving performance comparable to, and in some cases better than, human-data baselines, while substantially reducing the concentration of next-token probabilities. Overall, our results provide a unified, model-centric understanding of model collapse and suggest practical, scalable strategies for training generative AI systems in increasingly synthetic environments.