Hyungyu Shin, Jingyu Tang, Yoonjoo Lee, Nayoung Kim, Hyunseung Lim, Ji Yong Cho, Hwajung Hong, Moontae Lee, Juho Kim. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Patent examination remains an ongoing challenge in the NLP literature even after the advent of large language models (LLMs), as it requires an extensive yet nuanced human judgment on whether a submitted $\textit{claim}$ meets the statutory standards of $\textit{novelty}$ and $\textit{non-obviousness}$ against previously granted claims—$\textit{prior art}$—in expert domains. Previous NLP studies have approached this challenge as a prediction task (e.g., forecasting grant outcomes) with high-level proxies such as similarity metrics or classifiers trained on historical labels. However, this approach often overlooks the step-by-step evaluations that examiners must make with profound information, including rationales for the decisions provided in $\textit{office actions}$ documents, which also makes it harder to measure the current state of techniques in patent review processes. To fill this gap, we construct PANORAMA, a dataset of 8,143 U.S. patent examination records that preserves the full decision trails, including original applications, all cited references, $\textit{Non-Final Rejections}$, and $\textit{Notices of Allowance}$. Also, PANORAMA decomposes the trails into sequential benchmarks that emulate patent professionals' patent review processes and allow researchers to examine large language models' capabilities at each step of them. Our findings indicate that, although LLMs are relatively effective at retrieving relevant prior art and pinpointing the pertinent paragraphs, they struggle to assess the novelty and non-obviousness of patent claims. We discuss these results and argue that advancing NLP, including LLMs, in the patent domain requires a deeper understanding of real-world patent examination. Our dataset is openly available at https://huggingface.co/datasets/LG-AI-Research/PANORAMA.

Clients often partner with AI experts to develop AI applications tailored to their needs. In these partnerships, careful planning and clear communication are critical, as inaccurate or incomplete specifications can result in misaligned model characteristics, expensive reworks, and potential friction between collaborators. Unfortunately, given the complexity of requirements ranging from functionality, data, and governance, effective guidelines for collaborative specification of requirements in client-AI expert collaborations are missing. In this work, we introduce AINeedsPlanner, a workbook that AI experts and clients can use to facilitate effective interchange of clear specifications. The workbook is based on (1) an interview of 10 completed AI application project teams, which identifies and characterizes steps in AI application planning and (2) a study with 12 AI experts, which defines a taxonomy of AI experts’ information needs and dimensions that affect the information needs. Finally, we demonstrate the workbook’s utility with two case studies in real-world settings.

Research promotion enables researchers to share advanced knowledge with pertinent academic communities. The question-and-answer (QA) style articles are effective for researchers to promote their research by enabling readers to understand research on complex subjects. Recent advances in large language models (LLMs) have opened avenues for supporting researchers in creating QA-style articles for research promotion. However, without the authors’ involvement, these models may only partially capture the researcher’s intention and voice. We developed AQUA, a research probe that enables researchers to co-create QA-style articles with LLMs to promote their research papers. A user study (n=12) reveals that LLMs reduced authors’ burden and helped them understand the readers’ perspectives. Nevertheless, LLMs failed to capture the unique intent of the authors, and their automated generation discouraged authors from carefully revising their answers. Based on our findings, we discuss human-LLM interaction design to enable authors to create QA-style articles that reflect their intention.

This work investigates large language models (LLMs) as teachable agents for learning by teaching (LBT). LBT with teachable agents helps learners identify knowledge gaps and discover new knowledge. However, teachable agents require expensive programming of subject-specific knowledge. While LLMs as teachable agents can reduce the cost, LLMs’ expansive knowledge as tutees discourages learners from teaching. We propose a prompting pipeline that restrains LLMs’ knowledge and makes them initiate “why” and “how” questions for effective knowledge-building. We combined these techniques into TeachYou, an LBT environment for algorithm learning, and AlgoBo, an LLM-based tutee chatbot that can simulate misconceptions and unawareness prescribed in its knowledge state. Our technical evaluation confirmed that our prompting pipeline can effectively configure AlgoBo’s problem-solving performance. Through a between-subject study with 40 algorithm novices, we also observed that AlgoBo’s questions led to knowledge-dense conversations (effect size=0.71). Lastly, we discuss design implications, cost-efficiency, and personalization of LLM-based teachable agents.

Designing solution plans before writing code is critical for successful algorithmic problem-solving. Novices, however, often plan on-the-fly during implementation, resulting in unsuccessful problem-solving due to lack of mental organization of the solution. Research shows that subgoal learning helps learners develop more complete solution plans by enhancing their understanding of the high-level solution structure. However, expert-created materials such as subgoal labels are necessary to provide learning benefits from subgoal learning, which are a scarce resource in self-learning due to limited availability and high cost. We propose a learnersourcing workflow that collects high-quality subgoal labels from learners by helping them improve their label quality. We implemented the workflow into AlgoSolve, a prototype interface that supports subgoal learning for algorithmic problems. A between-subjects study with 63 problem-solving novices revealed that AlgoSolve helped learners create higher-quality labels and more complete solution plans, compared to a baseline method known to be effective in subgoal learning.

We introduce XDesign, a web-based interactive platform that guides learners through a multi-stage design process for creating user-centered explanations of AI models. Results from a course deployment show that students were able to identify concrete user needs in interacting with explanations, highlight user tasks to support the needs, and design a user interface that aids the tasks.

Planning a solution before writing code is essential in algorithmic problem-solving. However, novices often skip planning and jump straight into coding. Even if they set up a plan, some do not connect to their plan when writing code. Learners solving algorithmic problems often struggle with high-level components such as solution techniques and sub-problems, but existing representations that guide learners in planning, such as flowcharts, focus on presenting lower-level details. We use subgoal diagrams -- diagrams made of subgoal labels and the relationships between them -- as a representation that guides learners to focus on high-level plans when they develop solutions. We introduce AlgoPlan, an interface that enables learners to build their own subgoal diagram and use it to guide their problem-solving process. A preliminary study with seven students shows that subgoal diagrams help learners focus on high-level plans and connect these plans to their code.

Similar to a design process for designing graphical user interfaces, conversation designers often apply an iterative design process by defining a conversation flow, testing with users, reviewing user data, and improving the design. While it is possible to iterate on conversation design with existing chatbot prototyping tools, there still remain challenges in recruiting participants on-demand and collecting structured feedback on specific conversational components. These limitations hinder designers from running rapid iterations and making informed design decisions. We posit that involving a crowd in the conversation design process can address these challenges, and introduce ProtoChat, a crowd-powered chatbot design tool built to support the iterative process of conversation design. ProtoChat makes it easy to recruit crowd workers to test the current conversation within the design tool. ProtoChat's crowd-testing tool allows crowd workers to provide concrete and practical feedback and suggest improvements on specific parts of the conversation. With the data collected from crowd-testing, ProtoChat provides multiple types of visualizations to help designers analyze and revise their design. Through a three-day study with eight designers, we found that ProtoChat enabled an iterative design process for designing a chatbot. Designers improved their design by not only modifying the conversation design itself, but also adjusting the persona and getting UI design implications beyond the conversation design itself. The crowd responses were helpful for designers to explore user needs, contexts, and diverse response formats. With ProtoChat, designers can successfully collect concrete evidence from the crowd and make decisions to iteratively improve their conversation design.

Online instructional videos are ubiquitous, but it is difficult for instructors to gauge learners' experience and their level of comprehension or confusion regarding the lecture video. Moreover, learners watching the videos may become disengaged or fail to reflect and construct their own understanding. This paper explores instructor and learner perceptions of in-video prompting where learners answer reflective questions while watching videos. We conducted two studies with crowd workers to understand the effect of prompting in general, and the effect of different prompting strategies on both learners and instructors. Results show that some learners found prompts to be useful checkpoints for reflection, while others found them distracting. Instructors reported the collected responses to be generally more specific than what they have usually collected. Also, different prompting strategies had different effects on the learning experience and the usefulness of responses as feedback.

A dynamic graph is defined by an initial graph and a graph update stream consisting of edge insertions and deletions. Identifying and monitoring critical patterns in the dynamic graph is important in various application domains such as fraud detection, cyber security, and emergency response. Given a dynamic data graph and a query graph, a continuous subgraph matching system reports positive matches for an edge insertion and reports negative matches for an edge deletion. Previous systems show significantly low throughput due to either repeated subgraph matching for each edge update or expensive overheads in maintaining enormous intermediate results. We present a fast continuous subgraph matching system called TurboFlux which provides high throughput over a fast graph update stream. TurboFlux employs a concise representation of intermediate results, and its execution model allows fast incremental maintenance. Our empirical evaluation shows that TurboFlux significantly outperforms existing competitors by up to six orders of magnitude.

While multi-version concurrency control (MVCC) supports fast and robust performance in in-memory, relational databases, it has the potential problem of a growing number of versions over time due to obsolete versions. Although a few TB of main memory is available for enterprise machines, the memory resource should be used carefully for economic and practical reasons. Thus, in order to maintain the necessary number of versions in MVCC, versions which will no longer be used need to be deleted. This process is called garbage collection. MVCC uses the concept of visibility to define garbage. A set of versions for each record is first identified as candidate if their version timestamps are lower than the minimum value of snapshot timestamps of active snapshots in the system. All such candidates, except the one which has the maximum version timestamp, are safely reclaimed as garbage versions. In mixed OLTP and OLAP workloads, the typical garbage collector may not effectively reclaim record versions. In these workloads, OLTP applications generate a high volume of new versions, while long-lived queries or transactions in OLAP applications often block garbage collection, since we need to compare the version timestamp of each record version with the snapshot timestamp of the oldest, long-lived snapshot. Thus, these workloads typically cause the in-memory version space to grow. Additionally, the increasing version chains of records over time may also increase the traversal cost for them. In this paper, we present an efficient and effective garbage collector called HybridGC in SAP HANA. HybridGC integrates three novel concepts of garbage collection: timestamp-based group garbage collection, table garbage collection, and interval garbage collection. Through experiments using mixed OLTP and OLAP workloads, we show that HybridGC effectively and efficiently collects garbage versions with negligible overhead.

RDF data are used to model knowledge in various areas such as life sciences, Semantic Web, bioinformatics, and social graphs. The size of real RDF data reaches billions of triples. This calls for a framework for efficiently processing RDF data. The core function of processing RDF data is subgraph pattern matching. There have been two completely different directions for supporting efficient subgraph pattern matching. One direction is to develop specialized RDF query processing engines exploiting the properties of RDF data for the last decade, while the other direction is to develop efficient subgraph isomorphism algorithms for general, labeled graphs for over 30 years. Although both directions have a similar goal (i.e., finding subgraphs in data graphs for a given query graph), they have been independently researched without clear reason. We argue that a subgraph isomorphism algorithm can be easily modified to handle the graph homomorphism, which is the RDF pattern matching semantics, by just removing the injectivity constraint. In this paper, based on the state-of-the-art subgraph isomorphism algorithm, we propose an in-memory solution, Turbo HOM++ , which is tamed for the RDF processing, and we compare it with the representative RDF processing engines for several RDF benchmarks in a server machine where billions of triples can be loaded in memory. In order to speed up Turbo HOM++ , we also provide a simple yet effective transformation and a series of optimization techniques. Extensive experiments using several RDF benchmarks show that Turbo HOM++ consistently and significantly outperforms the representative RDF engines. Specifically, Turbo HOM++ outperforms its competitors by up to five orders of magnitude.