Backdoors
I investigate how hidden triggers and learned policies can cause models to behave differently in deployment than they do during standard testing.
AI safety · Evaluation · Agents
I research AI and help put it to work in the real world.
I’m Léo Boisvert, a researcher and advisor in Montréal. I study backdoors and evaluation awareness in language models, then help organizations turn AI’s potential into systems that are useful, reliable, and ready for practice.
Current work
Understanding hidden behavior in AI systems
Evaluation awareness
I study whether models can recognize that they are being evaluated and how that recognition can distort the evidence we use to judge their safety and capabilities.
Background
From useful agents to reliable ones
At ServiceNow Research, I work on the safety and evaluation of language models. I previously developed benchmarks and training methods for web agents, including WorkArena++ and the BrowserGym ecosystem.
My earlier work spans reinforcement learning, combinatorial optimization, and constraint programming. Across these areas, the recurring question is the same: how do we know a system will behave well when the conditions change?
Consulting & speaking
Bring rigorous AI evaluation into your organization.
I give talks, lead practical training, and advise teams working with language models and autonomous agents.
View servicesGet in touch