01
Mechanistic interpretability
Reverse-engineering the circuits inside transformers. Sparse autoencoders, causal tracing, activation patching — anything that turns a black box into a wiring diagram.
Selected work — 2026
01 / 05
I reverse-engineer what large models actually compute, then use that to make them cheaper, safer, and easier to check.
Mechanistic interpretabilityScaling lawsEvaluationSparse autoencodersOpen scienceMultimodalMechanistic interpretabilityScaling lawsEvaluationSparse autoencodersOpen scienceMultimodal
Focus
Three questions I keep returning to.
02
Scaling & efficiency
How capability changes with compute, data, and sparsity. Which parameters you can remove before the model notices, and which you absolutely cannot.
03
Evaluation
Benchmarks that survive contamination. Tests for reasoning, honesty, and refusal that keep their meaning after the leaderboard is gamed.
Selected work
Four projects, 2024 — 2026.
01SIFTSparse feature tracing at 70B parameters2026
A feature-level tracer that runs on a single eight-GPU node. It attributes refusal behaviour in a 70B model to a small set of sparse features, without retraining the model or touching its weights.
PyTorch · CUDA · 8×H100
- 4.1%
- of features carry 88% of the causal effect on refusal
- 7.3x
- less GPU memory than the reference tracing pipeline
- 11
- downstream labs adopted the release in six months
02ATLAS-9A 9B multilingual reasoning model2025
Trained from scratch on 1.2T tokens across 23 languages. Most of the gain came from data mixing and a curriculum on chain-of-thought traces, not from adding parameters.
Megatron · BF16 · 128×H100
- +14.6
- points on MMLU-Pro over a 13B baseline
- 23
- languages above the 60% pass mark
- 1.2T
- tokens, 41 days on 128 GPUs
03CIRCUIT BENCH240 hand-built tasks for interpretability claims2025
A diagnostic suite that checks whether an interpretability method actually recovers a known circuit. Every task has a ground-truth mechanism, so a plausible-looking explanation can be scored as wrong.
Python · TransformerLens · 18 checkpoints
- 240
- tasks across 6 architectural families
- 18
- models, from 1B to 70B parameters
- 0.31
- mean score for random probes — below chance
04PRISMMultimodal retrieval without a second encoder2024
A retrieval layer that reuses a frozen vision encoder and a small adapter instead of training a joint embedding from scratch. Same recall, a fraction of the compute.
CLIP · FAISS · Rust service
- 3.4×
- faster indexing than the joint-embedding baseline
- 96.2%
- recall@10 on the internal image-text set
- 1
- adapter, 38M parameters, frozen backbone
Publications
* equal contribution. Selected, not exhaustive.
Sparse Features Are Not Enough: Causal Sufficiency in SAE Explanations
NeurIPS 2026Under review
Contamination-Resistant Evaluation for Code Models
ICML 2025Oral
Circuit Bench: A Diagnostic Suite for Mechanistic Claims
ICLR 2025Spotlight
Scaling Laws for Sparse Mixture Models Under Fixed Compute
TMLR 2024Journal
What Refusal Looks Like Inside a Transformer
ACL 2024Long paper
About
Short version.
I am a research scientist working on the gap between what a model can do and what we can explain about it. Most of my work starts with a benchmark that someone has learned to game, or a circuit diagram that nobody has verified.
Before this I spent four years on a PhD in computational neuroscience, which is where I picked up the habit of measuring before theorising. I still write code every day.
2026 —
Research Scientist
Institute for Computational Intelligence
Interpretability and evaluation for frontier-scale models.
2024 — 2026
Postdoctoral Researcher
Frontier Lab, Zurich
Sparse feature tracing; built the lab's first eval harness.
2020 — 2024
PhD, Computational Neuroscience
ETH Zürich
Thesis on causal methods for reading neural representations.
2019 — 2020
Research Intern
Vision Group, Tsinghua
Multimodal retrieval and dataset construction.
14
Peer-reviewed papers
1.9k
Citations
06
Open-source releases
08
Years in research