Large language models (LLMs) excel in various capabilities but also pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment from the perspective of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose generation-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects. Experiments on multiple recent LLMs show that: (1) Safety neurons are sparse and effective. We can restore $90$% safety performance with intervention only on about $5$% of all the neurons. (2) Safety neurons encode transferrable mechanisms. They exhibit consistent effectiveness on different red-teaming datasets. The finding of safety neurons also interprets "alignment tax". We observe that the identified key neurons for safety and helpfulness significantly overlap, but they require different activation patterns of the shared neurons. Furthermore, we demonstrate an application of safety neurons in detecting unsafe outputs before generation. Our findings may promote further research on understanding LLM alignment. The source codes will be publicly released to facilitate future research.

This paper presents a novel application of explainable AI (XAI) for root-causing performance degradation in machine learning models that learn continuously from user engagement data. In such systems a single feature corruption can cause cascading feature, label and concept drifts. We have successfully applied this technique to improve the reliability of models used in personalized advertising. Performance degradation in such systems manifest as prediction anomalies in the models. These models are typically trained continuously using features that are produced by hundreds of real time data processing pipelines or derived from other upstream models. A failure in any of these pipelines or an instability in any of the upstream models can cause feature corruption, causing the model's predicted output to deviate from the actual output and the training data to become corrupted. The causal relationship between the features and the predicted output is complex, and root-causing is challenging due to the scale and dynamism of the system. We demonstrate how temporal shifts in the global feature importance distribution can effectively isolate the cause of a prediction anomaly, with better recall than model-to-feature correlation methods. The technique appears to be effective even when approximating the local feature importance using a simple perturbation-based method, and aggregating over a few thousand examples. We have found this technique to be a model-agnostic, cheap and effective way to monitor complex data pipelines in production and have deployed a system for continuously analyzing the global feature importance distribution of continuously trained models.

In-context learning (ICL) has become the default method for using large language models (LLMs), making the exploration of its limitations and understanding the underlying causes crucial. In this paper, we find that ICL falls short of handling specification-heavy tasks, which are tasks with complicated and extensive task specifications, requiring several hours for ordinary humans to master, such as traditional information extraction tasks. The performance of ICL on these tasks mostly cannot reach half of the state-of-the-art results. To explore the reasons behind this failure, we conduct comprehensive experiments on 18 specification-heavy tasks with various LLMs and identify three primary reasons: inability to specifically understand context, misalignment in task schema comprehension with humans, and inadequate long-text understanding ability. Furthermore, we demonstrate that through fine-tuning, LLMs can achieve decent performance on these tasks, indicating that the failure of ICL is not an inherent flaw of LLMs, but rather a drawback of existing alignment methods that renders LLMs incapable of handling complicated specification-heavy tasks via ICL. To substantiate this, we perform dedicated instruction tuning on LLMs for these tasks and observe a notable improvement. We hope the analyses in this paper could facilitate advancements in alignment methods enabling LLMs to meet more sophisticated human demands.

Understanding events in texts is a core objective of natural language understanding, which requires detecting event occurrences, extracting event arguments, and analyzing inter-event relationships. However, due to the annotation challenges brought by task complexity, a large-scale dataset covering the full process of event understanding has long been absent. In this paper, we introduce MAVEN-Arg, which augments MAVEN datasets with event argument annotations, making the first all-in-one dataset supporting event detection, event argument extraction (EAE), and event relation extraction. As an EAE benchmark, MAVEN-Arg offers three main advantages: (1) a comprehensive schema covering 162 event types and 612 argument roles, all with expert-written definitions and examples; (2) a large data scale, containing 98,591 events and 290,613 arguments obtained with laborious human annotation; (3) the exhaustive annotation supporting all task variants of EAE, which annotates both entity and non-entity event arguments in document level. Experiments indicate that MAVEN-Arg is quite challenging for both fine-tuned EAE models and proprietary large language models (LLMs). Furthermore, to demonstrate the benefits of an all-in-one dataset, we preliminarily explore a potential application, future event prediction, with LLMs. MAVEN-Arg and our code can be obtained from https://github.com/THU-KEG/MAVEN-Argument.

The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth of LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, and applicable evaluations. Given the importance of world knowledge to LLMs, we construct a Knowledge-oriented LLM Assessment benchmark (KoLA), in which we carefully design three crucial factors: (1) For ability modeling, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering $19$ tasks. (2) For data, to ensure fair comparisons, we use both Wikipedia, a corpus prevalently pre-trained by LLMs, along with continuously collected emerging corpora, aiming to evaluate the capacity to handle unseen data and evolving knowledge. (3) For evaluation criteria, we adopt a contrastive system, including overall standard scores for better numerical comparability across tasks and models and a unique self-contrast metric for automatically evaluating knowledge hallucination. We evaluate $21$ open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset and open-participation leaderboard are publicly released at https://kola.xlore.cn and will be continuously updated to provide references for developing LLMs and knowledge-related systems.

Multi-Armed Bandit (MAB) framework has been successfully applied in many web applications. However, many complex real-world applications that involve multiple content recommendations cannot fit into the traditional MAB setting. To address this issue, we consider an ordered combinatorial semi-bandit problem where the learner recommends S actions from a base set of K actions, and displays the results in S (out of M ) different positions. The aim is to maximize the cumulative reward with respect to the best possible subset and positions in hindsight. By the adaptation of a minimum-cost maximum-flow network, a practical algorithm based on Thompson sampling is derived for the (contextual) combinatorial problem, thus resolving the problem of computational intractability.With its potential to work with whole-page recommendation and any probabilistic models, to illustrate the effectiveness of our method, we focus on Gaussian process optimization and a contextual setting where click-through rate is predicted using logistic regression. We demonstrate the algorithms’ performance on synthetic Gaussian process problems and on large-scale news article recommendation datasets from Yahoo! Front Page Today Module.

A challenge for mining large-scale streaming data overlooked by most existing studies on online learning is the skewdistribution of examples over different classes. Many previous works have considered cost-sensitive approaches in an online setting for streaming data, where fixed costs are assigned to different classes, or ad-hoc costs are adapted based on the distribution of data received so far. However, it is not necessary for them to achieve optimal performance in terms of the measures suited for imbalanced data, such as Fmeasure, area under ROC curve (AUROC), area under precision and recall curve (AUPRC). This work proposes a general framework for online learning with imbalanced streaming data, where examples are coming sequentially and models are updated accordingly on-the-fly. By simultaneously learning multiple classifiers with different cost vectors, the proposed method can be adopted for different target measures for imbalanced data, including F-measure, AUROC and AUPRC. Moreover, we present a rigorous theoretical justification of the proposed framework for the F-measure maximization. Our empirical studies demonstrate the competitive if not better performance of the proposed method compared to previous cost-sensitive and resampling based online learning algorithms and those that are designed for optimizing certain measures.

We propose the convex factorization machine (CFM), which is a convex variant of the widely used Factorization Machines (FMs). Specifically, we employ a linear+quadratic model and regularize the linear term with the $\ell_2$-regularizer and the quadratic term with the trace norm regularizer. Then, we formulate the CFM optimization as a semidefinite programming problem and propose an efficient optimization procedure with Hazan's algorithm. A key advantage of CFM over existing FMs is that it can find a globally optimal solution, while FMs may get a poor locally optimal solution since the objective function of FMs is non-convex. In addition, the proposed algorithm is simple yet effective and can be implemented easily. Finally, CFM is a general factorization method and can also be used for other factorization problems including including multi-view matrix factorization and tensor completion problems. Through synthetic and movielens datasets, we first show that the proposed CFM achieves results competitive to FMs. Furthermore, in a toxicogenomics prediction task, we show that CFM outperforms a state-of-the-art tensor factorization method.

We consider stochastic strongly convex optimization with a complex inequality constraint. This complex inequality constraint may lead to computationally expensive projections in algorithmic iterations of the stochastic gradient descent~(SGD) methods. To reduce the computation costs pertaining to the projections, we propose an Epoch-Projection Stochastic Gradient Descent~(Epro-SGD) method. The proposed Epro-SGD method consists of a sequence of epochs; it applies SGD to an augmented objective function at each iteration within the epoch, and then performs a projection at the end of each epoch. Given a strongly convex optimization and for a total number of $T$ iterations, Epro-SGD requires only $\log(T)$ projections, and meanwhile attains an optimal convergence rate of $O(1/T)$, both in expectation and with a high probability. To exploit the structure of the optimization problem, we propose a proximal variant of Epro-SGD, namely Epro-ORDA, based on the optimal regularized dual averaging method. We apply the proposed methods on real-world applications; the empirical results demonstrate the effectiveness of our methods.

In multi-task learning, multiple related tasks are considered simultaneously, with the goal to improve the generalization performance by utilizing the intrinsic sharing of information across tasks. This paper presents a multi-task learning approach by modeling the task-feature relationships. Specifically, instead of assuming that similar tasks have similar weights on all the features, we start with the motivation that the tasks should be related in terms of subsets of features, which implies a co-cluster structure. We design a novel regularization term to capture this task-feature co-cluster structure. A proximal algorithm is adopted to solve the optimization problem. Convincing experimental results demonstrate the effectiveness of the proposed algorithm and justify the idea of exploiting the task-feature relationships.