Harshita Chopra

Long-horizon personalization requires dialog assistants to retrieve user-specific facts from extended interaction histories. However, many relevant facts have low semantic similarity to the query under dense retrieval. Standard Retrieval-Augmented Generation (RAG) and GraphRAG systems remain largely retrospective: they rely on embedding similarity or fixed graph traversals, often missing relevant facts that lie far from the query in embedding space. Inspired by prospection, the human ability to use imagined futures as cues for recall, we introduce Prospection-Guided Retrieval (PGR), which decouples retrieval from memory storage. Given a user query, PGR expands the goal into a linear chain or tree of thoughts (TOT) denoting plausible next steps, and uses these steps as retrieval probes instead of the original query alone. Retrieved facts are then used to personalize the next round of prospection, enabling PGR to uncover additional memories by grounding the simulation in the user’s history. We also introduce MemoryQuest, a challenging multi-session benchmark where each query is annotated with 3–5 dated reference facts under a low query-reference similarity constraint. Across 1,625 queries spanning 185 user profiles from 3 public datasets, PGR-TOT substantially improves retrieval, achieving nearly 3x recall on MemoryQuest over baselines. In pairwise LLM-as-judge evaluations, PGR-generated responses are preferred on 89–98% of queries, with blinded human annotations showing a similar trend. Overall, the results show that explicit prospection yields large gains in long-horizon retrieval and response quality over similarity-only baselines.

The ability to identify and acquire missing information is a critical component of effective decision making and problem solving. With the rise of conversational artificial intelligence (AI) systems, strategically formulating information-seeking questions becomes crucial and demands efficient methods to guide the search process. We introduce a novel approach to adaptive question-asking through a combination of Large Language Models (LLM) for generating questions that maximize information gain, Monte Carlo Tree Search (MCTS) for building and using a decision tree across multiple samples, and a hierarchical feedback mechanism to learn from past interactions. We present two key innovations: (1) an adaptive MCTS algorithm that balances exploration and exploitation for efficient search over potential questions; and (2) a clustering-based feedback algorithm that leverages prior experience to guide future interactions. Each incoming sample is assigned to a cluster based on its semantic similarity with previously observed samples. Our UCT (Upper Confidence bound for Trees) formulation selects optimal questions by combining expected rewards, a function of information gain, with a clusterspecific bonus that decays with depth, to emphasize the importance of early-stage questions that have proven effective for narrowing the solution space in similar samples. Experiments in medical diagnosis and troubleshooting domains demonstrate that our method leads to an average of 12% improvement in success rates and a 10x reduction in the average number of LLM calls made per conversation for the search process, in comparison to the state of the art.

A Large Language Model (LLM) offers versatility across domains and tasks, purportedly benefiting users with a wide variety of behaviors and preferences. We question this perception about an LLM when users have inherently subjective behaviors and preferences, as seen in their ubiquitous and idiosyncratic browsing of websites or apps. The sequential behavior logs of pages, thus generated, form something akin to each user's self-constructed "language", albeit without the structure and grammar imbued in natural languages. We ask: (i) Can a small LM represent the "language of browsing" better than a large LM? (ii) Can an LM with a single set of parameters (or, single LM) adequately capture myriad users' heterogeneous, subjective behaviors and preferences? (iii) Can a single LM with high average performance, yield low variance in performance to make alignment good at user level? We introduce clusterwise LM training, HeTLM (Heterogeneity aware Training of Language Model), appropriate for subjective behaviors. We find that (i) a small LM trained using a page-level tokenizer outperforms large pretrained or finetuned LMs; (ii) HeTLM with heterogeneous cluster specific set of parameters outperforms a single LM of the same family, controlling for the number of parameters; and (iii) a higher mean and a lower variance in generation ensues, implying improved alignment.

Users' behavioral footprints online enable firms to discover behavior-based user segments (or, segments) and deliver segment specific messages to users. Following the discovery of segments, delivery of messages to users through preferred media channels like Facebook and Google can be challenging, as only a portion of users in a behavior segment find match in a medium, and only a fraction of those matched actually see the message (exposure). Even high quality discovery becomes futile when delivery fails. Many sophisticated algorithms exist for discovering behavioral segments; however, these ignore the delivery component. The problem is compounded because (i) the discovery is performed on the behavior data space in firms' data (e.g., user clicks), while the delivery is predicated on the static data space (e.g., geo, age) as defined by media; and (ii) firms work under budget constraint. We introduce a stochastic optimization based algorithm for delivery optimized discovery of behavioral user segments and offer new metrics to address the joint optimization. We leverage optimization under a budget constraint for delivery combined with a learning-based component for discovery. Extensive experiments on a public dataset from Google and a proprietary dataset show the effectiveness of our approach by simultaneously improving delivery metrics, reducing budget spend and achieving strong predictive performance in discovery.

Online browsing on firms' sites generates user behavior logs (or, logs). These logs are mainstays that drive several user modeling tasks. The logs that inform user modeling are the ones that are attributed to each user, termed Attributed Behaviors (AB). But, a lot more logs are anonymous, upwards of 85%. For example, many users do not sign in while browsing. These logs are not attributed to users, termed Unattributed Behaviors (UB), and are not recognized in user modeling. We examine whether and how UB can benefit user modeling. We focus on a common task, that of user segmentation, for which the prior art uses only AB. We demonstrate that information from UBs, although unattributed to any individual, when used along with ABs, enriches performance of machine learning model for user segmentation. We perform predictive segmentation, whereby predicted outcomes for each segment are evaluated against actual outcomes. Multiple evaluations on two datasets, one of which is public, relative to state of the art baseline, show strong performance of our model in predicting outcomes and in reducing user segmentation error.

Exploring students’ discourse in academic settings over time can provide valuable insight into the evolution of learner engagement and participation in online learning. In this study, we propose an analytical framework to capture topics and the temporal progression of learner discourse. We employed a Contextualized Topic Modeling technique on messages posted by undergraduates in online discussion forums from Fall 2019 to Spring 2020. We further evaluated if topics were originating from specific courses or more generally distributed across multiple courses. Our results suggested a significant increase in the number of general topics after the onset of the pandemic, suggesting emergent topics being discussed in a range of courses. In addition, using Word Mover’s Distance, we examined the semantic similarity of topics in adjacent months and constructed topic chains. Our findings indicated that previously course-centric topics such as public health developed into more general discussions that emphasize inequities and healthcare during the pandemic. Furthermore, emergent topics around students’ lived experiences underscored the role of discussion forums in capturing educational experiences temporally. Finally, we discuss the implications of current findings for post-pandemic higher education and the effectiveness of our framework in exploring unstructured large-scale educational data.

Students’ conversations in academic settings evolve over time and can be affected by events such as the COVID-19 pandemic. In this paper, we employ a Contextualized Topic Modeling technique to detect coherent topics from students’ posts in online discussion forums. We construct topic chains by connecting semantically similar topics across months using Word Mover’s Distance. Consistent academic discourse and contemporary events such as the COVID-19 outbreak and the Black Lives Matter movement were found among prominent topics. In later months, new themes around students’ lived experiences emerged and evolved into discussions reflecting the shift in educational experiences. Results revealed a significant increase in more general topics after the onset of pandemic. Our proposed framework can also be applied to other contexts investigating temporal topic trends in large-scale text data.

Evidence from peer-reviewed literature is the cornerstone for designing responses to global threats such as COVID-19. In massive and rapidly growing corpuses, such as COVID-19 publications, assimilating and synthesizing information is challenging. Leveraging a robust computational pipeline that evaluates multiple aspects, such as network topological features, communities, and their temporal trends, can make this process more efficient.
Objective:
We aimed to show that new knowledge can be captured and tracked using the temporal change in the underlying unsupervised word embeddings of the literature. Further imminent themes can be predicted using machine learning on the evolving associations between words.
Methods:
Frequently occurring medical entities were extracted from the abstracts of more than 150,000 COVID-19 articles published on the World Health Organization database, collected on a monthly interval starting from February 2020. Word embeddings trained on each month’s literature were used to construct networks of entities with cosine similarities as edge weights. Topological features of the subsequent month’s network were forecasted based on prior patterns, and new links were predicted using supervised machine learning. Community detection and alluvial diagrams were used to track biomedical themes that evolved over the months.
Results:
We found that thromboembolic complications were detected as an emerging theme as early as August 2020. A shift toward the symptoms of long COVID complications was observed during March 2021, and neurological complications gained significance in June 2021. A prospective validation of the link prediction models achieved an area under the receiver operating characteristic curve of 0.87. Predictive modeling revealed predisposing conditions, symptoms, cross-infection, and neurological complications as dominant research themes in COVID-19 publications based on the patterns observed in previous months.
Conclusions:
Machine learning–based prediction of emerging links can contribute toward steering research by capturing themes represented by groups of medical entities, based on patterns of semantic relationships over time.

Social media plays a pivotal role in disseminating news globally and acts as a platform for people to express their opinions on various topics. A wide variety of views accompany COVID-19 vaccination drives across the globe, often colored by emotions that change along with rising cases, approval of vaccines, and multiple factors discussed online.
Objective:
This study aims to analyze the temporal evolution of different emotions and the related influencing factors in tweets belonging to 5 countries with vital vaccine rollout programs, namely India, the United States, Brazil, the United Kingdom, and Australia.
Methods:
We extracted a corpus of nearly 1.8 million Twitter posts related to COVID-19 vaccination and created 2 classes of lexical categories—emotions and influencing factors. Using cosine distance from selected seed words’ embeddings, we expanded the vocabulary of each category and tracked the longitudinal change in their strength from June 2020 to April 2021 in each country. Community detection algorithms were used to find modules in positive correlation networks.
Results:
Our findings indicated the varying relationship among emotions and influencing factors across countries. Tweets expressing hesitancy toward vaccines represented the highest mentions of health-related effects in all countries, which reduced from 41% to 39% in India. We also observed a significant change (P<.001) in the linear trends of categories like hesitation and contentment before and after approval of vaccines. After the vaccine approval, 42% of tweets coming from India and 45% of tweets from the United States represented the “vaccine_rollout” category. Negative emotions like rage and sorrow gained the highest importance in the alluvial diagram and formed a significant module with all the influencing factors in April 2021, when India observed the second wave of COVID-19 cases.
Conclusions:
By extracting and visualizing these tweets, we propose that such a framework may help guide the design of effective vaccine campaigns and be used by policy makers to model vaccine uptake and targeted interventions.