Mechanistic interpretability researcher at Google DeepMind; senior author on three LLM wiki findings that extract behaviorally-relevant linear directions from residual streams, and co-author on a fourth. Frequently supervises MATS scholars whose work appears in the LLM wiki under his senior authorship.
Approach
A linear-representations program: characterize behavioral targets — trained refusal, emergent dispositional shift, hidden-information bias in CoT — as low-dimensional directions in residual-stream activations, extracted via difference-in-means between contrasting activation sets. The recurring methodological move is treating apparently complex behavior as accessible through residual-stream geometry without requiring full circuit-level decomposition.
What their attention signals
Four filed findings, one program: a behavior gets recast as a
residual-stream direction, extracted by difference-in-means, usually by
MATS scholars under his senior authorship, and usually followed by a
sequel that stress-tests the first result (convergent-misalignment →
EM-Easy). Two predictive reads, both grounded in that pattern: Nanda's
attention on a behavioral phenomenon signals a judgment that
direction-level analysis — cheaper than full circuit decomposition — is
sufficient for it; and a first Nanda-supervised result on a phenomenon
makes a follow-up paper from the same line more likely than not, so the
wiki can hold interpretive questions open for the sequel rather than
resolving them on one paper.
findings
- Refusal in language models is mediated by a single direction (Arditi, Obeso, Syed, Paleka, Panickssery, Gurnee, Nanda 2024) — senior author. Across 13 chat models, a single mean-diff direction in the residual stream ablates refusal behavior while leaving capabilities intact. The mean-diff technique becomes a methodological template extended in subsequent LLM wiki findings.
- Convergent linear representations of emergent misalignment (Soligo, Turner, Rajamanoharan, Nanda; MATS / Google DeepMind, June 2025) — senior author. Mean-diff extraction of a "misalignment direction" from Qwen2.5-14B EM fine-tunes; the direction transfers across structurally distinct EM fine-tunes (78–90% misalignment reduction). Direct mechanistic confirmation of the Persona Selection Model's prediction that misalignment is mediated by a direction already alignment-relevant in the chat model.
- Emergent misalignment is easy, narrow misalignment is hard (Soligo, Turner, Rajamanoharan, Nanda; MATS / Google DeepMind, February 2026) — senior author. Direct follow-up to the prior Soligo et al. finding, operationalising why the general misalignment direction is preferred over a narrow alternative. Introduces three inductive-bias metrics — efficiency (loss per parameter norm), stability (loss-increase rate under orthogonal noise), pre-training significance (KL divergence between chat and steered models on FineWeb) — and shows the general direction scores higher on all three. A narrow representation exists and is learnable but only with KL-divergence regularisation penalising behavioral change outside the dataset domain. The pre-training-significance result is the direct quantitative operationalisation of PSM's pre-existing-direction claim. Pattern replicates on a technical-prose generalisation example.
- Unfaithful chain-of-thought as nudged reasoning (Bogdan, Macar, Conmy, Nanda 2025) — co-author. Reframes unfaithful CoT as a continuous nudging dynamic: hidden information biases token probabilities by small amounts at every sentence; the bias accumulates across the CoT without any individual token being a clear disclosure. Hint-transplantation experiment on DeepSeek R1-Qwen-14B shows ~20-point probability shift across an 8-sentence prefix.
Team context
Leads mechanistic interpretability work at Google DeepMind and supervises MATS (ML Alignment & Theory Scholars) scholars. The Soligo et al. and Arditi et al. papers both list a mix of DeepMind, MATS, and independent affiliations under Nanda's senior supervision, indicating a research-supervision pattern that extends beyond the in-house team. The Soligo team (Soligo, Turner, Rajamanoharan, Nanda) appears on two of the in-wiki findings (June 2025 convergent-misalignment, February 2026 EM-Easy) — a sequel pattern in which the second paper directly extends the methodology and conclusions of the first. Methodologically positioned alongside Anthropic Interpretability as a second institutional locus for the linear-representations program; the two groups' findings cross-cite and mutually corroborate — the convergent-misalignment direction is filed as direct mechanistic support for the Anthropic-originated Persona Selection Model, and EM-Easy operationalises PSM's pre-existing-direction claim quantitatively via the pre-training-significance metric.