On Calibration of LLM-based Guard Models for Reliable Content Moderation

Large language models (LLMs) pose significant risks due to the potential for generating harmful content or users attempting to evade guardrails. Existing studies have developed LLM-based guard models designed to moderate the input and output of threat LLMs, ensuring adherence to safety policies by blocking content that violates these protocols upon deployment. However, limited attention has been given to the reliability and calibration of such guard models. In this work, we empirically conduct comprehensive investigations of confidence calibration for 9 existing LLM-based guard models on 12 benchmarks in both user input and model output classification. Our findings reveal that current LLM-based guard models tend to 1) produce overconfident predictions, 2) exhibit significant miscalibration when subjected to jailbreak attacks, and 3) demonstrate limited robustness to the outputs generated by different types of response models. Additionally, we assess the effectiveness of post-hoc calibration methods to mitigate miscalibration. We demonstrate the efficacy of temperature scaling and, for the first time, highlight the benefits of contextual calibration for confidence calibration of guard models, particularly in the absence of validation sets. Our analysis and experiments underscore the limitations of current LLM-based guard models and provide valuable insights for the future development of well-calibrated guard models toward more reliable content moderation. We also advocate for incorporating reliability evaluation of confidence calibration when releasing future LLM-based guard models.

Advancing Adversarial Suffix Transfer Learning on Aligned Large Language Models

Language Language Models (LLMs) face safety concerns due to potential misuse by malicious users. Recent red-teaming efforts have identified adversarial suffixes capable of jailbreaking LLMs using the gradient-based search algorithm Greedy Coordinate Gradient (GCG). However, GCG struggles with computational inefficiency, limiting further investigations regarding suffix transferability and scalability across models and data. In this work, we bridge the connection between search efficiency and suffix transferability. We propose a two-stage transfer learning framework, DeGCG, which decouples the search process into behavior-agnostic pre-searching and behavior-relevant post-searching. Specifically, we employ direct first target token optimization in pre-searching to facilitate the search process. We apply our approach to cross-model, cross-data, and self-transfer scenarios. Furthermore, we introduce an interleaved variant of our approach, i-DeGCG, which iteratively leverages self-transferability to accelerate the search process. Experiments on HarmBench demonstrate the efficiency of our approach across various models and domains. Notably, our i-DeGCG outperforms the baseline on Llama2-chat-7b with ASRs of $43.9$ ($+22.2$) and $39.0$ ($+19.5$) on valid and test sets, respectively. Further analysis on cross-model transfer indicates the pivotal role of first target token optimization in leveraging suffix transferability for efficient searching.

Salutary Labeling with Zero Human Annotation

Active learning strategically selects informative unlabeled data points and queries their ground truth labels for model training. The prevailing assumption underlying this machine learning paradigm is that acquiring these ground truth labels will optimally enhance model performance. However, this assumption may not always hold true or maximize learning capacity, particularly considering the costly labor annotations required for ground truth labels. In contrast to traditional ground truth labeling, this paper proposes salutary labeling, which automatically assigns the most beneficial labels to the most informative samples without human annotation. Specifically, we utilize the influence function, a tool for estimating sample influence, to select newly added samples and assign their salutary labels by choosing the category that maximizes their positive influence. This process eliminates the need for human annotation. Extensive experiments conducted on nine benchmark datasets demonstrate the superior performance of our salutary labeling approach over traditional active learning strategies. Additionally, we provide several in-depth explorations and practical applications of large language model (LLM) fine-tuning.

Revisit, Extend, and Enhance Hessian-Free Influence Functions

Influence functions serve as crucial tools for assessing sample influence in model interpretation, subset training set selection, noisy label detection, and more. By employing the first-order Taylor extension, influence functions can estimate sample influence without the need for expensive model retraining. However, applying influence functions directly to deep models presents challenges, primarily due to the non-convex nature of the loss function and the large size of model parameters. This difficulty not only makes computing the inverse of the Hessian matrix costly but also renders it non-existent in some cases. Various approaches, including matrix decomposition, have been explored to expedite and approximate the inversion of the Hessian matrix, with the aim of making influence functions applicable to deep models. In this paper, we revisit a specific, albeit naive, yet effective approximation method known as TracIn. This method substitutes the inverse of the Hessian matrix with an identity matrix. We provide deeper insights into why this simple approximation method performs well. Furthermore, we extend its applications beyond measuring model utility to include considerations of fairness and robustness. Finally, we enhance TracIn through an ensemble strategy. To validate its effectiveness, we conduct experiments on synthetic data and extensive evaluations on noisy label detection, sample selection for large language model fine-tuning, and defense against adversarial attacks.

On the Inflation of KNN-Shapley Value

Shapley value-based data valuation methods, originating from cooperative game theory, quantify the usefulness of each individual sample by considering its contribution to all possible training subsets. Despite their extensive applications, these methods encounter the challenge of value inflation - while samples with negative Shapley values are detrimental, some with positive values can also be harmful. This challenge prompts two fundamental questions: the suitability of zero as a threshold for distinguishing detrimental from beneficial samples and the determination of an appropriate threshold. To address these questions, we focus on KNN-Shapley and propose Calibrated KNN-Shapley (CKNN-Shapley), which calibrates zero as the threshold to distinguish detrimental samples from beneficial ones by mitigating the negative effects of small-sized training subsets. Through extensive experiments, we demonstrate the effectiveness of CKNN-Shapley in alleviating data valuation inflation, detecting detrimental samples, and assessing data quality. We also extend our approach beyond conventional classification settings, applying it to diverse and practical scenarios such as learning with mislabeled data, online learning with stream data, and active learning for label annotation.

Outlier Gradient Analysis: Efficiently Improving Deep Learning Model Performance via Hessian-Free Influence Functions

Influence functions offer a robust framework for assessing the impact of each training data sample on model predictions, serving as a prominent tool in data-centric learning. Despite their widespread use in various tasks, the strong convexity assumption on the model and the computational cost associated with calculating the inverse of the Hessian matrix pose constraints, particularly when analyzing large deep models. This paper focuses on a classical data-centric scenario--trimming detrimental samples--and addresses both challenges within a unified framework. Specifically, we establish an equivalence transformation between identifying detrimental training samples via influence functions and outlier gradient detection. This transformation not only presents a straightforward and Hessian-free formulation but also provides profound insights into the role of the gradient in sample impact. Moreover, it relaxes the convexity assumption of influence functions, extending their applicability to non-convex deep models. Through systematic empirical evaluations, we first validate the correctness of our proposed outlier gradient analysis on synthetic datasets and then demonstrate its effectiveness in detecting mislabeled samples in vision models, selecting data samples for improving performance of transformer models for natural language processing, and identifying influential samples for fine-tuned Large Language Models.

Benchmarking Large Language Models on Communicative Medical Coaching: a Novel System and Dataset

Traditional applications of natural language processing (NLP) in healthcare have predominantly focused on patient-centered services, enhancing patient interactions and care delivery, such as through medical dialogue systems. However, the potential of NLP to benefit inexperienced doctors, particularly in areas such as communicative medical coaching, remains largely unexplored. We introduce ``ChatCoach,'' an integrated human-AI cooperative framework. Within this framework, both a patient agent and a coaching agent collaboratively support medical learners in practicing their medical communication skills during consultations. Unlike traditional dialogue systems, ChatCoach provides a simulated environment where a human doctor can engage in medical dialogue with a patient agent. Simultaneously, a coaching agent provides real-time feedback to the doctor. To construct the ChatCoach system, we developed a dataset and integrated Large Language Models such as ChatGPT and Llama2, aiming to assess their effectiveness in communicative medical coaching tasks. Our comparative analysis demonstrates that instruction-tuned Llama2 significantly outperforms ChatGPT's prompting-based approaches.

Advancing Test-Time Adaptation for Acoustic Foundation Models in Open-World Shifts

Test-Time Adaptation (TTA) is a critical paradigm for tackling distribution shifts during inference, especially in visual recognition tasks. However, while acoustic models face similar challenges due to distribution shifts in test-time speech, TTA techniques specifically designed for acoustic modeling in the context of open-world data shifts remain scarce. This gap is further exacerbated when considering the unique characteristics of acoustic foundation models: 1) they are primarily built on transformer architectures with layer normalization and 2) they deal with test-time speech data of varying lengths in a non-stationary manner. These aspects make the direct application of vision-focused TTA methods, which are mostly reliant on batch normalization and assume independent samples, infeasible. In this paper, we delve into TTA for pre-trained acoustic models facing open-world data shifts. We find that noisy, high-entropy speech frames, often non-silent, carry key semantic content. Traditional TTA methods might inadvertently filter out this information using potentially flawed heuristics. In response, we introduce a heuristic-free, learning-based adaptation enriched by confidence enhancement. Noting that speech signals' short-term consistency, we also apply consistency regularization during test-time optimization. Our experiments on synthetic and real-world datasets affirm our method's superiority over existing baselines.

Towards Informative Few-Shot Prompt with Maximum Information Gain for In-Context Learning

Large Language models (LLMs) possess the capability to engage In-context Learning (ICL) by leveraging a few demonstrations pertaining to a new downstream task as conditions. However, this particular learning paradigm suffers from high instability stemming from substantial variances induced by factors such as the input distribution of selected examples, their ordering, and prompt formats. In this work, we demonstrate that even when all these factors are held constant, the random selection of examples still results in high variance. Consequently, we aim to explore the informative ability of data examples by quantifying the Information Gain (IG) obtained in prediction after observing a given example candidate. Then we propose to sample those with maximum IG. Additionally, we identify the presence of template bias, which can lead to unfair evaluations of IG during the sampling process. To mitigate this bias, we introduce Calibration Before Sampling strategy. The experimental results illustrate that our proposed method can yield an average relative improvement of 14.3% across six classification tasks using three LLMs.

Dual Node and Edge Fairness-Aware Graph Partition

Fair graph partition of social networks is a crucial step toward ensuring fair and non-discriminatory treatments in unsupervised user analysis. Current fair partition methods typically consider node balance, a notion pursuing a proportionally balanced number of nodes from all demographic groups, but ignore the bias induced by imbalanced edges in each cluster. To address this gap, we propose a notion edge balance to measure the proportion of edges connecting different demographic groups in clusters. We analyze the relations between node balance and edge balance, then with line graph transformations, we propose a co-embedding framework to learn dual node and edge fairness-aware representations for graph partition. We validate our framework through several social network datasets and observe balanced partition in terms of both nodes and edges along with good utility. Moreover, we demonstrate our fair partition can be used as pseudo labels to facilitate graph neural networks to behave fairly in node classification and link prediction tasks.