Mechanistic-interpretability researcher at Google probing when SAE features are causally steerable, not merely decodable
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
3
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-13) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
No output score is computed for this profile. This axis would summarize total publication volume, venue types, and first/last-author share across the career; the available record only lists per-year work counts, which is insufficient.
Impact—
No impact score is available. This axis would measure citation counts, h-index trajectory, and downstream influence on later work; the record contains only a minimal h-index of 3 and i10-index of 3, too little to score reliably.
Focus0/100
Scores 0 on focus. The field distribution splits evenly (0.5/0.5) between 'machine learning — mechanistic interpretability' and 'mechanistic interpretability of large language models' — two adjacent topics — yet the record's five OpenAlex topics (topic modeling, adversarial robustness, multimodal applications, XAI, science education) spread thin and may reflect name-ambiguity noise rather than genuine breadth. The reviewed work itself is tightly concentrated on SAEs.
Collaboration48/100
Scores 48 on collaboration, with 10 distinct coauthors on record. Both reviewed 2026 papers are small-team arXiv efforts with Shah in middle and last authorship; recurring collaborator names include Muhammad Khan, Shlok Channawar, Akshaj Gurugubelli, and Maheep Chaudhary. This pattern suggests compact, hands-on interpretability working groups rather than large multi-institution consortia.
Activity52/100
Scores 52 on activity, driven by recency rather than volume: 5 works in the last three years, with the counts-by-year record peaking at 4 works in 2026 — including both Evolve-reviewed SAE papers. Earlier single-work entries (2006, 2012, 2023, 2025) suggest an early-career record, with the oldest entries likely homonym artifacts.
Reproducibility55/100
Scores 55 on reproducibility, a reproMean computed over 2 papers, both rated medium reproducibility by Evolve. Both are arXiv preprints describing controlled intervention experiments — rank-correlation evidence up to ρ = −0.546, and single-model Gemma-2-9B settings — which are checkable in principle but depend on model access and careful SAE training.
Research summary
Aditya Shah is a researcher at Google (United States) whose profile, name-resolved from public OpenAlex bibliographic data, centers on mechanistic interpretability of large language models — specifically sparse autoencoder (SAE)-based analysis of when features are causally actionable rather than merely readable. Two 2026 arXiv papers in the Evolve catalog define the record. 'Decodability is Not Causality' (middle author) shows that linear-probe weight geometry does not identify the Gemma-2-9B features that causally control a True/False truth probe's behavior, using SAE decomposition and feature ablation. 'Look Before You Steer' (last author) proposes decoder-geometry metrics that rank SAE feature steerability before intervention, with rank-correlation evidence up to ρ = −0.546. Evolve rates both as methodologically careful but setting-limited — single-model/layer/task and rank-only evidence — with medium reproducibility (reproMean 55 over 2 papers). Bibliometrics are early-stage: h-index 3, i10-index 3, 10 distinct coauthors, and 5 works in the last three years, peaking at 4 in 2026. Collaboration and activity scores (48 and 52) fit a compact preprint-team pattern; a focus score of 0 reflects topic spread in the OpenAlex record — topic modeling, adversarial robustness, multimodal applications, XAI, and science education — that may partly be name-ambiguity noise, since the reviewed work itself is tightly SAE-focused. Output and impact dimensions lack sufficient data to score. Overall, this reads as a concentrated early research program on making SAE features predictable, steerable, and honestly causal.
Collaboration pattern
Scores 48 on collaboration with 10 distinct coauthors on record. Both reviewed 2026 papers are small-team arXiv efforts with Shah in middle and last authorship; recurring collaborator names include Muhammad Khan, Shlok Channawar, Akshaj Gurugubelli, Girish Gupta, Devesh Tiwari, Camille Davis, Shivank Sinha, Talia Weaver, and Maheep Chaudhary. The pattern suggests compact, hands-on interpretability working groups rather than large multi-institution consortia.
Research keywords
Highlights
Probe decodability is not causality
In a workshop-scale but methodologically careful SAE study on Gemma-2-9B, linear-probe weight geometry fails to identify the features that causally control a True/False truth probe's behavior, dissociating probe readouts from behavioral drivers via SAE decomposition and feature ablation; convincing within its single model/layer/task setting.
Decoder geometry predicts SAE steerability
As last author, proposes decoder-geometry metrics that rank SAE feature steering cost before intervention. Evolve judges the evidence modest but well-controlled — rank correlation reaching ρ = −0.546, though rank-only — making pre-intervention steerability screening credible with caveats.
A coherent SAE-causality research program
Taken together, both 2026 papers press the same question: does representational geometry predict intervention outcomes? One negates it for probes, one affirms it for steering cost — a focused early agenda around making sparse-autoencoder features predictable and honestly causal.
Timeline
- 2026Publishes 'Decodability is Not Causality' on arXiv (middle author), showing via SAE decomposition on Gemma-2-9B that probe weight geometry does not identify causally driving features.
- 2026Publishes 'Look Before You Steer: Geometry Predicts SAE Feature Steerability' on arXiv as last author, with rank-correlation evidence up to ρ = −0.546 linking decoder geometry to steering cost.
- 2026Most productive year on record: the OpenAlex counts-by-year data peaks at 4 works in 2026, within 5 works over the last three years — including both Evolve-reviewed papers.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Name-resolved from public OpenAlex bibliographic data; 'Aditya Shah' is a common name, so the 2006 and 2012 entries in the counts-by-year record may be homonym artifacts.
- Bibliometrics come from the provided OpenAlex snapshot only; output and impact dimensions are unscored due to insufficient underlying data.
- Qualitative assessments draw on only two Evolve-reviewed papers, both 2026 arXiv preprints; peer-reviewed venue status is not recorded.
- Radar scores (collaboration 48, activity 52, focus 0, reproducibility 55) are computed from real bibliometric data but rest on a very small publication record.