Concepts or Skills? Rethinking Instruction Selection for Multi-modal Models

Vision-language instruction tuning achieves two main purposes: learning visual concepts and learning visual skills. In this paper, we found that vision-language benchmarks fall into the dichotomy of mainly benefiting from training on instructions with similar skills or visual concepts. Inspired by the discovery, we designed a simple targeted training data selection method to optimize the performance of a given benchmark. We first extract the concepts/skills from the benchmark, determine whether the benchmark predominantly benefits from similar concepts or skills, and finally select instructions with the most matching concepts/skills. Experiments on 10+ benchmarks validate the effectiveness of our targeted data selection method, showing +0.9\% over the best existing baseline averaged over all benchmarks and +1.5\% on the skill-focused subset. Our findings underscore the importance of recognizing the inherent trade-off within instruction selection, which requires balancing the acquisition of conceptual knowledge against visual skill.

DD-Ranking: Rethinking the Evaluation of Dataset Distillation

In recent years, dataset distillation has provided a reliable solution for data compression, where models trained on the resulting smaller synthetic datasets achieve performance comparable to those trained on the original datasets. To further improve the performance of synthetic datasets, various training pipelines and optimization objectives have been proposed, greatly advancing the field of dataset distillation. Recent decoupled dataset distillation methods introduce soft labels and stronger data augmentation during the post-evaluation phase and scale dataset distillation up to larger datasets (e.g., ImageNet-1K). However, this raises a question: Is accuracy still a reliable metric to fairly evaluate dataset distillation methods? Our empirical findings suggest that the performance improvements of these methods often stem from additional techniques rather than the inherent quality of the images themselves, with even randomly sampled images achieving superior results. Such misaligned evaluation settings severely hinder the development of DD. Therefore, we propose DD-Ranking, a unified evaluation framework, along with new general evaluation metrics to uncover the true performance improvements achieved by different methods. By refocusing on the actual information enhancement of distilled datasets, DD-Ranking provides a more comprehensive and fair evaluation standard for future research advancements.

Latent Video Dataset Distillation

Dataset distillation has demonstrated remarkable effectiveness in high-compression scenarios for image datasets. While video datasets inherently contain greater redundancy, existing video dataset distillation methods primarily focus on compression in the pixel space, overlooking advances in the latent space that have been widely adopted in modern text-to-image and text-to-video models. In this work, we bridge this gap by introducing a novel video dataset distillation approach that operates in the latent space using a state-of-the-art variational encoder. Furthermore, we employ a diversity-aware data selection strategy to select both representative and diverse samples. Additionally, we introduce a simple, training-free method to further compress the distilled latent dataset. By combining these techniques, our approach achieves a new state-of-the-art performance in dataset distillation, outperforming prior methods on all datasets, e.g. on HMDB51 IPC 1, we achieve a 2.6% performance increase; on MiniUCF IPC 5, we achieve a 7.8% performance increase.

Have we unified image generation and understanding yet? An empirical study of GPT-4o's image generation ability

OpenAI's multimodal GPT-4o has demonstrated remarkable capabilities in image generation and editing, yet its ability to achieve world knowledge-informed semantic synthesis--seamlessly integrating domain knowledge, contextual reasoning, and instruction adherence--remains unproven. In this study, we systematically evaluate these capabilities across three critical dimensions: (1) Global Instruction Adherence, (2) Fine-Grained Editing Precision, and (3) Post-Generation Reasoning. While existing benchmarks highlight GPT-4o's strong capabilities in image generation and editing, our evaluation reveals GPT-4o's persistent limitations: the model frequently defaults to literal interpretations of instructions, inconsistently applies knowledge constraints, and struggles with conditional reasoning tasks. These findings challenge prevailing assumptions about GPT-4o's unified understanding and generation capabilities, exposing significant gaps in its dynamic knowledge integration. Our study calls for the development of more robust benchmarks and training strategies that go beyond surface-level alignment, emphasizing context-aware and reasoning-grounded multimodal generation.

Ameliorate Spurious Correlations in Dataset Condensation

Dataset Condensation has emerged as a technique for compressing large datasets into smaller synthetic counterparts, facilitating downstream training tasks. In this paper, we study the impact of bias inside the original dataset on the performance of dataset condensation. With a comprehensive empirical evaluation on canonical datasets with color, corruption and background biases, we found that color and background biases in the original dataset will be amplified through the condensation process, resulting in a notable decline in the performance of models trained on the condensed dataset, while corruption bias is suppressed through the condensation process. To reduce bias amplification in dataset condensation, we introduce a simple yet highly effective approach based on a sample reweighting scheme utilizing kernel density estimation. Empirical results on multiple real-world and synthetic datasets demonstrate the effectiveness of the proposed method. Notably, on CMNIST with 5% bias-conflict ratio and IPC 50, our method achieves 91.5% test accuracy compared to 23.8% from vanilla DM, boosting the performance by 67.7%, whereas applying state-of-the-art debiasing method on the same dataset only achieves 53.7% accuracy. Our findings highlight the importance of addressing biases in dataset condensation and provide a promising avenue to address bias amplification in the process.

OR-Bench: An Over-Refusal Benchmark for Large Language Models

Large Language Models (LLMs) require careful safety alignment to prevent malicious outputs. While significant research focuses on mitigating harmful content generation, the enhanced safety often come with the side effect of over-refusal, where the LLMs may reject innocuous prompts and become less helpful. Although the issue of over-refusal has been empirically observed, a systematic measurement is challenging due to the difficulty of crafting prompts that appear harmful but are benign. This study proposes a novel method for automatically generating large-scale sets of ``seemingly toxic prompts'' (benign prompts likely rejected by LLMs). Leveraging this technique, we introduce OR-Bench, the first large-scale over-refusal benchmark. OR-Bench comprises 80,000 seemingly toxic prompts across 10 common rejection categories, a subset of around 1,000 hard prompts that are challenging even for state-of-the-art LLMs, and an additional 600 toxic prompts to prevent indiscriminate responses. We then conduct a comprehensive study to measure the over-refusal of 25 popular LLMs across 8 model families. Our datasets are available at https://huggingface.co/datasets/bench-llm/OR-Bench and the corresponding demo can be found at https://huggingface.co/spaces/bench-llm/or-bench. We hope this benchmark can help the community develop better safety aligned models.

Retrieval-Augmented Conversational Recommendation with Prompt-based Semi-Structured Natural Language State Tracking

Conversational recommendation (ConvRec) systems must understand rich and diverse natural language (NL) expressions of user preferences and intents, often communicated in an indirect manner (e.g., "I'm watching my weight"). Such complex utterances make retrieving relevant items challenging, especially if only using often incomplete or out-of-date metadata. Fortunately, many domains feature rich item reviews that cover standard metadata categories and offer complex opinions that might match a user's interests (e.g., "classy joint for a date"). However, only recently have large language models (LLMs) let us unlock the commonsense connections between user preference utterances and complex language in user-generated reviews. Further, LLMs enable novel paradigms for semi-structured dialogue state tracking, complex intent and preference understanding, and generating recommendations, explanations, and question answers. We thus introduce a novel technology RA-Rec, a Retrieval-Augmented, LLM-driven dialogue state tracking system for ConvRec, showcased with a video, open source GitHub repository, and interactive Google Colab notebook.

Scaling Up Dataset Distillation to ImageNet-1K with Constant Memory

Dataset distillation methods aim to compress a large dataset into a small set of synthetic samples, such that when being trained on, competitive performances can be achieved compared to regular training on the entire dataset. Among recently proposed methods, Matching Training Trajectories (MTT) achieves state-of-the-art performance on CIFAR-10/100, while having difficulty scaling to ImageNet-1k dataset due to the large memory requirement when performing unrolled gradient computation through back-propagation. Surprisingly, we show that there exists a procedure to exactly calculate the gradient of the trajectory matching loss with constant GPU memory requirement (irrelevant to the number of unrolled steps). With this finding, the proposed memory-efficient trajectory matching method can easily scale to ImageNet-1K with 6x memory reduction while introducing only around 2% runtime overhead than original MTT. Further, we find that assigning soft labels for synthetic images is crucial for the performance when scaling to larger number of categories (e.g., 1,000) and propose a novel soft label version of trajectory matching that facilities better aligning of model training trajectories on large datasets. The proposed algorithm not only surpasses previous SOTA on ImageNet-1K under extremely low IPCs (Images Per Class), but also for the first time enables us to scale up to 50 IPCs on ImageNet-1K. Our method (TESLA) achieves 27.9% testing accuracy, a remarkable +18.2% margin over prior arts.

DC-BENCH: Dataset Condensation Benchmark

Dataset Condensation is a newly emerging technique aiming at learning a tiny dataset that captures the rich information encoded in the original dataset. As the size of datasets contemporary machine learning models rely on becomes increasingly large, condensation methods become a prominent direction for accelerating network training and reducing data storage. Despite numerous methods have been proposed in this rapidly growing field, evaluating and comparing different condensation methods is non-trivial and still remains an open issue. The quality of condensed dataset are often shadowed by many critical contributing factors to the end performance, such as data augmentation and model architectures. The lack of a systematic way to evaluate and compare condensation methods not only hinders our understanding of existing techniques, but also discourages practical usage of the synthesized datasets. This work provides the first large-scale standardized benchmark on Dataset Condensation. It consists of a suite of evaluations to comprehensively reflect the generability and effectiveness of condensation methods through the lens of their generated dataset. Leveraging this benchmark, we conduct a large-scale study of current condensation methods, and report many insightful findings that open up new possibilities for future development. The benchmark library, including evaluators, baseline methods, and generated datasets, is open-sourced to facilitate future research and application.