We propose a new method to measure the task-specific accuracy of Retrieval-Augmented Large Language Models (RAG). Evaluation is performed by scoring the RAG on an automatically-generated synthetic exam composed of multiple choice questions based on the corpus of documents associated with the task. Our method is an automated, cost-efficient, interpretable, and robust strategy to select the optimal components for a RAG system. We leverage Item Response Theory (IRT) to estimate the quality of an exam and its informativeness on task-specific accuracy. IRT also provides a natural way to iteratively improve the exam by eliminating the exam questions that are not sufficiently informative about a model's ability. We demonstrate our approach on four new open-ended Question-Answering tasks based on Arxiv abstracts, StackExchange questions, AWS DevOps troubleshooting guides, and SEC filings. In addition, our experiments reveal more general insights into factors impacting RAG performance like size, retrieval mechanism, prompting and fine-tuning. Most notably, our findings show that choosing the right retrieval algorithms often leads to bigger performance gains than simply using a larger language model.
Large language models (LLMs) are widely recognized for their exceptional capacity to capture semantics meaning. Yet, there remains no established metric to quantify this capability. In this work, we introduce a quantitative metric, Information Emergence (IE), designed to measure LLMs' ability to extract semantics from input tokens. We formalize ``semantics'' as the meaningful information abstracted from a sequence of tokens and quantify this by comparing the entropy reduction observed for a sequence of tokens (macro-level) and individual tokens (micro-level). To achieve this, we design a lightweight estimator to compute the mutual information at each transformer layer, which is agnostic to different tasks and language model architectures. We apply IE in both synthetic in-context learning (ICL) scenarios and natural sentence contexts. Experiments demonstrate informativeness and patterns about semantics. While some of these patterns confirm the conventional prior linguistic knowledge, the rest are relatively unexpected, which may provide new insights.
Before deploying outputs from foundation models in high-stakes tasks, it is imperative to ensure that they align with human values. For instance, in radiology report generation, reports generated by a vision-language model must align with human evaluations before their use in medical decision-making. This paper presents Conformal Alignment, a general framework for identifying units whose outputs meet a user-specified alignment criterion. It is guaranteed that on average, a prescribed fraction of selected units indeed meet the alignment criterion, regardless of the foundation model or the data distribution. Given any pre-trained model and new units with model-generated outputs, Conformal Alignment leverages a set of reference data with ground-truth alignment status to train an alignment predictor. It then selects new units whose predicted alignment scores surpass a data-dependent threshold, certifying their corresponding outputs as trustworthy. Through applications to question answering and radiology report generation, we demonstrate that our method is able to accurately identify units with trustworthy outputs via lightweight training over a moderate amount of reference data. En route, we investigate the informativeness of various features in alignment prediction and combine them with standard models to construct the alignment predictor.
Weighted conformal prediction (WCP), a recently proposed framework, provides uncertainty quantification with the flexibility to accommodate different covariate distributions between training and test data. However, it is pointed out in this paper that the effectiveness of WCP heavily relies on the overlap between covariate distributions; insufficient overlap can lead to uninformative prediction intervals. To enhance the informativeness of WCP, we propose two methods for scenarios involving multiple sources with varied covariate distributions. We establish theoretical guarantees for our proposed methods and demonstrate their efficacy through simulations.
Active learning is considered a viable solution to alleviate the contradiction between the high dependency of deep learning-based segmentation methods on annotated data and the expensive pixel-level annotation cost of medical images. However, most existing methods suffer from unreliable uncertainty assessment and the struggle to balance diversity and informativeness, leading to poor performance in segmentation tasks. In response, we propose an efficient Predictive Accuracy-based Active Learning (PAAL) method for medical image segmentation, first introducing predictive accuracy to define uncertainty. Specifically, PAAL mainly consists of an Accuracy Predictor (AP) and a Weighted Polling Strategy (WPS). The former is an attached learnable module that can accurately predict the segmentation accuracy of unlabeled samples relative to the target model with the predicted posterior probability. The latter provides an efficient hybrid querying scheme by combining predicted accuracy and feature representation, aiming to ensure the uncertainty and diversity of the acquired samples. Extensive experiment results on multiple datasets demonstrate the superiority of PAAL. PAAL achieves comparable accuracy to fully annotated data while reducing annotation costs by approximately 50% to 80%, showcasing significant potential in clinical applications. The code is available at https://github.com/shijun18/PAAL-MedSeg.
Chain-of-Thought (CoT) reasoning often fails to faithfully reflect a language model's underlying decision process. We address this by making CoT text causally essential in a "Markovian" language model, factoring next-token prediction through an intermediate CoT and training it to predict future tokens independently of the original prompt. We formalize this via an "informativeness" objective that quantifies how much a trained CoT improves next-token predictions over a baseline. Using policy gradient, we show that Llama 3.1 8B achieves a 33.2% absolute accuracy improvement on GSM8K. Perturbation tests confirm stronger reliance on the CoT, while cross-model transfers indicate these reasoning traces generalize across interpreters. Our approach enhances both accuracy and interpretability, potentially extending CoT reasoning to arbitrarily long contexts and diverse tasks.
Deep Generative Models (DGMs) can learn flexible latent variable representations of images while avoiding intractable computations, common in Bayesian inference. However, investigating the properties of inference in Variational Autoencoders (VAEs), a major class of DGMs, reveals severe problems in their uncertainty representations. Here we draw inspiration from classical computer vision to introduce an inductive bias into the VAE by incorporating a global explaining-away latent variable, which remedies defective inference in VAEs. Unlike standard VAEs, the Explaing-Away VAE (EA-VAE) provides uncertainty estimates that align with normative requirements across a wide spectrum of perceptual tasks, including image corruption, interpolation, and out-of-distribution detection. We find that restored inference capabilities are delivered by developing a motif in the inference network (the encoder) which is widespread in biological neural networks: divisive normalization. Our results establish EA-VAEs as reliable tools to perform inference under deep generative models with appropriate estimates of uncertainty.
Large Vision-Language Models (LVLMs) suffer from hallucination issues, wherein the models generate plausible-sounding but factually incorrect outputs, undermining their reliability. A comprehensive quantitative evaluation is necessary to identify and understand the extent of hallucinations in these models. However, existing benchmarks are often limited in scope, focusing mainly on object hallucinations. Furthermore, current evaluation methods struggle to effectively address the subtle semantic distinctions between model outputs and reference data, as well as the balance between hallucination and informativeness. To address these issues, we introduce a multi-dimensional benchmark covering objects, attributes, and relations, with challenging images selected based on associative biases. Moreover, we propose a large language model (LLM)-based two-stage evaluation framework that generalizes the popular CHAIR metric and incorporates both faithfulness and coverage into the evaluation. Experiments on 10 established LVLMs demonstrate that our evaluation metric is more comprehensive and better correlated with humans than existing work when evaluating on our challenging human-annotated benchmark dataset. Our work also highlights the critical balance between faithfulness and coverage of model outputs, and encourages future works to address hallucinations in LVLMs while keeping their outputs informative.
Dataset distillation aims to compress information from a large-scale original dataset to a new compact dataset while striving to preserve the utmost degree of the original data informational essence. Previous studies have predominantly concentrated on aligning the intermediate statistics between the original and distilled data, such as weight trajectory, features, gradient, BatchNorm, etc. In this work, we consider addressing this task through the new lens of model informativeness in the compression stage on the original dataset pretraining. We observe that with the prior state-of-the-art SRe$^2$L, as model sizes increase, it becomes increasingly challenging for supervised pretrained models to recover learned information during data synthesis, as the channel-wise mean and variance inside the model are flatting and less informative. We further notice that larger variances in BN statistics from self-supervised models enable larger loss signals to update the recovered data by gradients, enjoying more informativeness during synthesis. Building on this observation, we introduce SC-DD, a simple yet effective Self-supervised Compression framework for Dataset Distillation that facilitates diverse information compression and recovery compared to traditional supervised learning schemes, further reaps the potential of large pretrained models with enhanced capabilities. Extensive experiments are conducted on CIFAR-100, Tiny-ImageNet and ImageNet-1K datasets to demonstrate the superiority of our proposed approach. The proposed SC-DD outperforms all previous state-of-the-art supervised dataset distillation methods when employing larger models, such as SRe$^2$L, MTT, TESLA, DC, CAFE, etc., by large margins under the same recovery and post-training budgets. Code is available at https://github.com/VILA-Lab/SRe2L/tree/main/SCDD/.
Large language models (LLMs) have become increasingly integrated with various applications. To ensure that LLMs do not generate unsafe responses, they are aligned with safeguards that specify what content is restricted. However, such alignment can be bypassed to produce prohibited content using a technique commonly referred to as jailbreak. Different systems have been proposed to perform the jailbreak automatically. These systems rely on evaluation methods to determine whether a jailbreak attempt is successful. However, our analysis reveals that current jailbreak evaluation methods have two limitations. (1) Their objectives lack clarity and do not align with the goal of identifying unsafe responses. (2) They oversimplify the jailbreak result as a binary outcome, successful or not. In this paper, we propose three metrics, safeguard violation, informativeness, and relative truthfulness, to evaluate language model jailbreak. Additionally, we demonstrate how these metrics correlate with the goal of different malicious actors. To compute these metrics, we introduce a multifaceted approach that extends the natural language generation evaluation method after preprocessing the response. We evaluate our metrics on a benchmark dataset produced from three malicious intent datasets and three jailbreak systems. The benchmark dataset is labeled by three annotators. We compare our multifaceted approach with three existing jailbreak evaluation methods. Experiments demonstrate that our multifaceted evaluation outperforms existing methods, with F1 scores improving on average by 17% compared to existing baselines. Our findings motivate the need to move away from the binary view of the jailbreak problem and incorporate a more comprehensive evaluation to ensure the safety of the language model.
Video summarization is a crucial research area that aims to efficiently browse and retrieve relevant information from the vast amount of video content available today. With the exponential growth of multimedia data, the ability to extract meaningful representations from videos has become essential. Video summarization techniques automatically generate concise summaries by selecting keyframes, shots, or segments that capture the video's essence. This process improves the efficiency and accuracy of various applications, including video surveillance, education, entertainment, and social media. Despite the importance of video summarization, there is a lack of diverse and representative datasets, hindering comprehensive evaluation and benchmarking of algorithms. Existing evaluation metrics also fail to fully capture the complexities of video summarization, limiting accurate algorithm assessment and hindering the field's progress. To overcome data scarcity challenges and improve evaluation, we propose an unsupervised approach that leverages video data structure and information for generating informative summaries. By moving away from fixed annotations, our framework can produce representative summaries effectively. Moreover, we introduce an innovative evaluation pipeline tailored specifically for video summarization. Human participants are involved in the evaluation, comparing our generated summaries to ground truth summaries and assessing their informativeness. This human-centric approach provides valuable insights into the effectiveness of our proposed techniques. Experimental results demonstrate that our training-free framework outperforms existing unsupervised approaches and achieves competitive results compared to state-of-the-art supervised methods.
As AI becomes more integral in our lives, the need for transparency and responsibility grows. While natural language explanations (NLEs) are vital for clarifying the reasoning behind AI decisions, evaluating them through human judgments is complex and resource-intensive due to subjectivity and the need for fine-grained ratings. This study explores the alignment between ChatGPT and human assessments across multiple scales (i.e., binary, ternary, and 7-Likert scale). We sample 300 data instances from three NLE datasets and collect 900 human annotations for both informativeness and clarity scores as the text quality measurement. We further conduct paired comparison experiments under different ranges of subjectivity scores, where the baseline comes from 8,346 human annotations. Our results show that ChatGPT aligns better with humans in more coarse-grained scales. Also, paired comparisons and dynamic prompting (i.e., providing semantically similar examples in the prompt) improve the alignment. This research advances our understanding of large language models' capabilities to assess the text explanation quality in different configurations for responsible AI development.
The scale and quality of a dataset significantly impact the performance of deep models. However, acquiring large-scale annotated datasets is both a costly and time-consuming endeavor. To address this challenge, dataset expansion technologies aim to automatically augment datasets, unlocking the full potential of deep models. Current data expansion techniques include image transformation and image synthesis methods. Transformation-based methods introduce only local variations, leading to limited diversity. In contrast, synthesis-based methods generate entirely new content, greatly enhancing informativeness. However, existing synthesis methods carry the risk of distribution deviations, potentially degrading model performance with out-of-distribution samples. In this paper, we propose DistDiff, a training-free data expansion framework based on the distribution-aware diffusion model. DistDiff constructs hierarchical prototypes to approximate the real data distribution, optimizing latent data points within diffusion models with hierarchical energy guidance. We demonstrate its capability to generate distribution-consistent samples, significantly improving data expansion tasks. DistDiff consistently enhances accuracy across a diverse range of datasets compared to models trained solely on original data. Furthermore, our approach consistently outperforms existing synthesis-based techniques and demonstrates compatibility with widely adopted transformation-based augmentation methods. Additionally, the expanded dataset exhibits robustness across various architectural frameworks. Our code is available at https://github.com/haoweiz23/DistDiff
Generalized Category Discovery (GCD) is a pragmatic and challenging open-world task, which endeavors to cluster unlabeled samples from both novel and old classes, leveraging some labeled data of old classes. Given that knowledge learned from old classes is not fully transferable to new classes, and that novel categories are fully unlabeled, GCD inherently faces intractable problems, including imbalanced classification performance and inconsistent confidence between old and new classes, especially in the low-labeling regime. Hence, some annotations of new classes are deemed necessary. However, labeling new classes is extremely costly. To address this issue, we take the spirit of active learning and propose a new setting called Active Generalized Category Discovery (AGCD). The goal is to improve the performance of GCD by actively selecting a limited amount of valuable samples for labeling from the oracle. To solve this problem, we devise an adaptive sampling strategy, which jointly considers novelty, informativeness and diversity to adaptively select novel samples with proper uncertainty. However, owing to the varied orderings of label indices caused by the clustering of novel classes, the queried labels are not directly applicable to subsequent training. To overcome this issue, we further propose a stable label mapping algorithm that transforms ground truth labels to the label space of the classifier, thereby ensuring consistent training across different active selection stages. Our method achieves state-of-the-art performance on both generic and fine-grained datasets. Our code is available at https://github.com/mashijie1028/ActiveGCD
In representation learning, a disentangled representation is highly desirable as it encodes generative factors of data in a separable and compact pattern. Researchers have advocated leveraging disentangled representations to complete downstream tasks with encouraging empirical evidence. This paper further investigates the necessity of disentangled representation in downstream applications. Specifically, we show that dimension-wise disentangled representations are unnecessary on a fundamental downstream task, abstract visual reasoning. We provide extensive empirical evidence against the necessity of disentanglement, covering multiple datasets, representation learning methods, and downstream network architectures. Furthermore, our findings suggest that the informativeness of representations is a better indicator of downstream performance than disentanglement. Finally, the positive correlation between informativeness and disentanglement explains the claimed usefulness of disentangled representations in previous works. The source code is available at https://github.com/Richard-coder-Nai/disentanglement-lib-necessity.git.
Masked Autoencoder (MAE) is a notable method for self-supervised pretraining in visual representation learning. It operates by randomly masking image patches and reconstructing these masked patches using the unmasked ones. A key limitation of MAE lies in its disregard for the varying informativeness of different patches, as it uniformly selects patches to mask. To overcome this, some approaches propose masking based on patch informativeness. However, these methods often do not consider the specific requirements of downstream tasks, potentially leading to suboptimal representations for these tasks. In response, we introduce the Multi-level Optimized Mask Autoencoder (MLO-MAE), a novel framework that leverages end-to-end feedback from downstream tasks to learn an optimal masking strategy during pretraining. Our experimental findings highlight MLO-MAE's significant advancements in visual representation learning. Compared to existing methods, it demonstrates remarkable improvements across diverse datasets and tasks, showcasing its adaptability and efficiency. Our code is available at: https://github.com/Alexiland/MLOMAE
Reward functions are central in specifying the task we want a reinforcement learning agent to perform. Given a task and desired optimal behavior, we study the problem of designing informative reward functions so that the designed rewards speed up the agent's convergence. In particular, we consider expert-driven reward design settings where an expert or teacher seeks to provide informative and interpretable rewards to a learning agent. Existing works have considered several different reward design formulations; however, the key challenge is formulating a reward informativeness criterion that adapts w.r.t. the agent's current policy and can be optimized under specified structural constraints to obtain interpretable rewards. In this paper, we propose a novel reward informativeness criterion, a quantitative measure that captures how the agent's current policy will improve if it receives rewards from a specific reward function. We theoretically showcase the utility of the proposed informativeness criterion for adaptively designing rewards for an agent. Experimental results on two navigation tasks demonstrate the effectiveness of our adaptive reward informativeness criterion.
The rapid proliferation of AI models has underscored the importance of thorough documentation, as it enables users to understand, trust, and effectively utilize these models in various applications. Although developers are encouraged to produce model cards, it's not clear how much information or what information these cards contain. In this study, we conduct a comprehensive analysis of 32,111 AI model documentations on Hugging Face, a leading platform for distributing and deploying AI models. Our investigation sheds light on the prevailing model card documentation practices. Most of the AI models with substantial downloads provide model cards, though the cards have uneven informativeness. We find that sections addressing environmental impact, limitations, and evaluation exhibit the lowest filled-out rates, while the training section is the most consistently filled-out. We analyze the content of each section to characterize practitioners' priorities. Interestingly, there are substantial discussions of data, sometimes with equal or even greater emphasis than the model itself. To evaluate the impact of model cards, we conducted an intervention study by adding detailed model cards to 42 popular models which had no or sparse model cards previously. We find that adding model cards is moderately correlated with an increase weekly download rates. Our study opens up a new perspective for analyzing community norms and practices for model documentation through large-scale data science and linguistics analysis.
In the web era, graph machine learning has been widely used on ubiquitous graph-structured data. As a pivotal component for bolstering web security and enhancing the robustness of graph-based applications, the significance of graph anomaly detection is continually increasing. While Graph Neural Networks (GNNs) have demonstrated efficacy in supervised and semi-supervised graph anomaly detection, their performance is contingent upon the availability of sufficient ground truth labels. The labor-intensive nature of identifying anomalies from complex graph structures poses a significant challenge in real-world applications. Despite that, the indirect supervision signals from other tasks (e.g., node classification) are relatively abundant. In this paper, we propose a novel MultItask acTIve Graph Anomaly deTEction framework, namely MITIGATE. Firstly, by coupling node classification tasks, MITIGATE obtains the capability to detect out-of-distribution nodes without known anomalies. Secondly, MITIGATE quantifies the informativeness of nodes by the confidence difference across tasks, allowing samples with conflicting predictions to provide informative yet not excessively challenging information for subsequent training. Finally, to enhance the likelihood of selecting representative nodes that are distant from known patterns, MITIGATE adopts a masked aggregation mechanism for distance measurement, considering both inherent features of nodes and current labeled status. Empirical studies on four datasets demonstrate that MITIGATE significantly outperforms the state-of-the-art methods for anomaly detection. Our code is publicly available at: https://github.com/AhaChang/MITIGATE.
Biased data can lead to unfair machine learning models, highlighting the importance of embedding fairness at the beginning of data analysis, particularly during dataset curation and labeling. In response, we propose Falcon, a scalable fair active learning framework. Falcon adopts a data-centric approach that improves machine learning model fairness via strategic sample selection. Given a user-specified group fairness measure, Falcon identifies samples from "target groups" (e.g., (attribute=female, label=positive)) that are the most informative for improving fairness. However, a challenge arises since these target groups are defined using ground truth labels that are not available during sample selection. To handle this, we propose a novel trial-and-error method, where we postpone using a sample if the predicted label is different from the expected one and falls outside the target group. We also observe the trade-off that selecting more informative samples results in higher likelihood of postponing due to undesired label prediction, and the optimal balance varies per dataset. We capture the trade-off between informativeness and postpone rate as policies and propose to automatically select the best policy using adversarial multi-armed bandit methods, given their computational efficiency and theoretical guarantees. Experiments show that Falcon significantly outperforms existing fair active learning approaches in terms of fairness and accuracy and is more efficient. In particular, only Falcon supports a proper trade-off between accuracy and fairness where its maximum fairness score is 1.8-4.5x higher than the second-best results.