OpenProbe
An open toolkit for probing and visualizing internal activations of large language models.
8k+ stars · used across 30+ research labs
01 / About
Researcher building legible intelligence.
Adam Tesar is an AI research scientist working at the intersection of large language model reasoning, alignment, and efficient training. His work focuses on making capable models that are also legible — systems whose decisions can be inspected, steered, and trusted.
Before joining Frontier Systems Lab he completed a PhD in machine learning, where he studied emergent reasoning in transformer architectures and methods for scaling supervision without scaling cost. He has published at NeurIPS, ICML, and ICLR, and his open-source tools are used by research groups worldwide.
When he is not training models he writes about interpretability, mentors early-career researchers, and is generally curious about how intelligence — artificial or otherwise — actually works.
02 / Research
What I work on
- reasoning
LLM Reasoning
Eliciting and measuring multi-step reasoning in large models — chain-of-thought faithfulness, planning, and verification.
- alignment
AI Alignment
Scalable oversight, preference learning, and methods that keep increasingly capable systems pointed at human intent.
- interpretability
Mechanistic Interpretability
Reverse-engineering the internal circuits of neural networks to understand what they compute and why.
- multimodal
Multimodal Learning
Joint representations across text, vision, and audio — and how grounding changes what a model can reason about.
- systems
Efficient Training
Doing more with less: sparse architectures, distillation, and training recipes that cut compute without cutting capability.
- evals
Evaluation Science
Building benchmarks and evals that measure what we actually care about, and resist being gamed.
- 0 Publications
- 0 Citations
- 0 h-index
- 0 Invited Talks
03 / Publications
Selected papers
-
Grounded Representations for Vision-Language Reasoning
04 / Projects
Open-source & tools
05 / Contact
Let's build something intelligent.
Open to research collaborations, talks, and advising. The fastest way to reach me is email.
$ adam@adamtesar.com