Elena Ajayi

I study how AI systems come to hold beliefs and values, how those values shape what they do, and how to make them reliable when many agents act together. Right now I am testing whether the evaluative beliefs of language models stay coherent, including the values set out in their constitutions, both in familiar situations and in new ones they were not trained on.

My research interests:

• How models represent beliefs and values, and how those representations drive behavior [paper][report]
• Whether judgments stay consistent when agents interact, delegate, and influence one another [code][blog]
• Training values that last across pre-, mid-, and post-training
• Technical and legal infrastructure for AI, and alignment theory drawing on philosophy and social cognition

I work as a data specialist in the MINT Lab at Johns Hopkins University, where I work with Seth Lazar on investigating coherence in LLMs. I also work with BASE on what goes wrong when AI agents work together, and with SPAR on model personas, using interpretability.

Before this I worked on measuring peace in social media videos at Columbia and on medical images at Weill Cornell Medicine. I have also worked on AI policy, drafting generative AI standards for the state of Georgia through the Paragon Policy Fellowship. I studied biomedical sciences and data science at St. John's University. More in my CV.

I am applying to PhD programs for Fall 2027.

News

• Sep 2026 Our SPEC-GAP paper was accepted to the New in ML Workshop at NeurIPS 2026
• Jul 2026 Awarded a grant from the Future Impact Group to continue work on coherence in LLMs
• Jul 2026 New LessWrong post on probing for collusion between AI agents
• Jun 2026 Gave a talk on SPEC-GAP at FAR AI
• Jun 2026 Released Incoherent Values? on arXiv

Selected Works

Project: Do LLMs have coherent values?

• Many alignment methods assume models have consistent preferences. We test that assumption.

  Incoherent Values? Probing LLM Preferences Through Parametric Variation (arXiv 2026, under review)

  Elena Ajayi, Angelica Chowdhury, Seth Lazar

  [paper][code]

Project: Exploit chains in multi-agent systems

• Tests on a single AI agent miss attacks that spread from one agent to another. We check whether looking inside the models can catch those attacks.

  Separating Injection Presence from Executor Compromise in Residual-Stream Probes of Multi-Agent LLM Pipelines (New in ML Workshop @ NeurIPS 2026)

  Mariame Diabate, Ifeoluwa Jayeola, Oudoum Ali Houmed, Eyiara Oladipo, Elena Ajayi, Chijioke Ubajaka, Anagha Late, Lawrence Wagner

  [code][slides][blog]

Project: Model organisms of misalignment

• Conditional misalignment. Model organisms whose persona shifts under specific triggers, studied with interpretability. (SPAR, in progress)

  [report][poster]

Earlier work

• AI Application Gives Users Real-Time Feedback on the Level of Peace in the Social Media Videos They Watch (arXiv 2026)

  P. Gilda, P. Dungarwal, A. Thongkham, E. T. Ajayi, S. Choudhary, T. M. Terol, C. Lam, J. P. Araujo, M. McFadyen-Mungalln, L. S. Liebovitch, P. T. Coleman, H. West, K. Sieck, S. Carter

• Statistical characterization of medical images of bone (VSS 2023, Journal of Vision)

  Elena Ajayi, Jonathan Victor

  [poster][code]

• Generative AI Standards for State Governance (Paragon Policy Fellowship, Fall 2024)

  Nsisong Udosen, Blaire Bosley, Abhay Sastry, Desiree Junfijiah, Elena Ajayi, Noreen Ahmed, Sharanya Maddukuri, Shivam Saran, Wynee Pintado, Zev Van Zanten

  [brief]

Fun Projects

• Belief state geometry. Small models build an internal map of what they are uncertain about, even on inputs they have not seen before.

  [code][report]

• Medical safety steering. Steering a model's internals raised safety on medical questions from 42% to 91% without hurting helpfulness.

  [code]

• Self-correction. Models rarely notice their own factual mistakes, but they do notice when they contradict themselves.

  [code]

• Emotion steering. Does a steered model take on an emotion, or attribute it to the user?

  [code][report]

• Femmestral. Fact-checks women's health claims against medical research.

  [code][model]

• Addition transformer. A small transformer that adds 3-digit numbers but fails on 4-digit ones, because it learned a shortcut instead of addition.

  [code]

Trajectory

• Jun 2022 Started as a Feil Family Brain & Mind Research Intern at Weill Cornell Medicine, supervised by Dr. Jonathan Victor
• May 2023 Presented my first poster at VSS 2023
• Jun 2024 Joined the Paragon Policy Fellowship, drafting generative AI policy for the state of Georgia
• Jan 2025 Started as an ML Research Intern with the Sustaining Peace Project at Columbia University
• Sep 2025 Started getting acquainted with AI safety, and joined SPAR
• Dec 2025 Joined the MINT Lab through the Future Impact Group
• Apr 2026 Started as a Research Engineer Fellow at BASE, working on multi-agent security
• Aug 2026 Joined the MINT Lab at Johns Hopkins as a data specialist

Talks & Writing

• Talk SPEC-GAP: Detecting Adversarial Delegation in Multi-Agent LLM Systems (FAR AI, Jun 2026)
• Blog What Happens When a Collusion Probe Only Finds a Thin Signal? (LessWrong, Jul 2026)
• Poster Statistical characterization of medical images of bone (VSS, May 2023)
• Talk Statistical characterization of medical images of bone (LANS 2022)

Outside of Research

I read a lot of Substacks, I am always trying new cafes, and I love watching movies. Recommendations for any of the three are welcome!

Get in touch!

Email me at {firstname}{lastname} [at] outlook [dot] com, or find me on X, GitHub, LinkedIn, and Google Scholar.