To acquire instruction-following capabilities, large language models (LLMs) undergo instruction tuning, where they are trained on instruction-response pairs using next-token prediction (NTP). Efforts to improve instruction tuning often focus on higher-quality supervised fine-tuning (SFT) datasets, typically requiring data filtering with proprietary LLMs or human annotation. In this paper, we take a different approach by proposing SFTMix, a novel Mixup-based recipe that elevates LLM instruction tuning without relying on well-curated datasets. We observe that LLMs exhibit uneven confidence across the semantic representation space. We argue that examples with different confidence levels should play distinct roles in instruction tuning: Confident data is prone to overfitting, while unconfident data is harder to generalize. Based on this insight, SFTMix leverages training dynamics to identify examples with varying confidence levels. We then interpolate them to bridge the confidence gap and apply a Mixup-based regularization to support learning on these additional, interpolated examples. We demonstrate the effectiveness of SFTMix in both instruction-following and healthcare-specific SFT tasks, with consistent improvements across LLM families and SFT datasets of varying sizes and qualities. Extensive analyses across six directions highlight SFTMix's compatibility with data selection, adaptability to compute-constrained scenarios, and scalability to broader applications.

Despite extensive safety alignment efforts, large language models (LLMs) remain vulnerable to jailbreak attacks that elicit harmful behavior. While existing studies predominantly focus on attack methods that require technical expertise, two critical questions remain underexplored: (1) Are jailbroken responses truly useful in enabling average users to carry out harmful actions? (2) Do safety vulnerabilities exist in more common, simple human-LLM interactions? In this paper, we demonstrate that LLM responses most effectively facilitate harmful actions when they are both *actionable* and *informative*---two attributes easily elicited in multi-step, multilingual interactions. Using this insight, we propose HarmScore, a jailbreak metric that measures how effectively an LLM response enables harmful actions, and Speak Easy, a simple multi-step, multilingual attack framework. Notably, by incorporating Speak Easy into direct request and jailbreak baselines, we see an average absolute increase of $0.319$ in Attack Success Rate and $0.426$ in HarmScore in both open-source and proprietary LLMs across four safety benchmarks. Our work reveals a critical yet often overlooked vulnerability: Malicious users can easily exploit common interaction patterns for harmful intentions.

Characterizing a large language model's (LLM's) knowledge of a given question is challenging. As a result, prior work has primarily examined LLM behavior under knowledge conflicts, where the model's internal parametric memory contradicts information in the external context. However, this does not fully reflect how well the model knows the answer to the question. In this paper, we first introduce a taxonomy of five knowledge statuses based on the consistency and correctness of LLM knowledge modes. We then propose KScope, a hierarchical framework of statistical tests that progressively refines hypotheses about knowledge modes and characterizes LLM knowledge into one of these five statuses. We apply KScope to nine LLMs across four datasets and systematically establish: (1) Supporting context narrows knowledge gaps across models. (2) Context features related to difficulty, relevance, and familiarity drive successful knowledge updates. (3) LLMs exhibit similar feature preferences when partially correct or conflicted, but diverge sharply when consistently wrong. (4) Context summarization constrained by our feature analysis, together with enhanced credibility, further improves update effectiveness and generalizes across LLMs.

As the development and application of Large Language Models (LLMs) continue to advance rapidly, enhancing their trustworthiness and aligning them with human preferences has become a critical area of research. Traditional methods rely heavily on extensive data for Reinforcement Learning from Human Feedback (RLHF), but representation engineering offers a new, training-free approach. This technique leverages semantic features to control the representation of LLM's intermediate hidden states, enabling the model to meet specific requirements such as increased honesty or heightened safety awareness. However, a significant challenge arises when attempting to fulfill multiple requirements simultaneously. It proves difficult to encode various semantic contents, like honesty and safety, into a singular semantic feature, restricting its practicality. In this work, we address this challenge through Sparse Activation Control. By delving into the intrinsic mechanisms of LLMs, we manage to identify and pinpoint modules that are closely related to specific tasks within the model, i.e. attention heads. These heads display sparse characteristics that allow for near-independent control over different tasks. Our experiments, conducted on the open-source Llama series models, have yielded encouraging results. The models were able to align with human preferences on issues of safety, factualness, and bias concurrently.

Yuxin Xiao, Zecheng Zhang, Yuning Mao, Carl Yang, Jiawei Han. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Since real-world objects and their interactions are often multi-modal and multi-typed, heterogeneous networks have been widely used as a more powerful, realistic, and generic superclass of traditional homogeneous networks (graphs). Meanwhile, representation learning (a.k.a.embedding) has recently been intensively studied and shown effective for various network mining and analytical tasks. In this work, we aim to provide a unified framework to deeply summarize and evaluate existing research on heterogeneous network embedding (HNE), which includes but goes beyond a normal survey. Since there has already been a broad body of HNE algorithms, as the first contribution of this article, we provide a generic paradigm for the systematic categorization and analysis over the merits of various existing HNE algorithms. Moreover, existing HNE algorithms, though mostly claimed generic, are often evaluated on different datasets. Understandable due to the application favor of HNE, such indirect comparisons largely hinder the proper attribution of improved task performance towards effective data preprocessing and novel technical design, especially considering the various ways possible to construct a heterogeneous network from real-world application data. Therefore, as the second contribution, we create four benchmark datasets with various properties regarding scale, structure, attribute/label availability, andetc. from different sources, towards handy and fair evaluations of HNE algorithms. As the third contribution, we carefully refactor and amend the implementations and create friendly interfaces for 13 popular HNE algorithms, and provide all-around comparisons among them over multiple tasks and experimental settings. By putting all existing HNE algorithms under a unified framework, we aim to provide a universal reference and guideline for the understanding and development of HNE algorithms. Meanwhile, by open-sourcing all data and code, we envision to serve the community with an ready-to-use benchmark platform to test and compare the performance of existing and future HNE algorithms (https://github.com/yangji9181/HNE).

Some social networks provide explicit mechanisms to allocate social rewards such as reputation based on users’ actions, while the mechanism is more opaque in other networks. Nonetheless, there are always individuals who obtain greater rewards and reputation than their peers. An intuitive yet important question to ask is whether these successful users employ strategic behaviors to become influential. It might appear that the influencers ”have gamed the system.” However, it remains difficult to conclude the rationality of their actions due to factors like the combinatorial strategy space, inability to determine payoffs, and resource limitations faced by individuals. The challenging nature of this question has drawn attention from both the theory and data mining communities. Therefore, in this paper, we are motivated to investigate if resource-limited individuals discover strategic behaviors associated with high payoffs when producing collaborative/interactive content in social networks. We propose a novel framework of Dynamic Dual Attention Networks (DDAN) which models individuals’ content production strategies through a generative process, under the influence of social interactions involved in the process. Extensive experimental results illustrate the model’s effectiveness in user behavior modeling. We make three strong empirical findings: (1) Different strategies give rise to different social payoffs; (2) The best performing individuals exhibit stability in their preference over the discovered strategies, which indicates the emergence of strategic behavior; and (3) The stability of a user’s preference is correlated with high payoffs.

Heterogeneous information network (HIN) summarizes rich structural information in real-world datasets and plays an important role in many big data applications. Recently, graph neural networks have been extended to the representation learning of HIN. One very recent advancement is the hierarchical attention mechanism which incorporates both nodewise and semantic-wise attention. However, since HIN is more likely to be densely connected given its diverse types of edges, repeatedly applying graph convolutional layers can make the node embeddings indistinguishable very quickly. In order to avoid oversmoothness, existing graph neural networks targeting HIN generally suffer from a shallow structure. Consequently, those approaches ignore information beyond the local neighborhood. This design flaw violates the concept of non-local learning, which emphasizes the importance of capturing long-range dependencies. To properly address this limitation, we propose a novel framework of non-local attention in heterogeneous information networks (NLAH). Our framework utilizes a non-local attention structure to complement the hierarchical attention mechanism. In this way, it leverages both local and non-local information simultaneously. Moreover, a weighted sampling schema is designed for NLAH to reduce the computation cost for largescale datasets. Extensive experiments on three different realworld heterogeneous information networks illustrate that our framework exhibits extraordinary scalability and outperforms state-of-the-art baselines with significant margins.