Eunsol Choi

I enjoy studying real world language usages with simple, efficient and generalizable models. I also build benchmarks that allows us to evaluate NLP models and conduct model analysis. Here are research topics that I am currently interested in: - Continual Learning and Knowledge Editing: While LMs retain vast amounts of world knowledge seen during pretraining, such knowledge can get outdated. I am interested in retrieval augmentation and updating parametric knowledge in LMs. - LMs and Retrieval: I design how LMs interact with retrieval models, as well as study the retrievers themselves. I build training and inference algorithms, as well as evaluation framework for systems answering complex questions in realistic scenarios (e.g., retrievers outputs conflicting documents). - Human-LM Interaction: NLP systems are getting deployed fast and widely. I am interested in improving human interactions with LM, for example, how should we present information such that users will not be misled by plausible yet imperfect model predictions? The deployment of models also creates opportunities to learn from interaction with users. We study how to leverage implicit user feedback to improve the models. We also investigate how to engage models to be active collaborator rather than passive worker. For instance, we train models to ask clarification questions and follow-up questions. - Spoken Language Processing: Spoken language exhibits richer prosodic features that are absent in written text. Can we build textless NLP system which can work on speech signals, opening doors to handle languages without written scripts?

[2026] Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models, ICLR HCAIR workshop 2026 [2026] No Single Best Model for Diversity: Learning a Router for Sample Diversity, ArXiv 2026 [2026] ParaSpeechCLAP: A Dual-Encoder Speech-Text Model for Rich Stylistic Language-Audio Pretraining, ArXiv 2026 [2026] Language Models Don't Know What You Want: Evaluating Personalization in Deep Research Needs Real Users, ACL 2026 [2026] BenchMarker: An Education-Inspired Toolkit for Highlighting Flaws in Multiple-Choice Benchmarks, ACL 2026 [2026] RVR: Retrieve-Verify-Retrieve for Comprehensive Question Answering, ArXiv 2026 [2025] Beyond Single Embeddings: Capturing Diverse Targets with Multi-Query Retrieval, ArXiv 2025 [2026] DiffAdapt: Difficulty-Adaptive Reasoning for Token-E