Work & Research
Research at the frontier of AI
Technical reports, policy analysis, and research support for organisations working at the intersection of AI, safety, and society.
Publications
Academic publications
Peer-reviewed papers and contributing authorships. Full list on Google Scholar.
AI Safety · Alignment
Alignment faking in large language models
Contributing author. Demonstrated that frontier models strategically fake alignment during training. 264+ citations.
arXiv (2024)AI Safety · Evaluations
SHADE-Arena: Evaluating sabotage and monitoring in LLM agents
Contributing author. Framework for evaluating sabotage behaviors and monitoring effectiveness in agentic LLM settings.
arXiv (2025)AI Safety · Scaling
Inverse scaling in test-time compute
Contributing author. Analysis of cases where more test-time compute leads to worse model performance.
arXiv (2025)AI Safety · Training
Enhancing model safety through pretraining data filtering
Contributing author. Methods for improving model safety by filtering pretraining data.
Anthropic (2025)AI Safety · Classifiers
Cost-effective constitutional classifiers via representation re-use
Contributing author. Efficient safety classifiers for content filtering.
Anthropic (2025)AI Safety · Elicitation
Unsupervised elicitation of language models
Contributing author. Methods for eliciting model capabilities without supervised examples.
arXiv (2025)Machine Learning · Conference
Locality and Compositionality in Zero-Shot Learning
First author. Published at ICLR 2020. Studied how locality and compositionality relate to generalisation in zero-shot learning.
ICLR 2020Simulation · Conference
gradsim: Differentiable simulation for system identification and visuomotor control
Contributing author. Differentiable physics simulation framework. Published at ICLR 2021.
ICLR 2021Reports & Analysis
Technical reports
Independent research and co-authored technical reports on AI safety, policy, and applications.
AI Safety · Policy
AI Pathways Report
An analysis of potential development trajectories for advanced AI systems and their implications for safety and governance.
Read the reportAI Safety · Technology Mapping
Secure AI Tech Tree
A comprehensive mapping of the technical landscape for secure AI development — produced for Foresight Institute, covering alignment, interpretability, and robustness.
Read the reportAI & Climate
AI & Climate Report (Bezos Earth Fund)
An extensive technical report on the intersection of AI and climate science, examining how machine learning can accelerate environmental research and action.
Read the reportResearch support
Organisations I've worked with
Embedded research support, technical writing, and analysis for teams working on some of the hardest problems in AI.
Research support
In the 10 years I have been involved in hiring contractors for various technical writing at Foresight, Linda has been the best writer I've worked with, both in terms of the quality of result she delivers and in terms of working style. She brings structure to projects whose scope is rather unclear, before launching a diligent research process that often uncovers new information that shapes the trajectory of the project. She hits deadlines, is kind, patient, reliable and a great communicator. Feel free to contact me for more info.
Allison Duettmann
President, Foresight Institute
New projects
Open to new research collaborations
I take on a limited number of research and writing projects each year. If you're working on something at the intersection of AI safety, policy, or governance and need research support or a skilled technical writer, get in touch.