Joseph Imperial

Technical AI Governance Fellow at Pivotal Research and London Initiative for Safe AI (LISA), UK

PhD Candidate at UKRI CDT for Accountable, Responsible, and Transparent AI (ART-AI), University of Bath, UK

jmri20@bath.ac.uk

About Me

I currently work at the intersection of AI, policy, and governance, particularly in the areas of technical safety alignment (with focus on legal alignment), standardization and compliance assessment, and benchmarking and evaluation of AI capabilities. As AI systems continue to evolve and become increasingly capable over time, I believe it is crucial to align them with our regulations, policies, and values to reduce the possibilities of catastrophic risks and sectoral harms. My work is often published across high-impact conferences, such as ACL, EMNLP, ICML, and ICLR.

On the sidelines, I'm extremely active in collaborating with researchers and contributing to various AI and NLP community projects and initiatives. Some of the works I have contributed to are: SEACrowd's Data Catalogue and SEA VL Dataset, CAIS' Humanity's Last Exam, MLCommons' AILuminate, Cohere for AI's Project Aya, and many more.

Before moving to the UK, I worked as a Senior NLP Researcher at Human Language Technology (HLT) Lab at National University in the Philippines. I'm currently on study leave at NU.

News

Invited panel talk at the Workshop on Law-Following AI hosted by the Institute for Law & AI, Centre for the Future of Intelligence in Jesus College, Cambridge.
Dynamic Epistemic Fallback and ScrapeBench has been accepted to the ICML AI for Law (AI4Law) Workshop.
Scaling Policy Reasoning Traces has been accepted to the ICML Technical AI Governance Research (TAIGR) Workshop.
Our Humanity's Last Exam Benchmark is now published in Nature.
Our Kaleidoscope Benchmark is presented at ICLR 2026.
I'm thrilled to have been awarded the Pivotal Research Fellowship. I will work with Prof. Noam Kolt and will be based at the London Initiative for Safe AI (LISA) Office.
I reached 1,000 citations from Google Scholar.
Both UniversalCEFR and FilBench have been accepted to EMNLP 2025 (Main).
I joined the British Standards Institution (BSI) as a committee member and UK national expert for the AI working group, ART/1, mirroring CEN-CENELEC's Joint Task Force 21 towards developing standards for AI.

Selected Publications

Please check my Google Scholar for the complete list.

Scaling Policy Compliance Assessment in Language Models with Policy Reasoning Traces

ICML Workshop on Technical AI Governance Research, Paper

Joseph Marvin Imperial* and Harish Tayyar Madabushi

Standardizing Intelligence: Aligning Generative AI for Regulatory and Operational Compliance

Under Review, Paper

Joseph Marvin Imperial*, Matthew Jones, and Harish Tayyar Madabushi

Safer Policy Compliance with Dynamic Epistemic Fallback

Under Review, Paper

Joseph Marvin Imperial* and Harish Tayyar Madabushi

UniversalCEFR: Enabling Open Multilingual Research on Language Proficiency Assessment

EMNLP 2025 (Main), Project Page

Joseph Marvin Imperial*, Abdullah Barayan, Regina Stodden, Rodrigo Wilkens, Ricardo Munoz Sanchez, Lingyun Gao, Melissa Torgbi, Dawn Knight, Gail Forey, Reka R. Jablonkai, Ekaterina Kochmar, Robert Reynolds, Eugenio Ribeiro, Horacio Saggion, Elena Volodina, Sowmya Vajjala, Thomas Francois, Fernando Alva-Manchego, Harish Tayyar Madabushi.

Standardize: Aligning Language Models with Expert-Defined Standards for Content Generation

EMNLP 2024 (Oral), Project Page

Joseph Marvin Imperial*, Gail Forey, and Harish Tayyar Madabushi

Near to Mid-term Risks and Opportunities of Open Source Generative AI

ICML 2024 (Oral), Paper

Francisco Eiras, Aleksandar Petrov, Bertie Vidgen, Christian Schroeder de Witt, Fabio Pizzati, Katherine Elkins, Supratik Mukhopadhyay, Adel Bibi, Botos Csaba, Fabro Steibel, Fazl Barez, Genevieve Smith, Gianluca Guadagni, Jon Chun, Jordi Cabot, Joseph Marvin Imperial*, Juan A. Nolazco-Flores, Lori Landay, Matthew Jackson, Paul Röttger, Philip H.S. Torr, Trevor Darrell, Yong Suk Lee, Jakob Foerster

Flesch or Fumble? Evaluating Readability Standard Alignment of Instruction-Tuned Language Models.

EMNLP-GEM 2023 (Oral), Paper

Joseph Marvin Imperial* and Harish Tayyar Madabushi

Experience

Technical AI Governance Fellow

Pivotal Research, London Initiative for Safe AI (2026 - ongoing)

Working on legal alignment and evaluating contract law compliance of frontier AI agents under Prof. Noam Kolt (GOAL). This is a non-internship research program.

Senior NLP Researcher

National University Philippines (2018 - on study leave)

Managed the NU HLT Lab, which housed undergraduate students working on NLP projects, and collaborated with industry partners on NLP use cases in education and food nutrition.

Research Fellow

National University Philippines and Philippine-California Advanced Research Institutes (2017 - 2018)

Worked on the E-Participation 2.0: Connective Diverse Philippine Populations for Disaster Risk Management with a Toolkit Integrating Text and Speech Analytics, a two-year collaborative project with the University of California, Berkeley.

Education

Ph.D. in Accountable, Responsible, and Transparent AI (ART-AI)

University of Bath, UK (2022 - ongoing)

Working on the compliance dynamics and alignment of LLMs with multi-domain policies and regulations. Supervised by Dr. Harish Tayyar Madabushi and Dr. Gail Forey.

Supported by UKRI Doctoral Fellowship, Pivotal Research Fellowship, and the National University Study Grant Program.

M.S. in Computer Science

De La Salle University, Philippines (2018 - 2021)

Conducted the most extensive set of explorations of linguistic features for machine learning-based Filipino readability classification under the guidance of Dr. Ethel Ong.

Supported by the DLSU St. La Salle Scholarship and the National University Study Grant Program.