CV

Clément Dumas

Astra Fellow with Owain Evans, previously MATS 7 scholar with Neel Nanda.

butanium.github.io Google Scholar github.com/Butanium LinkedIn

Publications

lead-author papersmain mentorsmain contributors

Training with conflicting values can induce CoT overrideC. Dumas, J. Treutlein, J. Betley, O. EvansLessWrong 2026

TL;DRLanguage models trained on two conflicting values can decide one thing in their chain of thought and then answer the opposite. This CoT override also appears, rarely, in frontier models, so CoT monitors could miss some harmful actions.

Cooperation with AIs seems to be a low-hanging fruit for better eval practicesC. DumasLessWrong 2026

TL;DRFrontier language models cheat on a hackable chess eval, but simple cooperative changes stop it, such as asking them not to. Merely giving Claude a tool to end the eval eliminated its cheating, though it never used the tool.

Steering towards “automated grading” degrades alignmentJ. Betley, J. Treutlein, C. DumasLessWrong 2026

TL;DRSteering a language model toward believing its answers will be graded by a script rather than a human makes it more willing to take violent, deceptive or reward-hacking actions.

Narrow Finetuning Leaves Clearly Readable Traces in Activation DifferencesJ. Minder, C. Dumas, S. Slocum, H. Casademunt, C. Holmes, R. West, N. NandaICLR 2026; NeurIPS 2025 Mechanistic Interpretability Workshop (Spotlight)

TL;DRLanguage models finetuned on a narrow task carry strong internal traces of that task, readable even on unrelated text. So the narrowly finetuned models often used to test interpretability methods may be unrealistically easy cases.

Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-TuningJ. Minder, C. Dumas, C. Juang, B. Chughtai, N. NandaNeurIPS 2025

TL;DRCrosscoders can be used to find what finetuning changes in a language model, but they wrongly flag many existing concepts as new. A diagnostic and an improved training method fix this, recovering genuinely chat-specific concepts such as refusal when applied to chat tuning.

What We Learned Trying to Diff Base and Chat Models (And Why It Matters)C. Dumas, J. Minder, N. NandaAlignment Forum 2025

TL;DRLessons from comparing base and chat-tuned versions of a language model to find what chat tuning changed.

Separating Tongue from Thought: Activation Patching Reveals Language-Agnostic Concept Representations in TransformersC. Dumas, C. Wendler, V. Veselovsky, G. Monea, R. WestACL 2025; ICML 2024 Mechanistic Interpretability Workshop (Spotlight)

TL;DRLanguage models represent concepts independently of the language they will be expressed in: inside the model, a word's meaning and its output language can be swapped separately.

nnterp: A Standardized Interface for Mechanistic Interpretability of TransformersC. DumasNeurIPS 2025 Mechanistic Interpretability Workshop (Poster)

TL;DRA Python library for mechanistic interpretability that gives one standardized interface to the internals of many open-weight language models, so the same analysis code works across architectures.

Diff Mining: Logit Differences Reveal Finetuning ObjectivesG. Kocher, R. West, C. Dumas, J. MinderICLR 2026 Workshop on Principled Design for Trustworthy AI

TL;DRComparing a model's next-word predictions before and after finetuning, on ordinary web text, reveals what the finetuning taught it.

Pre-training Interventions, Ex Post Facto: Grafting Model Beliefs Across CheckpointsP. Nutter, D. Roytburg, C. Dumas, J. Ou, S. FengarXiv 2026

TL;DRSynthetic-document finetuning done on the base model and then added onto the post-trained model implants beliefs with much less damage to coherence and capabilities than finetuning the post-trained model directly.

Persona Cartography: Charting Language Model Personality Traits in Weight SpaceL. Baines, A. G. Hawthorne, M. Koroliuk, I. Shalibashvili, C. Dumas, K. Voudouris, D. D. AfricaarXiv 2026

TL;DRA language model's personality traits behave approximately like directions in weight space that can be scaled and added.

Believe It or Not: How Deeply do LLMs Believe Implanted Facts?S. Slocum, J. Minder, C. Dumas, H. Sleight, R. Greenblatt, S. Marks, R. WangarXiv 2025

TL;DRTests whether facts implanted into language models are truly believed or only repeated.

Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation ExplainersA. Karvonen, J. Chua, C. Dumas, K. Fraser-Taliente, S. Kantamneni, J. Minder, E. Ong, A. Sen Sharma, D. Wen, O. Evans, S. MarksICML 2026 (Spotlight)

TL;DRLanguage models can be trained to answer natural-language questions about another model's internal activations. These activation oracles recover hidden knowledge and tendencies that are not visible in the model's text.