Jonathan Zhang

Haoming Li, Zhaoliang Chen, Jonathan Zhang et al.

Effective planning is essential for the success of any task, from organizing a vacation to routing autonomous vehicles and developing corporate strategies. It involves setting goals, formulating plans, and allocating resources to achieve them. LLMs are particularly well-suited for automated planning due to their strong capabilities in commonsense reasoning. They can deduce a sequence of actions needed to achieve a goal from a given state and identify an effective course of action. However, it is frequently observed that plans generated through direct prompting often fail upon execution. Our survey aims to highlight the existing challenges in planning with language models, focusing on key areas such as embodied environments, optimal scheduling, competitive and cooperative games, task decomposition, reasoning, and planning. Through this study, we explore how LLMs transform AI planning and provide unique insights into the future of LM-assisted planning.

Haoming Li, Zhaoliang Chen, Jonathan Zhang et al.

Planning is central to agents and agentic AI. The ability to plan, e.g., creating travel itineraries within a budget, holds immense potential in both scientific and commercial contexts. Moreover, optimal plans tend to require fewer resources compared to ad-hoc methods. To date, a comprehensive understanding of existing planning benchmarks appears to be lacking. Without it, comparing planning algorithms' performance across domains or selecting suitable algorithms for new scenarios remains challenging. In this paper, we examine a range of planning benchmarks to identify commonly used testbeds for algorithm development and highlight potential gaps. These benchmarks are categorized into embodied environments, web navigation, scheduling, games and puzzles, and everyday task automation. Our study recommends the most appropriate benchmarks for various algorithms and offers insights to guide future benchmark development.

Leo Lu, Jonathan Zhang, Sean Chua et al.

Chain-of-Thought (CoT) prompting has significantly advanced the reasoning capabilities of large language models (LLMs). While prior work focuses on improving model performance through internal reasoning strategies, little is known about the interchangeability of reasoning across different models. In this work, we explore whether a partially completed reasoning chain from one model can be reliably continued by another model, either within the same model family or across families. We achieve this by assessing the sufficiency of intermediate reasoning traces as transferable scaffolds for logical coherence and final answer accuracy. We interpret this interchangeability as a means of examining inference-time trustworthiness, probing whether reasoning remains both coherent and reliable under model substitution. Using token-level log-probability thresholds to truncate reasoning at early, mid, and late stages from our baseline models, Gemma-3-4B-IT and LLaMA-3.1-70B-Instruct, we conduct continuation experiments with Gemma-3-1B-IT and LLaMA-3.1-8B-Instruct to test intra-family and cross-family behaviors. Our evaluation pipeline leverages truncation thresholds with a Process Reward Model (PRM), providing a reproducible framework for assessing reasoning stability via model interchange. Evaluations with a PRM reveal that hybrid reasoning chains often preserve, and in some cases even improve, final accuracy and logical structure. Our findings point towards interchangeability as an emerging behavioral property of reasoning models, offering insights into new paradigms for reliable modular reasoning in collaborative AI systems.

Nayoung Choi, Jonathan Zhang, Jinho D. Choi

Large Language Models (LLMs) often exhibit increased response latency and degraded answer quality as dialogue length grows, making effective context management essential. However, existing methods rely on extra LLM calls to build memory or perform offline memory construction without considering the current user utterance, which can introduce inefficiencies or disrupt conversational continuity. We introduce DyCP, a lightweight context management method that dynamically segment and retrieve relevant memory at query time. It preserves the sequential structure of dialogue without predefined topic boundaries and supports efficient, adaptive context retrieval. Across three long-form dialogue benchmarks, LoCoMo, MT-Bench+, and SCM4LLMs, and multiple LLMs, DyCP consistently improves answer quality while reducing response latency. We also examine the gap between modern LLMs' expanded context windows and their actual long-context processing capacity, highlighting the continued importance of effective context management.

Jonathan Zhang, Bowen Xie, Xin Wu et al.

In this work, deep learning algorithms are used to classify fundus images in terms of diabetic retinopathy severity. Six different combinations of two model architectures, the Dense Convolutional Network-121 and the Residual Neural Network-50 and three image types, RGB, Green, and High Contrast, were tested to find the highest performing combination. We achieved an average validation loss of 0.17 and a max validation accuracy of 85 percent. By testing out multiple combinations, certain combinations of parameters performed better than others, though minimal variance was found overall. Green filtration was shown to perform the poorest, while amplified contrast appeared to have a negligible effect in comparison to RGB analysis. ResNet50 proved to be less of a robust model as opposed to DenseNet121.

David Borland, Jonathan Zhang, Smiti Kaul et al.

The collection and visual analysis of large-scale data from complex systems, such as electronic health records or clickstream data, has become increasingly common across a wide range of industries. This type of retrospective visual analysis, however, is prone to a variety of selection bias effects, especially for high-dimensional data where only a subset of dimensions is visualized at any given time. The risk of selection bias is even higher when analysts dynamically apply filters or perform grouping operations during ad hoc analyses. These bias effects threatens the validity and generalizability of insights discovered during visual analysis as the basis for decision making. Past work has focused on bias transparency, helping users understand when selection bias may have occurred. However, countering the effects of selection bias via bias mitigation is typically left for the user to accomplish as a separate process. Dynamic reweighting (DR) is a novel computational approach to selection bias mitigation that helps users craft bias-corrected visualizations. This paper describes the DR workflow, introduces key DR visualization designs, and presents statistical methods that support the DR process. Use cases from the medical domain, as well as findings from domain expert user interviews, are also reported.

David Borland, Wenyuan Wang, Jonathan Zhang et al.

The collection of large, complex datasets has become common across a wide variety of domains. Visual analytics tools increasingly play a key role in exploring and answering complex questions about these large datasets. However, many visualizations are not designed to concurrently visualize the large number of dimensions present in complex datasets (e.g. tens of thousands of distinct codes in an electronic health record system). This fact, combined with the ability of many visual analytics systems to enable rapid, ad-hoc specification of groups, or cohorts, of individuals based on a small subset of visualized dimensions, leads to the possibility of introducing selection bias--when the user creates a cohort based on a specified set of dimensions, differences across many other unseen dimensions may also be introduced. These unintended side effects may result in the cohort no longer being representative of the larger population intended to be studied, which can negatively affect the validity of subsequent analyses. We present techniques for selection bias tracking and visualization that can be incorporated into high-dimensional exploratory visual analytics systems, with a focus on medical data with existing data hierarchies. These techniques include: (1) tree-based cohort provenance and visualization, with a user-specified baseline cohort that all other cohorts are compared against, and visual encoding of the drift for each cohort, which indicates where selection bias may have occurred, and (2) a set of visualizations, including a novel icicle-plot based visualization, to compare in detail the per-dimension differences between the baseline and a user-specified focus cohort. These techniques are integrated into a medical temporal event sequence visual analytics tool. We present example use cases and report findings from domain expert user interviews.

David Gotz, Jonathan Zhang, Wenyuan Wang et al.

Temporal event data are collected across a broad range of domains, and a variety of visual analytics techniques have been developed to empower analysts working with this form of data. These techniques generally display aggregate statistics computed over sets of event sequences that share common patterns. Such techniques are often hindered, however, by the high-dimensionality of many real-world event sequence datasets because the large number of distinct event types within such data prevents effective aggregation. A common coping strategy for this challenge is to group event types together as a pre-process, prior to visualization, so that each group can be represented within an analysis as a single event type. However, computing these event groupings as a pre-process also places significant constraints on the analysis. This paper presents a dynamic hierarchical aggregation technique that leverages a predefined hierarchy of dimensions to computationally quantify the informativeness of alternative levels of grouping within the hierarchy at runtime. This allows users to dynamically explore the hierarchy to select the most appropriate level of grouping to use at any individual step within an analysis. Key contributions include an algorithm for interactively determining the most informative set of event groupings from within a large-scale hierarchy of event types, and a scatter-plus-focus visualization that supports interactive hierarchical exploration. While these contributions are generalizable to other types of problems, we apply them to high-dimensional event sequence analysis using large-scale event type hierarchies from the medical domain. We describe their use within a medical cohort analysis tool called Cadence, demonstrate an example in which the proposed technique supports better views of event sequence data, and report findings from domain expert interviews.