Supervised Fine-Tuning (SFT) serves as a crucial phase in aligning Large Language Models (LLMs) to specific task prerequisites. The selection of fine-tuning data profoundly influences the model's performance, whose principle is traditionally grounded in data quality and distribution. In this paper, we introduce a new dimension in SFT data selection: learnability. This new dimension is motivated by the intuition that SFT unlocks capabilities acquired by a LLM during the pretraining phase. Given that different pretrained models have disparate capabilities, the SFT data appropriate for one may not suit another. Thus, we introduce the term learnability to define the suitability of data for effective learning by the model. We present the Loss Based SFT Data Selection (LoBaSS) method, utilizing data learnability as the principal criterion for the selection SFT data. This method provides a nuanced approach, allowing the alignment of data selection with inherent model capabilities, ensuring optimal compatibility and learning efficiency. In experimental comparisons involving 7B and 13B models, our LoBaSS method is able to surpass full-data fine-tuning at merely 6% of the total training data. When employing 16.7% of the data, LoBaSS harmonizes the model's capabilities across conversational and mathematical domains, proving its efficacy and adaptability.

Recent years have seen considerable advancements in multi-step reasoning with Large Language Models (LLMs). The previous studies have elucidated the merits of integrating feedback or search mechanisms during model inference to improve the reasoning accuracy. The Process-Supervised Reward Model (PRM), typically furnishes LLMs with step-by-step feedback during the training phase, akin to Proximal Policy Optimization (PPO) or reject sampling. Our objective is to examine the efficacy of PRM in the inference phase to help discern the optimal solution paths for multi-step tasks such as mathematical reasoning and code generation. To this end, we propose a heuristic greedy search algorithm that employs the step-level feedback from PRM to optimize the reasoning pathways explored by LLMs. This tailored PRM demonstrated enhanced results compared to the Chain of Thought (CoT) on mathematical benchmarks like GSM8K and MATH. Additionally, to explore the versatility of our approach, we develop a novel method to automatically generate step-level reward dataset for coding tasks and observed similar improved performance in the code generation tasks. In the exciting evolution of Large Language Models (LLMs) such as GPT (OpenAI, 2023; Brown et al., 2020), LLaMA (Touvron et al., 2023a;b), OPT (Zhang et al., 2022a), Falcon (Penedo et al., 2023), and PaLM (Anil et al., 2023; Chowdhery et al., 2022), a consistent ability to handle tasks from conversation to text generation has been evident. However, when it comes to reasoning, especially multi-step reasoning, current LLMs, even with sophisticated prompting techniques like the Chain of Thought (CoT)(Wei et al., 2023), are still prone to a cascade of errors in their generation processes. As the number of reasoning steps increases, these LLMs face challenges in providing and integrating effective feedback, resulting in one error leading to another. Achieving a refined multi-step reasoning capability for LLMs can unlock their potential across an even broader array of applications, ranging from complex problem-solving to high-level intellectual tasks.

Training on large-scale graphs has achieved remarkable results in graph representation learning, but its cost and storage have attracted increasing concerns. Existing graph condensation methods primarily focus on optimizing the feature matrices of condensed graphs while overlooking the impact of the structure information from the original graphs. To investigate the impact of the structure information, we conduct analysis from the spectral domain and empirically identify substantial Laplacian Energy Distribution (LED) shifts in previous works. Such shifts lead to poor performance in cross-architecture generalization and specific tasks, including anomaly detection and link prediction. In this paper, we propose a novel Structure-broadcasting Graph Dataset Distillation (SGDD) scheme for broadcasting the original structure information to the generation of the synthetic one, which explicitly prevents overlooking the original structure information. Theoretically, the synthetic graphs by SGDD are expected to have smaller LED shifts than previous works, leading to superior performance in both cross-architecture settings and specific tasks. We validate the proposed SGDD across 9 datasets and achieve state-of-the-art results on all of them: for example, on the YelpChi dataset, our approach maintains 98.6% test accuracy of training on the original graph dataset with 1,000 times saving on the scale of the graph. Moreover, we empirically evaluate there exist 17.6% ~ 31.4% reductions in LED shift crossing 9 datasets. Extensive experiments and analysis verify the effectiveness and necessity of the proposed designs. The code is available in the GitHub repository: https://github.com/RingBDStack/SGDD.

Learning to predict masked tokens in a sequence has been shown to be a powerful pretraining objective for large language models. After training, such masked language models can provide distributions of tokens conditioned on bidirectional context. In this paper, we show that contrary to popular assumptions, such bidirectional conditionals often demonstrate considerable inconsistencies, i.e., they cannot be derived from a coherent joint distribution when considered together. We empirically quantify such inconsistencies in the simple scenario of bigram comparison for two common styles of masked language models: T5-style and BERT-style. For example, we show that T5 models often confuse their own preference regarding two similar bigrams. We show that inconsistencies exist ubiquitously in masked language models of diverse sizes and configurations, from RoBERTa-base to GLM-130B. As an initial attempt to address this issue during the inference phase, we propose Ensemble of Conditionals, a self-ensemble algorithm that jointly considers many inconsistent conditionals directly produced by the MLM to synthesize a distribution that is used as the model's final output. Such ensembling improves open-source SOTA results on LAMBADA.

Recent research has highlighted the importance of dataset size in scaling language models. However, large language models (LLMs) are notoriously token-hungry during pre-training, and high-quality text data on the web is likely to be approaching its scaling limit for LLMs. To further enhance LLMs, a straightforward approach is to repeat the pre-training data for additional epochs. In this study, we empirically investigate three key aspects under this approach. First, we explore the consequences of repeating pre-training data, revealing that the model is susceptible to overfitting, leading to multi-epoch degradation. Second, we examine the key factors contributing to multi-epoch degradation, finding that significant factors include dataset size, model parameters, and training objectives, while less influential factors consist of dataset quality and model FLOPs. Finally, we explore whether widely used regularization can alleviate multi-epoch degradation. Most regularization techniques do not yield significant improvements, except for dropout, which demonstrates remarkable effectiveness but requires careful tuning when scaling up the model size. Additionally, we discover that leveraging mixture-of-experts (MoE) enables cost-effective and efficient hyper-parameter tuning for computationally intensive dense LLMs with comparable trainable parameters, potentially impacting efficient LLM development on a broader scale.

Unsupervised contrastive learning methods have recently seen significant improvements, particularly through data augmentation strategies that aim to produce robust and generalizable representations. However, prevailing data augmentation methods, whether hand designed or based on foundation models, tend to rely heavily on prior knowledge or external data. This dependence often compromises their effectiveness and efficiency. Furthermore, the applicability of most existing data augmentation strategies is limited when transitioning to other research domains, especially science-related data. This limitation stems from the paucity of prior knowledge and labeled data available in these domains. To address these challenges, we introduce DiffAug-a novel and efficient Diffusion-based data Augmentation technique. DiffAug aims to ensure that the augmented and original data share a smoothed latent space, which is achieved through diffusion steps. Uniquely, unlike traditional methods, DiffAug first mines sufficient prior semantic knowledge about the neighborhood. This provides a constraint to guide the diffusion steps, eliminating the need for labels, external data/models, or prior knowledge. Designed as an architecture-agnostic framework, DiffAug provides consistent improvements. Specifically, it improves image classification and clustering accuracy by 1.6%~4.5%. When applied to biological data, DiffAug improves performance by up to 10.1%, with an average improvement of 5.8%. DiffAug shows good performance in both vision and biological domains.

Object pose estimation constitutes a critical area within the domain of 3D vision. While contemporary state-of-the-art methods that leverage real-world pose annotations have demonstrated commendable performance, the procurement of such real-world training data incurs substantial costs. This paper focuses on a specific setting wherein only 3D CAD models are utilized as a priori knowledge, devoid of any background or clutter information. We introduce a novel method, CPPF++, designed for sim-to-real pose estimation. This method builds upon the foundational point-pair voting scheme of CPPF, reconceptualizing it through a probabilistic lens. To address the challenge of voting collision, we model voting uncertainty by estimating the probabilistic distribution of each point pair within the canonical space. This approach is further augmented by iterative noise filtering, employed to eradicate votes associated with backgrounds or clutters. Additionally, we enhance the context provided by each voting unit by introducing $N$-point tuples. In conjunction with this methodological contribution, we present a new category-level pose estimation dataset, DiversePose 300. This dataset is specifically crafted to facilitate a more rigorous evaluation of current state-of-the-art methods, encompassing a broader and more challenging array of real-world scenarios. Empirical results substantiate the efficacy of our proposed method, revealing a significant reduction in the disparity between simulation and real-world performance.

State-of-the-art deep neural networks are trained with large amounts (millions or even billions) of data. The expensive computation and memory costs make it difficult to train them on limited hardware resources, especially for recent popular large language models (LLM) and computer vision models (CV). Recent popular dataset distillation methods are thus developed, aiming to reduce the number of training samples via synthesizing small-scale datasets via gradient matching. However, as the gradient calculation is coupled with the specific network architecture, the synthesized dataset is biased and performs poorly when used for training unseen architectures. To address these limitations, we present dataset quantization (DQ), a new framework to compress large-scale datasets into small subsets which can be used for training any neural network architectures. Extensive experiments demonstrate that DQ is able to generate condensed small datasets for training unseen network architectures with state-of-the-art compression ratios for lossless model training. To the best of our knowledge, DQ is the first method that can successfully distill large-scale datasets such as ImageNet-1k with a state-of-the-art compression ratio. Notably, with 60% data from ImageNet and 20% data from Alpaca's instruction tuning data, the models can be trained with negligible or no performance drop for both vision tasks (including classification, semantic segmentation, and object detection) as well as language tasks (including instruction tuning tasks such as BBH and DROP).

Despite Graph Neural Networks demonstrating considerable promise in graph representation learning tasks, GNNs predominantly face significant issues with over-fitting and over-smoothing as they go deeper as models of computer vision realm. In this work, we conduct a systematic study of deeper GNN research trajectories. Our findings indicate that the current success of deep GNNs primarily stems from (I) the adoption of innovations from CNNs, such as residual/skip connections, or (II) the tailor-made aggregation algorithms like DropEdge. However, these algorithms often lack intrinsic interpretability and indiscriminately treat all nodes within a given layer in a similar manner, thereby failing to capture the nuanced differences among various nodes. To this end, we introduce the Snowflake Hypothesis -- a novel paradigm underpinning the concept of ``one node, one receptive field''. The hypothesis draws inspiration from the unique and individualistic patterns of each snowflake, proposing a corresponding uniqueness in the receptive fields of nodes in the GNNs. We employ the simplest gradient and node-level cosine distance as guiding principles to regulate the aggregation depth for each node, and conduct comprehensive experiments including: (1) different training schemes; (2) various shallow and deep GNN backbones, and (3) various numbers of layers (8, 16, 32, 64) on multiple benchmarks (six graphs including dense graphs with millions of nodes); (4) compare with different aggregation strategies. The observational results demonstrate that our hypothesis can serve as a universal operator for a range of tasks, and it displays tremendous potential on deep GNNs. It can be applied to various GNN frameworks, enhancing its effectiveness when operating in-depth, and guiding the selection of the optimal network depth in an explainable and generalizable way.

Continual learning (CL) can help pre-trained vision-language models efficiently adapt to new or under-trained data distributions without re-training. Nevertheless, during the continual training of the Contrastive Language-Image Pre-training (CLIP) model, we observe that the model's zero-shot transfer ability significantly degrades due to catastrophic forgetting. Existing CL methods can mitigate forgetting by replaying previous data. However, since the CLIP dataset is private, replay methods cannot access the pre-training dataset. In addition, replaying data of previously learned downstream tasks can enhance their performance but comes at the cost of sacrificing zero-shot performance. To address this challenge, we propose a novel method ZSCL to prevent zero-shot transfer degradation in the continual learning of vision-language models in both feature and parameter space. In the feature space, a reference dataset is introduced for distillation between the current and initial models. The reference dataset should have semantic diversity but no need to be labeled, seen in pre-training, or matched image-text pairs. In parameter space, we prevent a large parameter shift by averaging weights during the training. We propose a more challenging Multi-domain Task Incremental Learning (MTIL) benchmark to evaluate different methods, where tasks are from various domains instead of class-separated in a single dataset. Our method outperforms other methods in the traditional class-incremental learning setting and the MTIL by 9.7% average score. Our code locates at https://github.com/Thunderbeee/ZSCL.