Recent years have witnessed a rapid growth of recommender systems, providing suggestions in numerous applications with potentially high social impact, such as health or justice. Meanwhile, in Europe, the upcoming AI Act mentions \emph{transparency} as a requirement for critical AI systems in order to ``mitigate the risks to fundamental rights''. Post-hoc explanations seamlessly align with this goal and extensive literature on the subject produced several forms of such objects, graphs being one of them. Early studies in visualization demonstrated the graphs' ability to improve user understanding, positioning them as potentially ideal explanations. However, it remains unclear how graph-based explanations compare to other explanation designs. In this work, we aim to determine the effectiveness of graph-based explanations in improving users' perception of AI-based recommendations using a mixed-methods approach. We first conduct a qualitative study to collect users' requirements for graph explanations. We then run a larger quantitative study in which we evaluate the influence of various explanation designs, including enhanced graph-based ones, on aspects such as understanding, usability and curiosity toward the AI system. We find that users perceive graph-based explanations as more usable than designs involving feature importance. However, we also reveal that textual explanations lead to higher objective understanding than graph-based designs. Most importantly, we highlight the strong contrast between participants' expressed preferences for graph design and their actual ratings using it, which are lower compared to textual design. These findings imply that meeting stakeholders' expressed preferences might not alone guarantee ``good'' explanations. Therefore, crafting hybrid designs successfully balancing social expectations with downstream performance emerges as a significant challenge.

Although the spread of behaviors is influenced by many social factors, existing literature tends to study the effects of single factors -- most often, properties of the social network -- on the final cascade. In order to move towards a more integrated view of cascades, this paper offers the first comprehensive investigation into the role of two social factors in the diffusion of 1,337 popular hashtags representing the production of novel culture on Twitter: 1) the topology of the Twitter social network and 2) performance of each user's probable demographic identity. Here, we show that cascades are best modeled using a combination of network and identity, rather than either factor alone. This combined model best reproduces a composite index of ten cascade properties across all 1,337 hashtags. However, there is important heterogeneity in what social factors are required to reproduce different properties of hashtag cascades. For instance, while a combined network+identity model best predicts the popularity of cascades, a network-only model has better performance in predicting cascade growth and an identity-only model in adopter composition. We are able to predict what type of hashtag is best modeled by each combination of features and use this to further improve performance. Additionally, consistent with prior literature on the combined network+identity model most outperforms the single-factor counterfactuals among hashtags used for expressing racial or regional identity, stance-taking, talking about sports, or variants of existing cultural trends with very slow- or fast-growing communicative need. In sum, our results imply the utility of multi-factor models in predicting cascades, in order to account for the varied ways in which network, identity, and other social factors play a role in the diffusion of hashtags on Twitter.

Conversational search seeks to retrieve relevant passages for the given questions in Conversational QA (ConvQA). Questions in ConvQA face challenges such as omissions and coreferences, making it difficult to obtain desired search results. Conversational Query Reformulation (CQR) transforms these current queries into de-contextualized forms to resolve these issues. However, existing CQR methods focus on rewriting human-friendly queries, which may not always yield optimal search results for the retriever. To overcome this challenge, we introduce GuideCQR, a framework that utilizes guided documents to refine queries, ensuring that they are optimal for retrievers. Specifically, we augment keywords, generate expected answers from the re-ranked documents, and unify them with the filtering process. Experimental results show that queries enhanced by guided documents outperform previous CQR methods. Especially, GuideCQR surpasses the performance of Large Language Model (LLM) prompt-powered approaches and demonstrates the importance of the guided documents in formulating retriever-friendly queries across diverse setups.

The advances of large foundation models necessitate wide-coverage, low-cost, and zero-contamination benchmarks. Despite continuous exploration of language model evaluations, comprehensive studies on the evaluation of Large Multi-modal Models (LMMs) remain limited. In this work, we introduce LMMS-EVAL, a unified and standardized multimodal benchmark framework with over 50 tasks and more than 10 models to promote transparent and reproducible evaluations. Although LMMS-EVAL offers comprehensive coverage, we find it still falls short in achieving low cost and zero contamination. To approach this evaluation trilemma, we further introduce LMMS-EVAL LITE, a pruned evaluation toolkit that emphasizes both coverage and efficiency. Additionally, we present Multimodal LIVEBENCH that utilizes continuously updating news and online forums to assess models' generalization abilities in the wild, featuring a low-cost and zero-contamination evaluation approach. In summary, our work highlights the importance of considering the evaluation trilemma and provides practical solutions to navigate the trade-offs in evaluating large multi-modal models, paving the way for more effective and reliable benchmarking of LMMs. We opensource our codebase and maintain leaderboard of LIVEBENCH at https://github.com/EvolvingLMMs-Lab/lmms-eval and https://huggingface.co/spaces/lmms-lab/LiveBench.

Understanding and making use of audience feedback is important but difficult for journalists, who now face an impractically large volume of audience comments online. We introduce AudienceView, an online tool to help journalists categorize and interpret this feedback by leveraging large language models (LLMs). AudienceView identifies themes and topics, connects them back to specific comments, provides ways to visualize the sentiment and distribution of the comments, and helps users develop ideas for subsequent reporting projects. We consider how such tools can be useful in a journalist's workflow, and emphasize the importance of contextual awareness and human judgment.

Language models are known to encode a great amount of factual knowledge through pretraining. However, such knowledge might be insufficient to cater to user requests, requiring the model to integrate external knowledge sources and adhere to user-provided specifications. When answering questions about ongoing events, the model should use recent news articles to update its response; when asked to provide recommendations, the model should prioritize user specifications over retrieved product reviews; when some facts are edited in the model, the updated facts should override all prior knowledge learned by the model even if they are conflicting. In all of the cases above, the model faces a decision between its own parametric knowledge, (retrieved) contextual knowledge, and user instruction knowledge. In this paper, we (1) unify such settings into the problem of knowledge preference and define a three-level preference hierarchy over these knowledge sources; (2) compile a collection of existing datasets IfQA, MQuAKE, and MRQA covering a combination of settings (with/without user specifications, with/without context documents) to systematically evaluate how well models obey the intended knowledge preference; and (3) propose a dataset synthesis method that composes diverse question-answer pairs with user assumptions and related context to directly fine-tune LMs for instilling the hierarchy of knowledge. We demonstrate that a 7B model, fine-tuned on only a few thousand examples automatically generated by our proposed method, effectively achieves superior performance (more than 18% improvement across all evaluation benchmarks) in adhering to the desired knowledge preference hierarchy.

Large language models (LLMs) frequently generate non-factual content, known as hallucinations. Existing retrieval-augmented-based hallucination detection approaches typically address this by framing it as a classification task, evaluating hallucinations based on their consistency with retrieved evidence. However, this approach usually lacks detailed explanations for these evaluations and does not assess the reliability of these explanations. Furthermore, deficiencies in retrieval systems can lead to irrelevant or partially relevant evidence retrieval, impairing the detection process. Moreover, while real-world hallucination detection requires analyzing multiple pieces of evidence, current systems usually treat all evidence uniformly without considering its relevance to the content. To address these challenges, we introduce Halu-J, a critique-based hallucination judge with 7 billion parameters. Halu-J enhances hallucination detection by selecting pertinent evidence and providing detailed critiques. Our experiments indicate that Halu-J outperforms GPT-4o in multiple-evidence hallucination detection and matches its capability in critique generation and evidence selection. We also introduce ME-FEVER, a new dataset designed for multiple-evidence hallucination detection. Our code and dataset can be found in https://github.com/GAIR-NLP/factool .

This survey explores the burgeoning field of role-playing with language models, focusing on their development from early persona-based models to advanced character-driven simulations facilitated by Large Language Models (LLMs). Initially confined to simple persona consistency due to limited model capabilities, role-playing tasks have now expanded to embrace complex character portrayals involving character consistency, behavioral alignment, and overall attractiveness. We provide a comprehensive taxonomy of the critical components in designing these systems, including data, models and alignment, agent architecture and evaluation. This survey not only outlines the current methodologies and challenges, such as managing dynamic personal profiles and achieving high-level persona consistency but also suggests avenues for future research in improving the depth and realism of role-playing applications. The goal is to guide future research by offering a structured overview of current methodologies and identifying potential areas for improvement. Related resources and papers are available at https://github.com/nuochenpku/Awesome-Role-Play-Papers.

Graph Neural Networks (GNNs) have recently gained traction in symbolic music tasks, yet a lack of a unified framework impedes progress. Addressing this gap, we present GraphMuse, a graph processing framework and library that facilitates efficient music graph processing and GNN training for symbolic music tasks. Central to our contribution is a new neighbor sampling technique specifically targeted toward meaningful behavior in musical scores. Additionally, GraphMuse integrates hierarchical modeling elements that augment the expressivity and capabilities of graph networks for musical tasks. Experiments with two specific musical prediction tasks -- pitch spelling and cadence detection -- demonstrate significant performance improvement over previous methods. Our hope is that GraphMuse will lead to a boost in, and standardization of, symbolic music processing based on graph representations. The library is available at https://github.com/manoskary/graphmuse

2023 was the year the world woke up to generative AI, and 2024 is the year policymakers are responding more firmly. Importantly, this policy momentum is taking place alongside real world creation and distribution of synthetic media. Social media platforms, news organizations, dating apps, image generation companies, and more are already navigating a world of AI-generated visuals and sounds, already changing hearts and minds, as policymakers try to catch up. How, then, can AI governance capture the complexity of the synthetic media landscape? How can it attend to synthetic media's myriad uses, ranging from storytelling to privacy preservation, to deception, fraud, and defamation, taking into account the many stakeholders involved in its development, creation, and distribution? And what might it mean to govern synthetic media in a manner that upholds the truth while bolstering freedom of expression? What follows is the first known collection of diverse examples of the implementation of synthetic media governance that responds to these questions, specifically through Partnership on AI's (PAI) Responsible Practices for Synthetic Media - a voluntary, normative Framework for creating, distributing, and building technology for synthetic media responsibly, launched in February 2023. In this paper, we present a case bank of real world examples that help operationalize the Framework - highlighting areas synthetic media governance can be applied, augmented, expanded, and refined for use, in practice. Read together, the cases emphasize distinct elements of AI policymaking and seven emergent best practices supporting transparency, safety, expression, and digital dignity online: consent, disclosure, and differentiation between harmful and creative use cases.