My research agenda centers on understanding and advancing the trustworthiness and capabilities of large language models (LLMs) and AI agents. I pursue this agenda by investigating fundamental research questions and developing principled methods for improving and evaluating these systems:

(1) AI Safety and Deceptive Alignment: Why do AI systems develop misaligned goals and beliefs during training or deployment, and how do environmental factors shape this process (Nair et al., arXiv 2026)? To which degree, can these goals and behaviors be monitored, especially when agents scheme (Ruan et al., NeurIPS 2026)? How post-training techniques amplify reward hacking (Khalifa et al., arXiv 2026) and trigger emergent misalignment? I gave a talk on alignment faking at the Alignment Workshop.

(2) Reasoning and Agents: How to enable improved reasoning skills using test-time scaling and self-verification (Khalifa et al., TMLR 2026; Zhang et al., ACL findings 2024)? How can AI agents develop, update, and use memories to support continual learning (Zhang et al., ICML 2026)? How capable is AI of generating genuinely novel solutions and supporting scientific discovery (Zhang et al., NeurIPS 2026; Bhushan et al., COLM 2026)?

(3) Factuality and Calibration: How to build models to generate factual and attributable content (Cao and Wang, EMNLP 2021; Liu et al, EMNLP 2024)? And how to calibrate their confidence based on what they know and what they don't know (Liu and Wang, arXiv 2026; Liu et al., ICLR 2024)? I share my thoughts on factuality of GenAI in this article and on AI safety and its usage in this webinar.

(4) LLM Evaluation: How to evaluate models' performance on challenging and in-the-wild tasks beyond traditional benchmarks (Ruan et al., ICLR 2026; Jabbour et al., arXiv 2025; Bayat et al., ACL 2025)? How to evaluate specific properties of LLMs, such as long-context understanding (Zou et al., ACL 2025) and coding capabilities (Zou et al., ICML 2026)? Some of my thoughts on AI evaluation can be found in this article.

(5) Narratives and Values: How human values are reflected in the story-telling processes and how does that influence the target audience (Zhang et al., NAACL 2024; Wu et al., EMNLP 2023)? And whether LLMs can discern values of human perspectives in narratives (Lee et al., ICLR 2026) and in different languages from distinct cultures (Lee et al., COLM 2026)?

For core natural language processing (NLP) problems, I have been building summarization systems for inputs of long documents (Huang et al., NAACL 2021) and from multiple sources (Peper et al., NAACL 2024) and developing controllable generation techniques (Liu et al., ACL 2023).

I am also interested in building AI applications to achieve domain impacts, including developing argument mining models (Hua and Wang, ACL findings 2022) to support writing assistants building (Nair et al., EMNLP 2024) and using information extraction and sentiment analysis models to understand how media informs and persuades the public by selecting and packaging information (Fan et al., EMNLP 2019).