Mengting Wan

Taiwei Shi, Zhuoer Wang, Longqi Yang et al. · amazon-science, microsoft-research

As large language models (LLMs) continue to advance, aligning these models with human preferences has emerged as a critical challenge. Traditional alignment methods, relying on human or LLM annotated datasets, are limited by their resource-intensive nature, inherent subjectivity, misalignment with real-world user preferences, and the risk of feedback loops that amplify model biases. To overcome these limitations, we introduce WildFeedback, a novel framework that leverages in-situ user feedback during conversations with LLMs to create preference datasets automatically. Given a corpus of multi-turn user-LLM conversation, WildFeedback identifies and classifies user feedback to LLM responses between conversation turns. The user feedback is then used to create examples of preferred and dispreferred responses according to users' preference. Our experiments demonstrate that LLMs fine-tuned on WildFeedback dataset exhibit significantly improved alignment with user preferences, as evidenced by both traditional benchmarks and our proposed checklist-guided evaluation. By incorporating in-situ feedback from actual users, WildFeedback addresses the scalability, subjectivity, and bias challenges that plague existing approaches, marking a significant step toward developing LLMs that are more responsive to the diverse and evolving needs of their users.

Chirag Shah, Ryen W. White, Reid Andersen et al.

Log data can reveal valuable information about how users interact with Web search services, what they want, and how satisfied they are. However, analyzing user intents in log data is not easy, especially for emerging forms of Web search such as AI-driven chat. To understand user intents from log data, we need a way to label them with meaningful categories that capture their diversity and dynamics. Existing methods rely on manual or machine-learned labeling, which are either expensive or inflexible for large and dynamic datasets. We propose a novel solution using large language models (LLMs), which can generate rich and relevant concepts, descriptions, and examples for user intents. However, using LLMs to generate a user intent taxonomy and apply it for log analysis can be problematic for two main reasons: (1) such a taxonomy is not externally validated; and (2) there may be an undesirable feedback loop. To address this, we propose a new methodology with human experts and assessors to verify the quality of the LLM-generated taxonomy. We also present an end-to-end pipeline that uses an LLM with human-in-the-loop to produce, refine, and apply labels for user intent analysis in log data. We demonstrate its effectiveness by uncovering new insights into user intents from search and chat logs from the Microsoft Bing commercial search engine. The proposed work's novelty stems from the method for generating purpose-driven user intent taxonomies with strong validation. This method not only helps remove methodological and practical bottlenecks from intent-focused research, but also provides a new framework for generating, validating, and applying other kinds of taxonomies in a scalable and adaptable way with reasonable human effort.

Jing Ma, Mengting Wan, Longqi Yang et al.

Hypergraphs provide an effective abstraction for modeling multi-way group interactions among nodes, where each hyperedge can connect any number of nodes. Different from most existing studies which leverage statistical dependencies, we study hypergraphs from the perspective of causality. Specifically, in this paper, we focus on the problem of individual treatment effect (ITE) estimation on hypergraphs, aiming to estimate how much an intervention (e.g., wearing face covering) would causally affect an outcome (e.g., COVID-19 infection) of each individual node. Existing works on ITE estimation either assume that the outcome on one individual should not be influenced by the treatment assignments on other individuals (i.e., no interference), or assume the interference only exists between pairs of connected individuals in an ordinary graph. We argue that these assumptions can be unrealistic on real-world hypergraphs, where higher-order interference can affect the ultimate ITE estimations due to the presence of group interactions. In this work, we investigate high-order interference modeling, and propose a new causality learning framework powered by hypergraph neural networks. Extensive experiments on real-world hypergraphs verify the superiority of our framework over existing baselines.

Sarkar Snigdha Sarathi Das, Chirag Shah, Mengting Wan et al.

The traditional Dialogue State Tracking (DST) problem aims to track user preferences and intents in user-agent conversations. While sufficient for task-oriented dialogue systems supporting narrow domain applications, the advent of Large Language Model (LLM)-based chat systems has introduced many real-world intricacies in open-domain dialogues. These intricacies manifest in the form of increased complexity in contextual interactions, extended dialogue sessions encompassing a diverse array of topics, and more frequent contextual shifts. To handle these intricacies arising from evolving LLM-based chat systems, we propose joint dialogue segmentation and state tracking per segment in open-domain dialogue systems. Assuming a zero-shot setting appropriate to a true open-domain dialogue system, we propose S3-DST, a structured prompting technique that harnesses Pre-Analytical Recollection, a novel grounding mechanism we designed for improving long context tracking. To demonstrate the efficacy of our proposed approach in joint segmentation and state tracking, we evaluate S3-DST on a proprietary anonymized open-domain dialogue dataset, as well as publicly available DST and segmentation datasets. Across all datasets and settings, S3-DST consistently outperforms the state-of-the-art, demonstrating its potency and robustness the next generation of LLM-based chat systems.

Sheshera Mysore, Zhuoran Lu, Mengting Wan et al.

Powerful large language models have facilitated the development of writing assistants that promise to significantly improve the quality and efficiency of composition and communication. However, a barrier to effective assistance is the lack of personalization in LLM outputs to the author's communication style, specialized knowledge, and values. In this paper, we address this challenge by proposing Pearl, a LLM writing assistant personalized with a retriever that is trained to be generation-calibrated for personalization. Generation calibration ensures that our retriever selects historic user authored documents to augment an LLM prompt such that they are likely to help an LLM generation better adhere to a users' preferences. We propose two key novelties for training such a retriever: (1) A training data selection method that identifies user requests likely to benefit from personalization and documents that provide that benefit; and (2) A scale-calibrating KL-divergence objective that ensures that our retriever scores remain proportional to the downstream generation quality from using the document for personalized generation. In a series of holistic evaluations, we demonstrate the effectiveness of Pearl in generating long-form texts on multiple social media datasets. Finally, we demonstrate how a generation-calibrated retriever can double as a performance predictor -- detecting low quality retrieval, and improving potentially under-performing outputs via revision with LLMs.

Wenyue Hua, Mengting Wan, Shashank Vadrevu et al.

Agents, as user-centric tools, are increasingly deployed for human task delegation, assisting with a broad spectrum of requests by generating thoughts, engaging with user proxies, and producing action plans. However, agents based on large language models (LLMs) often face substantial planning latency due to two primary factors: the efficiency limitations of the underlying LLMs due to their large size and high demand, and the structural complexity of the agents due to the extensive generation of intermediate thoughts to produce the final output. Given that inefficiency in service provision can undermine the value of automation for users, this paper presents a human-centered efficient agent planning method -- Interactive Speculative Planning -- aiming at enhancing the efficiency of agent planning through both system design and human-AI interaction. Our approach advocates for the co-design of the agent system and user interface, underscoring the importance of an agent system that can fluidly manage user interactions and interruptions. By integrating human interruptions as a fundamental component of the system, we not only make it more user-centric but also expedite the entire process by leveraging human-in-the-loop interactions to provide accurate intermediate steps. Code and data will be released.

Tobias Schnabel, Mengting Wan, Longqi Yang

With information systems becoming larger scale, recommendation systems are a topic of growing interest in machine learning research and industry. Even though progress on improving model design has been rapid in research, we argue that many advances fail to translate into practice because of two limiting assumptions. First, most approaches focus on a transductive learning setting which cannot handle unseen users or items and second, many existing methods are developed for static settings that cannot incorporate new data as it becomes available. We argue that these are largely impractical assumptions on real-world platforms where new user interactions happen in real time. In this survey paper, we formalize both concepts and contextualize recommender systems work from the last six years. We then discuss why and how future work should move towards inductive learning and incremental updates for recommendation model design and evaluation. In addition, we present best practices and fundamental open challenges for future research.

Weiwei Sun, Xuhui Zhou, Jiarui Liu et al.

Humans learn social norms and behaviors from verbal feedback (e.g., a parent saying "that was rude" or a friend explaining "here's why that hurt"). Yet, learning from feedback for LLMs has largely focused on domains like code and math, where RL rewards are directly verifiable and condensed into scalar values. As LLMs are increasingly used to simulate human behavior, e.g., standing in for users, patients, students, and other personas, there is a pressing need to make them more human-like, which requires embracing a fundamentally different kind of signal: feedback that is verbal, subjective, and multi-faceted. We present DITTO, a model trained by treating verbal feedback as a first-class signal in reinforcement learning. After each rollout, DITTO receives verbal feedback and generates a feedback-conditioned improved rollout; both outputs are jointly optimized with GRPO, distilling verbal guidance into the base policy without requiring feedback at test time. We also introduce SOUL (Simulation gym Of hUman-Like behavior), a unified benchmark and training data suite spanning 10 tasks across six categories: Theory of Mind, character role play, social skill, learner simulation, user simulation, and persona simulation. DITTO achieves an average 36% improvement over the base model and exceeds GPT-5.4 on 6 of 10 SOUL benchmarks, demonstrating that RL with verbal feedback is a promising direction for training LLMs to simulate human behavior.

Ryo Kamoi, Ameya Godbole, Longqi Yang et al.

Simulating human conversations using large language models (LLMs) has emerged as a scalable methodology for modeling human social interaction. However, simulating human conversations is challenging because they inherently involve inconsistent and uncollaborative behaviors, such as misunderstandings and interruptions. Analysis comparing inconsistent and uncollaborative behaviors in human- and LLM-generated conversations remains limited, although reproducing these behaviors is integral to simulating human-like and complex social interaction. In this work, we introduce CoCoEval, an evaluation framework that analyzes LLM-simulated conversations by detecting 10 types of inconsistent and uncollaborative behaviors at the turn level using an LLM-as-a-Judge. Using CoCoEval, we evaluate GPT-4.1, GPT-5.1, and Claude Opus 4 by comparing the frequencies of detected behaviors in conversations simulated by each model and in human conversations across academic, business, and governmental meetings, as well as debates. Our analysis shows that (1) under vanilla prompting, LLM-simulated conversations exhibit far fewer inconsistent and uncollaborative behaviors than human conversations; (2) prompt engineering does not provide reliable control over these behaviors, as our results show that different prompts lead to their under- or overproduction; and (3) supervised fine-tuning on human conversations can lead LLMs to overproduce a narrow set of behaviors, such as repetition. Our findings highlight the difficulty of simulating human conversations, raising concerns about the use of LLMs as a proxy for human social interaction.

Zhiwei Liu, Mengting Wan, Stephen Guo et al.

Within-basket recommendation reduces the exploration time of users, where the user's intention of the basket matters. The intent of a shopping basket can be retrieved from both user-item collaborative filtering signals and multi-item correlations. By defining a basket entity to represent the basket intent, we can model this problem as a basket-item link prediction task in the User-Basket-Item~(UBI) graph. Previous work solves the problem by leveraging user-item interactions and item-item interactions simultaneously. However, collectivity and heterogeneity characteristics are hardly investigated before. Collectivity defines the semantics of each node which should be aggregated from both directly and indirectly connected neighbors. Heterogeneity comes from multi-type interactions as well as multi-type nodes in the UBI graph. To this end, we propose a new framework named \textbf{BasConv}, which is based on the graph convolutional neural network. Our BasConv model has three types of aggregators specifically designed for three types of nodes. They collectively learn node embeddings from both neighborhood and high-order context. Additionally, the interactive layers in the aggregators can distinguish different types of interactions. Extensive experiments on two real-world datasets prove the effectiveness of BasConv. Our code is available online at https://github.com/JimLiu96/basConv.

Mengting Wan, Tara Safavi, Sujay Kumar Jauhar et al.

Transforming unstructured text into structured and meaningful forms, organized by useful category labels, is a fundamental step in text mining for downstream analysis and application. However, most existing methods for producing label taxonomies and building text-based label classifiers still rely heavily on domain expertise and manual curation, making the process expensive and time-consuming. This is particularly challenging when the label space is under-specified and large-scale data annotations are unavailable. In this paper, we address these challenges with Large Language Models (LLMs), whose prompt-based interface facilitates the induction and use of large-scale pseudo labels. We propose TnT-LLM, a two-phase framework that employs LLMs to automate the process of end-to-end label generation and assignment with minimal human effort for any given use-case. In the first phase, we introduce a zero-shot, multi-stage reasoning approach which enables LLMs to produce and refine a label taxonomy iteratively. In the second phase, LLMs are used as data labelers that yield training samples so that lightweight supervised classifiers can be reliably built, deployed, and served at scale. We apply TnT-LLM to the analysis of user intent and conversational domain for Bing Copilot (formerly Bing Chat), an open-domain chat-based search engine. Extensive experiments using both human and automatic evaluation metrics demonstrate that TnT-LLM generates more accurate and relevant label taxonomies when compared against state-of-the-art baselines, and achieves a favorable balance between accuracy and efficiency for classification at scale. We also share our practical experiences and insights on the challenges and opportunities of using LLMs for large-scale text mining in real-world applications.

Jie He, Jennifer Neville, Mengting Wan et al.

Large Language Models (LLMs) can enhance their capabilities as AI assistants by integrating external tools, allowing them to access a wider range of information. While recent LLMs are typically fine-tuned with tool usage examples during supervised fine-tuning (SFT), questions remain about their ability to develop robust tool-usage skills and can effectively generalize to unseen queries and tools. In this work, we present GenTool, a novel training framework that prepares LLMs for diverse generalization challenges in tool utilization. Our approach addresses two fundamental dimensions critical for real-world applications: Zero-to-One Generalization, enabling the model to address queries initially lacking a suitable tool by adopting and utilizing one when it becomes available, and Weak-to-Strong Generalization, allowing models to leverage enhanced versions of existing tools to solve queries. To achieve this, we develop synthetic training data simulating these two dimensions of tool usage and introduce a two-stage fine-tuning approach: optimizing tool ranking, then refining tool selection. Through extensive experiments across four generalization scenarios, we demonstrate that our method significantly enhances the tool-usage capabilities of LLMs ranging from 1B to 8B parameters, achieving performance that surpasses GPT-4o. Furthermore, our analysis also provides valuable insights into the challenges LLMs encounter in tool generalization.

Ishani Mondal, Jack W. Stokes, Sujay Kumar Jauhar et al.

LLMs often fail to meet the specialized needs of distinct user groups due to their one-size-fits-all training paradigm \cite{lucy-etal-2024-one} and there is limited research on what personalization aspects each group expect. To address these limitations, we propose a group-aware personalization framework, Group Preference Alignment (GPA), that identifies context-specific variations in conversational preferences across user groups and then steers LLMs to address those preferences. Our approach consists of two steps: (1) Group-Aware Preference Extraction, where maximally divergent user-group preferences are extracted from real-world conversation logs and distilled into interpretable rubrics, and (2) Tailored Response Generation, which leverages these rubrics through two methods: a) Context-Tuned Inference (GAP-CT), that dynamically adjusts responses via context-dependent prompt instructions, and b) Rubric-Finetuning Inference (GPA-FT), which uses the rubrics to generate contrastive synthetic data for personalization of group-specific models via alignment. Experiments demonstrate that our framework significantly improves alignment of the output with respect to user preferences and outperforms baseline methods, while maintaining robust performance on standard benchmarks.

Tenghao Huang, Sihao Chen, Muhao Chen et al.

Large language models (LLMs) are increasingly expected to function as collaborative partners, engaging in back-and-forth dialogue to solve complex, ambiguous problems. However, current LLMs often falter in real-world settings, defaulting to passive responses or narrow clarifications when faced with incomplete or under-specified prompts, falling short of proactively gathering the missing information that is crucial for high-quality solutions. In this work, we introduce a new task paradigm: proactive information gathering, where LLMs must identify gaps in the provided context and strategically elicit implicit user knowledge through targeted questions. To systematically study and train this capability, we design a scalable framework that generates partially specified, real-world tasks, masking key information and simulating authentic ambiguity. Within this setup, our core innovation is a reinforcement finetuning strategy that rewards questions that elicit genuinely new, implicit user information -- such as hidden domain expertise or fine-grained requirements -- that would otherwise remain unspoken. Experiments demonstrate that our trained Qwen-2.5-7B model significantly outperforms o3-mini by 18% on automatic evaluation metrics. More importantly, human evaluation reveals that clarification questions and final outlines generated by our model are favored by human annotators by 42% and 28% respectively. Together, these results highlight the value of proactive clarification in elevating LLMs from passive text generators to genuinely collaborative thought partners.

Siddharth Suri, Scott Counts, Leijie Wang et al.

Until recently, search engines were the predominant method for people to access online information. The recent emergence of large language models (LLMs) has given machines new capabilities such as the ability to generate new digital artifacts like text, images, code etc., resulting in a new tool, a generative search engine, which combines the capabilities of LLMs with a traditional search engine. Through the empirical analysis of Bing Copilot (Bing Chat), one of the first publicly available generative search engines, we analyze the types and complexity of tasks that people use Bing Copilot for compared to Bing Search. Findings indicate that people use the generative search engine for more knowledge work tasks that are higher in cognitive complexity than were commonly done with a traditional search engine.

Ying-Chun Lin, Jennifer Neville, Jack W. Stokes et al.

Accurate and interpretable user satisfaction estimation (USE) is critical for understanding, evaluating, and continuously improving conversational systems. Users express their satisfaction or dissatisfaction with diverse conversational patterns in both general-purpose (ChatGPT and Bing Copilot) and task-oriented (customer service chatbot) conversational systems. Existing approaches based on featurized ML models or text embeddings fall short in extracting generalizable patterns and are hard to interpret. In this work, we show that LLMs can extract interpretable signals of user satisfaction from their natural language utterances more effectively than embedding-based approaches. Moreover, an LLM can be tailored for USE via an iterative prompting framework using supervision from labeled examples. The resulting method, Supervised Prompting for User satisfaction Rubrics (SPUR), not only has higher accuracy but is more interpretable as it scores user satisfaction via learned rubrics with a detailed breakdown.

Jing Ma, Ruocheng Guo, Mengting Wan et al.

Fair machine learning aims to mitigate the biases of model predictions against certain subpopulations regarding sensitive attributes such as race and gender. Among the many existing fairness notions, counterfactual fairness measures the model fairness from a causal perspective by comparing the predictions of each individual from the original data and the counterfactuals. In counterfactuals, the sensitive attribute values of this individual had been modified. Recently, a few works extend counterfactual fairness to graph data, but most of them neglect the following facts that can lead to biases: 1) the sensitive attributes of each node's neighbors may causally affect the prediction w.r.t. this node; 2) the sensitive attributes may causally affect other features and the graph structure. To tackle these issues, in this paper, we propose a novel fairness notion - graph counterfactual fairness, which considers the biases led by the above facts. To learn node representations towards graph counterfactual fairness, we propose a novel framework based on counterfactual data augmentation. In this framework, we generate counterfactuals corresponding to perturbations on each node's and their neighbors' sensitive attributes. Then we enforce fairness by minimizing the discrepancy between the representations learned from the original graph and the counterfactuals for each node. Experiments on both synthetic and real-world graphs show that our framework outperforms the state-of-the-art baselines in graph counterfactual fairness, and also achieves comparable prediction performance.

Mengting Wan, Jianmo Ni, Rishabh Misra et al.

Modern collaborative filtering algorithms seek to provide personalized product recommendations by uncovering patterns in consumer-product interactions. However, these interactions can be biased by how the product is marketed, for example due to the selection of a particular human model in a product image. These correlations may result in the underrepresentation of particular niche markets in the interaction data; for example, a female user who would potentially like motorcycle products may be less likely to interact with them if they are promoted using stereotypically 'male' images. In this paper, we first investigate this correlation between users' interaction feedback and products' marketing images on two real-world e-commerce datasets. We further examine the response of several standard collaborative filtering algorithms to the distribution of consumer-product market segments in the input interaction data, revealing that marketing strategy can be a source of bias for modern recommender systems. In order to protect recommendation performance on underrepresented market segments, we develop a framework to address this potential marketing bias. Quantitative results demonstrate that the proposed approach significantly improves the recommendation fairness across different market segments, with a negligible loss (or better) recommendation accuracy.

An Yan, Shuo Cheng, Wang-Cheng Kang et al.

Sequential patterns play an important role in building modern recommender systems. To this end, several recommender systems have been built on top of Markov Chains and Recurrent Models (among others). Although these sequential models have proven successful at a range of tasks, they still struggle to uncover complex relationships nested in user purchase histories. In this paper, we argue that modeling pairwise relationships directly leads to an efficient representation of sequential features and captures complex item correlations. Specifically, we propose a 2D convolutional network for sequential recommendation (CosRec). It encodes a sequence of items into a three-way tensor; learns local features using 2D convolutional filters; and aggregates high-order interactions in a feedforward manner. Quantitative results on two public datasets show that our method outperforms both conventional methods and recent sequence-based approaches, achieving state-of-the-art performance on various evaluation metrics.

Mengting Wan, Rishabh Misra, Ndapa Nakashole et al.

This paper presents computational approaches for automatically detecting critical plot twists in reviews of media products. First, we created a large-scale book review dataset that includes fine-grained spoiler annotations at the sentence-level, as well as book and (anonymized) user information. Second, we carefully analyzed this dataset, and found that: spoiler language tends to be book-specific; spoiler distributions vary greatly across books and review authors; and spoiler sentences tend to jointly appear in the latter part of reviews. Third, inspired by these findings, we developed an end-to-end neural network architecture to detect spoiler sentences in review corpora. Quantitative and qualitative results demonstrate that the proposed method substantially outperforms existing baselines.

Mengting Wan, Xin Chen

We study the problem of providing recommended responses to customer service agents in live-chat dialogue systems. Smart-reply systems have been widely applied in real-world applications (e.g. Gmail, LinkedIn Messaging), where most of them can successfully recommend reactive responses. However, we observe a major limitation of current methods is that they generally have difficulties in suggesting proactive investigation act (e.g. "Do you perhaps have another account with us?") due to the lack of long-term context information, which indeed act as critical steps for customer service agents to collect information and resolve customers' issues. Thus in this work, we propose an end-to-end method with special focus on suggesting proactive investigative questions to customer agents in Airbnb's customer service live-chat system. Effectiveness of our proposed method can be validated through qualitative and quantitative results.

Wang-Cheng Kang, Mengting Wan, Julian McAuley

Recommender Systems have proliferated as general-purpose approaches to model a wide variety of consumer interaction data. Specific instances make use of signals ranging from user feedback, item relationships, geographic locality, social influence (etc.). Typically, research proceeds by showing that making use of a specific signal (within a carefully designed model) allows for higher-fidelity recommendations on a particular dataset. Of course, the real situation is more nuanced, in which a combination of many signals may be at play, or favored in different proportion by individual users. Here we seek to develop a framework that is capable of combining such heterogeneous item relationships by simultaneously modeling (a) what modality of recommendation is a user likely to be susceptible to at a particular point in time; and (b) what is the best recommendation from each modality. Our method borrows ideas from mixtures-of-experts approaches as well as knowledge graph embeddings. We find that our approach naturally yields more accurate recommendations than alternatives, while also providing intuitive `explanations' behind the recommendations it provides.

Mengting Wan, Julian McAuley

Product review websites provide an incredible lens into the wide variety of opinions and experiences of different people, and play a critical role in helping users discover products that match their personal needs and preferences. To help address questions that can't easily be answered by reading others' reviews, some review websites also allow users to pose questions to the community via a question-answering (QA) system. As one would expect, just as opinions diverge among different reviewers, answers to such questions may also be subjective, opinionated, and divergent. This means that answering such questions automatically is quite different from traditional QA tasks, where it is assumed that a single `correct' answer is available. While recent work introduced the idea of question-answering using product reviews, it did not account for two aspects that we consider in this paper: (1) Questions have multiple, often divergent, answers, and this full spectrum of answers should somehow be used to train the system; and (2) What makes a `good' answer depends on the asker and the answerer, and these factors should be incorporated in order for the system to be more personalized. Here we build a new QA dataset with 800 thousand questions---and over 3.1 million answers---and show that explicitly accounting for personalization and ambiguity leads both to quantitatively better answers, but also a more nuanced view of the range of supporting, but subjective, opinions.