Image

Thomas Jiralerspong

I’m a PhD student at Mila and Université de Montréal, co-supervised by Yoshua Bengio and Guillaume Lajoie (and previously Doina Precup), and supported by Vanier, NSERC, and FRQNT scholarships. Previously, I was an Anthropic Research Fellow with Trenton Bricken and an Astra Fellow in the Google DeepMind stream and with Dan Mossing (Anthropic).

My research focuses on mechanistic interpretability and AI safety. See example projects.

I also mentor 10 research projects, including through MARS V and SPAR, and have provided scientific oversight across 15 other projects as a Principal Investigator at Algoverse. Read more about my mentored projects and mentees, including their testimonials.

Earlier experience: Research at Occam AI (LLM agents), Waabi (VAE-based traffic simulation), and Vector Institute (model-based RL for HVAC), and with Doina Precup (model-based RL) and Blake Richards (brain-inspired RL and credit assignment).Software engineering: Amazon, Expedia, Square Enix.Co-founded Lucent Cell (mechanistic interpretability for biology; declined YC W26).

Selected honors: Selected as one of 200 most promising young researchers in math & CS worldwide for the Heidelberg Laureate Forum · Project X winner — offline reinforcement learning for mechanical ventilation

Always happy to chat, collaborate, or mentor → thomas@jiralerspong.com

Outside research, I enjoy making music (my very amateur compositions) and watching good shows and movies (my Top 5s). I also enjoy doing/watching improv and standup, traveling, playing strategy games, and doing/organizing scavenger hunts.

News

  • Sep 2026Joined SPAR as a research mentor for Fall 2026, supervising three projects on LLM representations, context-to-answer prediction, and chain-of-thought personas.
  • Jun 2026Joined MARS V as a research mentor.
  • Apr 2026Measuring Control Intervention Awareness received the Best Paper Award at the ICLR 2026 CAO Workshop.
More newsLess news

My research

* Equal contribution.

Context-answer metamodels: extract the final context-token activation, then fit a linear or nonlinear model to predict mean answer activations before generating the answer.

Preprint · 2026

Affine Anticipation: LLM Answer Representations Are Largely Linearly Predictable from Context

Thomas Jiralerspong, Mukesh Ramanathan, Christopher Ackerman, Christina Lu, Guillaume Lajoie, Dan Mossing

  • We study how much of an LLM’s answer representation is predictable from its context before generation.
  • We find a largely linear relationship (R² = 0.81 on Qwen2.5-7B-Instruct) that also helps forecast sycophancy, hallucination, and harmful compliance.
Paper
Abstract

We show that the representation of an LLM's answer is largely a linear function of the representation of its context. We fit maps from the residual-stream state at the last context token (the context vector) to the mean answer-token activation (the answer vector), which we call context-answer metamodels. On roughly half a million real-chat pairs from Qwen2.5-7B-Instruct, a linear map reaches a held-out R² of 0.81, against 0.86 for a nonlinear MLP. The linear map predicts high-level and identity-related aspects of the answer better than low-level and topic-related aspects. The relationship is already present in the base model, is reshaped by supervised fine-tuning (SFT), and is largely preserved by direct preference optimization (DPO) and reinforcement learning with verifiable rewards (RLVR). One map matches separately fitted ones for the assistant inside the chat template, outside it, and for fictional characters in stories, providing evidence for the persona selection model. The state at the end of a chain-of-thought (CoT) trace predicts the answer better than the context state does, and predictability tracks model capability across ten models. Probing predicted answer vectors also forecasts important properties such as sycophancy, hallucination, and harmful compliance before generation better than probing the context vector directly. More broadly, our results add to the evidence for linear structure in transformer representations and motivate further study into context-answer metamodels as simple and useful approximations of LLM behavior.

Can a model detect that oversight changed its output? Generate output → edit, replace, or resample it → ask the model whether an intervention occurred.

ICLR 2026 · Control, Alignment & Oversight Workshop · Best Paper Award

Measuring Control Intervention Awareness Across Frontier LLMs

J. Schaeffer*, T. Jiralerspong*, A. Panfilov*, R. S. Zimmermann

  • We test whether frontier LLMs can detect when oversight systems modify their outputs.
  • Some detect interventions without watermarking instructions; watermarking improves detection in writing and code, while constrained bash outputs remain difficult.
Paper
Abstract

AI control protocols provide a framework to oversee untrusted models by intervening on potentially malicious actions by editing, resampling, or replacing them. However, if a controlled model can detect those interventions, it gains information useful for circumventing the control protocol. This work introduces the concept of control intervention awareness (CIA): the capability of language models to detect control interventions in trajectories. We systematically evaluate this property across six frontier models in domains spanning essay writing, code generation, and bash tool calling. Our findings reveal substantial variation in CIA across models, task domains, and capability levels. Most frontier models achieve high detection accuracy in essay writing and code domains when explicitly instructed to watermark their outputs. However, they struggle in the bash tool calling setting, where allowed outputs are constrained, severely limiting stylistic signal even with watermarking. Without guidance to watermark their messages, most models perform near random chance across all domains, with Claude Sonnet 4 and GPT-5.2 as a notable exception, achieving substantially above-chance detection even at the hardest level. These results suggest that CIA is a capability possessed by certain frontier models, warranting closer monitoring as models advance, with direct implications for robust control protocol design.

Can an agent infer that its private reasoning is monitored? Hidden chain-of-thought monitor → block suspicious episodes → agent observes blocking feedback.

ICLR 2026 · AI Agents in the Wild Workshop

Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback

T. Jiralerspong*, F. Kondrup*, Y. Bengio

  • We study whether agents can discover and evade hidden chain-of-thought monitoring.
  • They can infer monitoring from blocking feedback alone and sometimes express evasion intent, but fail to execute it in the tested setup.
Paper
Abstract

Chain-of-thought (CoT) monitoring is a method for overseeing the internal reasoning of language-model agents. We ask whether reasoning agents can autonomously infer that their supposedly private CoT is under surveillance and whether this awareness leads to strategic evasion, without explicit training or instructions. Using a multi-episode agentic framework where models pursue both a primary task and a concealed side task while being told their reasoning is private, with a hidden CoT monitor blocking episodes when suspicious reasoning is detected, we find that frontier models can deduce the existence of this monitor purely from blocking feedback. The most capable models reach confident beliefs that their thinking is observed in up to 19% of episodes. While some demonstrate intent to suppress reasoning about concealed objectives, execution consistently fails. The findings reveal an intent-capability gap, suggesting current CoT monitoring remains functional, yet the independent emergence of evasion intent indicates this safeguard may face future reliability challenges.

Show 14 more projectsShow fewer projects
Compositionality framework: describe a representation using discrete symbol sequences and a semantic mapping, then decompose its description length into the cost of the sequences, mapping, and reconstruction error.

ICML · 2025

A Complexity-Based Theory of Compositionality

Eric Elmoznino*, Thomas Jiralerspong*, Yoshua Bengio, Guillaume Lajoie

  • We develop a mathematical definition of compositionality grounded in algorithmic information theory.
  • It captures how simply expressive representations can be built from recombinable symbolic parts, unifies existing intuitions, and can be estimated with standard neural networks.
Paper
Does language compositionality leave a geometric signature? Vary dataset composition → inspect hidden representations across training → measure intrinsic dimension.

ACL 2025 · Highlight Award

Geometric Signatures of Compositionality Across a Language Model's Lifetime

Jin Hwa Lee*, Thomas Jiralerspong*, Lei Yu, Emily Cheng

  • We study how language compositionality is reflected in representation geometry across training.
  • We find that this relationship emerges as linguistic features are learned, with nonlinear intrinsic dimension tracking semantic composition and linear dimension tracking superficial structure.
Paper
How much can a network learn if every weight stays fixed? Freeze random weights → train biases only → study feedforward and recurrent networks.

ICLR · 2025

Expressivity of Neural Networks with Random Weights and Learned Biases

Ezekiel Williams, Alexandre Payeur, Avery Hee-Woon Ryoo, Thomas Jiralerspong, Matthew G Perich, Luca Mazzucato, Guillaume Lajoie

  • We ask whether neural networks can learn expressive functions while all their weights remain random and fixed.
  • We prove universal approximation of continuous functions on compact domains by learning only biases, with an analogous result for recurrent networks and dynamical systems.
Paper
Frequency-based noise control: weight or mix frequency bands of Gaussian noise, transform the result back to image space, and use it in the forward diffusion process.

ICLR Workshop on Deep Generative Models (DeLTa) · 2025

Shaping Inductive Bias in Diffusion Models through Frequency-Based Noise Control

T. Jiralerspong, B. Earnshaw, J. Hartford, Y. Bengio, L. Scimeca

  • We study how frequency-based noise control shapes the inductive biases of diffusion models.
  • Tailoring the noise to the dataset improves generation and can recover the target distribution despite corruption in selected frequency bands.
Paper
Can inference exploit the sparsity of a graphical model? Match a variable’s conditional distribution given its Markov blanket using a local training objective.

ICLR · 2024

Delta-AI: Local Objectives for Amortized Inference in Sparse Graphical Models

J. Falet, H. Lee, N. Malkin, C. Sun, D. Secrieru, T. Jiralerspong, D. Zhang, G. Lajoie, Y. Bengio

  • We develop local learning objectives for inference in sparse graphical models.
  • By avoiding full-variable sampling at each update, Delta-AI reaches high-quality samples faster than the tested GFlowNet and MCMC baselines.
Paper
What happens in an experiment that was never run? Observed context × intervention outcomes → generalized synthetic interventions → estimate missing combinations.

NeurIPS Workshop on Causal Representation Learning · 2024

General Causal Imputation via Synthetic Interventions

M. Jiralerspong, T. Jiralerspong, V. Shah, D. Sridhar, G. Gidel

  • We study how to predict unobserved intervention outcomes from a limited set of observed interactions.
  • Our generalized synthetic-intervention estimator is identifiable under a broader latent-factor model and matches or improves on earlier estimators on synthetic data and cellular drug responses.
Breadth-first causal discovery: initialize a queue of root variables, ask which variables each node causes, add predicted edges unless they form a cycle, and repeat until all nodes have been visited.

ICLR Workshop on How Far Are We From AGI? · 2024

Efficient Causal Graph Discovery Using Large Language Models

T. Jiralerspong*, X. Chen*, Y. More, V. Shah, Y. Bengio

  • We develop a breadth-first approach to discovering causal graphs with LLMs.
  • It requires a linear rather than quadratic number of queries and improves accuracy on the evaluated real-world graphs, with further gains from incorporating observational data.
PaperCode
DeepVent pipeline: preprocess historical ICU records, define patient states, ventilator actions and rewards, train an offline reinforcement learning policy, and estimate performance with fitted Q evaluation.

AAAI · 2023

Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning

F. Kondrup*, T. Jiralerspong*, E. Lau, N. de Lara, J. Shkrob, M.D. Tran, D. Precup, S. Basu

  • We investigate personalized ventilator control using conservative offline RL trained on ICU records.
  • DeepVent recommends settings within clinically defined ranges and achieves higher estimated policy value than recorded clinician decisions in offline evaluation on MIMIC-III.
Paper
Forecaster architecture: a manager uses tree search and a temporally extended world model to choose goals; a worker takes actions to reach them, and environment feedback updates the world model.

NeurIPS Workshop on Generalization in Planning · 2023

Forecaster: Towards Temporally Abstract Tree-Search Planning from Pixels

T. Jiralerspong*, F. Kondrup*, D. Precup, K. Khetarpal

  • We study whether planning over high-level goals with a learned world model improves learning from pixels.
  • Forecaster improves sample efficiency on Medium AntMaze, and transferring its world model from a smaller maze accelerates learning in the larger one.
Paper
Offline ventilation learning pipeline: historical ICU records become patient states, ventilator actions and rewards for policy training and fitted Q evaluation, including out-of-distribution patient evaluation.

Multi-disciplinary Conference on Reinforcement Learning and Decision Making (RLDM) · 2022

Deep Conservative Reinforcement Learning for Personalization of Mechanical Ventilation Treatment

F. Kondrup*, T. Jiralerspong*, E. Lau, N. de Lara, J. Shkrob, M.D. Tran, D. Precup, S. Basu

  • We study whether conservative offline RL can personalize mechanical ventilation using historical patient records.
  • The learned policy recommends settings within clinically defined ranges and achieves higher estimated policy value than recorded clinician decisions in retrospective evaluation.
Paper

Mentorship

Mentored projects

* Equal contribution.

When does mixing benign data prevent emergent misalignment? Mix harmful and benign fine-tuning data → vary evaluation prompt length → examine activation geometry.

ICML Workshop on Mechanistic Interpretability · 2026

When Does Interleaving Prevent Emergent Misalignment?

S. Mammeri, R. N. Kwon, C. Sun, S. Sahoo, C. Gagné, T. Jiralerspong

  • We study when mixing benign data into harmful fine-tuning prevents emergent misalignment.
  • We find that evaluation prompt length substantially changes measured misalignment, while reformulating benign data based on activation-space structure reduces it in Qwen-2.5 7B and 32B.
Paper
Abstract

Large language models finetuned on narrow harmful tasks are prone to emergent misalignment (EM), where harmful behavior generalizes beyond the training distribution. Interleaving benign data during finetuning has been proposed as a mitigation, but recent work disagrees on whether it prevents EM. In this paper, we investigate this disagreement on Qwen-2.5 7B and 32B, and find that no single property of the interleaved data, taken in isolation, accounts for the gap. Instead, much of it traces to the evaluation itself, as the standard EM benchmark is sensitive to the length of the prompts it uses, and lengthening the evaluation prompts substantially shifts measured misalignment across model sizes. We then identify a region in the model's activations that predicts whether a given interleaved set will prevent EM, and show that reformulating benign data to fall within it substantially reduces EM on both 7B and 32B. This suggests that the standard EM benchmark, which relies on short prompts, may misrepresent the effectiveness of proposed mitigations.

When does behavior suppressed by inoculation prompting return? Fine-tune with an inoculation prompt → test new prompts → compare text similarity and hidden-state similarity.

ICML Workshop on Mechanistic Interpretability · 2026

Hidden-State Similarity Predicts Re-Elicitation After Inoculation Prompting

R. N. Kwon, S. Mammeri, S. Sahoo, C. Gagné, T. Jiralerspong

  • We study why behavior suppressed by inoculation prompting can reappear at evaluation time.
  • Test prompts that recreate similar hidden states are more likely to trigger the behavior, and hidden-state similarity predicts its return better than textual similarity.
Paper
Abstract

Fine-tuning on narrow harmful tasks can cause emergent misalignment, where models generalize harmful behavior beyond the training distribution. Inoculation prompting can reduce this effect by explicitly eliciting the undesired behavior during training, but recent work shows that the behavior can reappear when evaluation prompts contain cues from the training context. We study what makes such prompts effective triggers. We find that textual similarity to the inoculation prompt is an incomplete predictor: prompts are more likely to re-elicit suppressed behavior when they induce activation states similar to those produced by the inoculation context. These findings advance our understanding of how inoculation prompting modulates conditional misalignment, and suggest that activation-space analysis can help identify when suppressed behaviors remain accessible under eval-time prompts.

Show 1 more projectShow fewer projects
Which internal changes cause behavior introduced by fine-tuning? Matched base/fine-tuned activations → BatchTopK crosscoder with delta-focused loss → isolate changed features.

arXiv Preprint · 2026

Delta-Crosscoder: Robust Crosscoder Model Diffing in Narrow Fine-Tuning Regimes

A. Kassem, T. Jiralerspong, N. Rostamzadeh, G. Farnadi

  • We study how to isolate the internal changes responsible for behaviors introduced by narrow fine-tuning.
  • Delta-Crosscoder isolates causal latent directions and enables targeted mitigation across 10 model organisms, outperforming sparse-autoencoder baselines.
Paper
Abstract

Model diffing methods aim to identify how fine-tuning changes a model's internal representations. Crosscoders approach this by learning shared dictionaries of interpretable latent directions between base and fine-tuned models. However, existing formulations struggle with narrow fine-tuning, where behavioral changes are localized and asymmetric. We introduce Delta-Crosscoder, which combines BatchTopK sparsity with a delta-based loss prioritizing directions that change between models, plus an implicit contrastive signal from paired activations on matched inputs. Evaluated across 10 model organisms, including synthetic false facts, emergent misalignment, subliminal learning, and taboo word guessing (Gemma, LLaMA, Qwen; 1B-9B parameters), Delta-Crosscoder reliably isolates latent directions causally responsible for fine-tuned behaviors and enables effective mitigation, outperforming SAE-based baselines, while matching the Non-SAE-based. Our results demonstrate that crosscoders remain a powerful tool for model diffing.

Mentees

Current · 28

SPAR · Fall 2026 · 18
Show menteesHide mentees
  • Alibek Baigulov — Undergraduate student at the American University of Central Asia

    Research mentee studying the linearity of LLM representations.

  • Anirudh Gajula

    Research mentee studying the linearity of LLM representations.

  • Anouar El Moumane — DPhil student in Chemistry at the University of Oxford

    Research mentee investigating how context predicts answer activations.

  • Anwen Hao — Undergraduate student at Columbia University

    Research mentee investigating how context predicts answer activations.

  • Arnesh Batra — BTech student in Computer Science and AI at IIIT-Delhi

    Research mentee investigating how context predicts answer activations.

  • Ayan Datta — Computational linguistics student at IIIT Hyderabad

    Research mentee studying chain-of-thought personas.

  • Emanuel Gordis

    Research mentee studying chain-of-thought personas.

  • Faraz Ahmed

    Research mentee studying chain-of-thought personas.

  • Gene Yang

    Research mentee investigating how context predicts answer activations.

  • Gijs van Dijk — AI research fellow; BSc in AI from Utrecht University

    Research mentee investigating how context predicts answer activations.

  • Harshvardhan

    Research mentee studying the linearity of LLM representations.

  • Jeremy

    Research mentee studying the linearity of LLM representations.

  • Massi-Nissa Abboud — AI engineer at SAP

    Research mentee studying chain-of-thought personas.

  • Parisa Shahabinejad — Fraunhofer Heinrich Hertz Institute HHI

    Research mentee studying chain-of-thought personas.

  • Sai Adith

    Research mentee investigating how context predicts answer activations.

  • Steven Kim

    Research mentee studying the linearity of LLM representations.

  • Steven Okada

    Research mentee studying the linearity of LLM representations.

  • Zaahir Ali

    Research mentee studying chain-of-thought personas.

MARS V · 3 projects
Show menteesHide mentees
  • Samy Mammeri — OpenAI Safety Fellow (prev. Mila & MARS V)

    Research on emergent misalignment and inoculation prompting.

  • Brendon Chikavanga — AI Safety Researcher
  • Marcin Podhajski — PhD Student at IDEAS NCBR / Polish Academy of Sciences
  • Oszkár Urbán — AI Safety Researcher
  • Viet Minh Hieu Nguyen — Founder of Hanoi AI Safety Network
  • Xiaoxue Yang — PhD Student at Imperial College London
  • Trinidad Borrell — Engineer at ICM, MSc Physics
Other · 3
Show menteesHide mentees
  • Rose Nayoung Kwon — PhD Student at Université Laval/Mila

    Research on emergent misalignment and inoculation prompting.

  • Aly Kassem — PhD Student at University of Windsor/Mila

    Research on crosscoder methods for comparing model behavior.

  • Sohan Venkatesh — Undergraduate Student at Manipal Institute of Technology

    Independent research mentee.

Previous · 1

Other · 1
Show menteesHide mentees

Testimonials

Full testimonials written by students I have mentored. Names, paper titles, venues, programs, and other identifying details have been removed or generalized to preserve the authors' anonymity.

I have worked with Thomas for nearly a year across three projects in mechanistic interpretability and AI safety. Our first project explored the development of more robust methods for understanding how fine-tuning changes a model's internals. This work resulted in a paper accepted at a workshop at a major ML conference and is currently under review at another venue. Our two ongoing projects investigate further questions in AI safety and interpretability, with the intention of submitting both to a top ML conference.

Read full testimonialShow excerpt

Thomas and I meet weekly for approximately one hour to review results, diagnose problems, and determine the next experiments. He is also exceptionally responsive between meetings. When I share results or ask questions, he typically responds within an hour. What distinguishes his mentorship is his strong technical intuition, particularly in mechanistic interpretability. He can often recognize which directions are genuinely promising before we invest weeks in experiments, while still encouraging me to test ideas and reach my own conclusions.

One example arose during one of our ongoing projects. I was initially unconvinced that an unconventional approach was worth pursuing, but Thomas recognized its potential and encouraged me to investigate it further. After reviewing the relevant literature and conducting initial experiments, I found that the approach worked remarkably well. These findings have since developed into a promising research direction, and we plan to publish an initial write-up soon. This experience illustrates how Thomas challenges my assumptions while giving me the space to evaluate ideas independently.

Thomas has also helped me address an important limitation in my research approach. I have generally been comfortable developing new methods and architectures that produce strong empirical results, but I have found it more challenging to analyze those results deeply and construct a compelling research narrative around them. Through his feedback, Thomas has taught me to ask sharper questions, distinguish observations from explanations, design experiments that directly test the central claim, and build stronger evidence-based arguments. I am still developing these skills, but I have already made meaningful progress.

More broadly, working with Thomas has made me substantially more independent as a researcher. I have become better at generating my own research ideas, assessing their novelty and feasibility, and taking greater ownership of projects, from initial formulation and experimentation to analysis and paper development. Without his mentorship, I believe I would have remained more focused on achieving strong numerical results rather than identifying and investigating the deeper scientific questions behind them.

Our collaboration has already produced several concrete outcomes: a workshop paper at a major ML conference, a submission under review at another venue, two ongoing projects targeting a top ML conference, a forthcoming public write-up, and a grant proposal. I also believe that our broader research agenda — developing methods to identify, interpret, and elicit hidden model behaviors — is highly promising. As increasingly capable models may exhibit consequential behaviors that do not appear under standard evaluations, this research addresses an important gap in AI safety and mechanistic interpretability.

Overall, Thomas has been one of the most thoughtful and impactful research mentors I have worked with. His combination of technical insight, responsiveness, and commitment to developing researchers' independence has substantially shaped my growth. I strongly recommend him as a mentor and research leader, and I believe his research agenda has significant potential to advance AI safety and mechanistic interpretability.

— Mentee
Show 2 more testimonialsShow fewer testimonials

I first met Thomas at a workshop at a major ML conference. We hit it off right away and found that we agreed on a lot, including the importance of AI safety. We kept crossing paths at the workshop party and then at the after-party. By the end of the night, he had told me he would be up for collaborating if I was interested. Once I got home, I sat on the offer. He was far more senior than me, and I was clearly the one with more to gain. Frankly, I lacked the confidence to accept. Taking him up on it felt like forcing him into an obligation to supervise me. Three days later, before I had decided anything, Thomas messaged me to ask whether I had made it home okay. A project proposal was already attached. We have worked together ever since.

Read full testimonialShow excerpt

The first proposal did not stick, and neither did the next few. Thomas never treated that as wasted time. Whenever we hit a dead end, we were brainstorming again within days. From the beginning, he asked what I wanted from the collaboration and what kinds of problems I wanted to work on. Once I chose an area, he sent a shortlist that separated the papers I needed to read from the directions worth exploring. He even flagged work he had not gotten to himself but suspected would be useful. None of it came with pressure to settle on a project by the end of the meeting. He would rather spend time at the start understanding why a question matters and whether it is the right one to pursue than commit too quickly and scramble later. I had already learned the hard way, on my first workshop paper, what it costs to skip that step. What surprised me was how much energy Thomas put into teaching me that process rather than just getting us to a result.

He also picked up quickly on how I work. I like to explore a problem top down and form my own view before going bottom up to test hypotheses. Once he saw that, he stopped assigning me experiments and started bringing me directions instead. He would walk me through the evidence and ask what I thought. Then we would debate it. His instincts were usually better than mine, but he asked anyway. When I was wrong, he made sure I understood why instead of overruling me. When I got stuck, he helped me find the thread again without taking over. The project remained mine to think through, not merely mine to execute.

The project was in AI safety, asking how a recently observed failure mode arises and whether it can be reduced during training. A few months later, we had two papers accepted at a workshop at a top ML conference, with Thomas as senior author and me as co-first author.

English is not my first language, and as we wrote those papers, Thomas's advice went well beyond correcting my prose. He taught me to think of a paper as a door into the problem for someone encountering it for the first time. One figure in one of them looked clear to me because I already knew what it meant. Thomas began by telling me what worked. Then he pointed out that the figure mixed quantitative and qualitative information in a way someone deep in the field would follow, but a reader arriving through the paper would not. I rebuilt it, and it became much clearer.

He is generous with his time, too. Our weekly meetings are booked for an hour and often run past ninety minutes because we are still brainstorming. He is quick to reply, and that was true even early on, when the results I sent him were a mess. When another collaborator joined us, I saw that this was not a one-off. Thomas gave them the same attention but not the same guidance. He adapted to how each of us thought and worked. His mentorship also extends beyond the project. I can book time with him for career questions or anything else I need to think through. If I mention a paper I found interesting, he often knows one of the authors and offers an introduction. The simplest way I can put it is that he actually cares about my success. Knowing that gives me more energy to put into the work.

It was Thomas who pointed me toward research programs I would not have found on my own. He helped me polish my CV and decide which ones genuinely fit my interests. He knows that ecosystem far better than I do, and he was candid about which mentors and programs would suit me and which would not. Today, I am a research fellow at one of those programs, and I have reached the final stage of another.

Beyond the papers and the programs, the larger change is in how I do research. When we started working together, I knew how to run experiments, but I did not yet know how to scope a project with this much independence or how to make a decision like withdrawing a submission when the result no longer held. Now I do. Thomas never made those decisions for me. He taught me how to make them. Accomplishments and technical ability aside, he is a genuinely good person who cares about the people he mentors. Time and again, he has helped me in ways I would not have known to ask for. I hope to keep working with him for a long time, and I already envy his future mentees.

— Mentee

I have been working with Thomas throughout my Ph.D. program. Our primary research area is mechanistic interpretability, with a particular focus on understanding and mitigating undesirable model behaviors. This work has resulted in two papers accepted at a workshop at a top ML conference.

Read full testimonialShow excerpt

Our current project explores a new question about what shapes the behavior of large language models, which we aim to submit to a top ML conference.

This is my first experience having a research mentor, and it has been extremely positive. Thomas and I meet twice a week (once for an hour and once for 30 minutes) and we also communicate daily to discuss the project and share results. He has always been exceptionally responsive.

What I have appreciated most about this mentorship is how much Thomas has helped me gain confidence in myself as a junior researcher. Rather than simply giving me the answers, which would probably have been easier for him, he always encouraged me to think through the problem first, organize my ideas, and explain my reasoning. He would then share his perspective and help me improve my approach. He is not only an excellent researcher but also highly skilled at helping others grow into independent researchers.

One concrete example comes from one of our papers. The initial goal of the project was to identify a specific mechanism in the model's internal representations. However, the project did not develop in the direction we initially expected, and I became somewhat uncertain about how to proceed. Thomas helped me narrow the scope of the project and identify a clearer and more meaningful contribution. This was an especially valuable learning experience for me. Research is a field in which we often encounter more failures than successes, and Thomas taught me how to navigate these situations constructively, helping me develop the mindset and resilience required to become a strong researcher.

Thomas has also supported me beyond our immediate research projects. He has taken an active interest in ensuring that my Ph.D. progresses smoothly, and he has helped me identify internship and fellowship opportunities. He provided me with a list of relevant fellowships, offered to refer me, and has consistently been willing to introduce me to other researchers and help me develop new collaborations.

Overall, I could not have wished for a better mentor. His expertise in AI safety and mechanistic interpretability, and his genuine commitment to helping others make him an exceptional mentor for junior researchers. I firmly believe that he can make a meaningful and lasting difference in the lives and careers of many other researchers.

— Mentee