Mechanistic interpretability researcher separating probe readouts from causally driving features in LLM truthfulness
1
papers here
1
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-20) 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—
Insufficient data: no publication-volume metrics are available yet. This axis would measure the size and shape of the publication record — paper count, venues, and authorship roles — none of which can be estimated from a single reviewed paper.
Impact—
Insufficient data: citation and influence signals are not available for this profile. The impact axis would measure how the work propagates — citations, uptake in follow-up interpretability studies, and community reference — and cannot be scored from the current catalog.
Focus100/100
This is an exceptionally concentrated profile on the available evidence: the field distribution assigns 100% of the reviewed output to a single field, mechanistic interpretability of large language models. The cataloged first-author paper applies that lens to linear probes, sparse autoencoder decomposition, and causal interventions on Gemma-2-9B. With only one reviewed paper, the perfect focus score reflects a narrow but coherent specialization rather than a long-term record.
Collaboration36/100
The collaboration score of 36 rests on a small pool: five distinct coauthors appear in the data — Camille Davis, Shivank Sinha, Talia Weaver, Aditya Shah, and Maheep Chaudhary. The cataloged paper is first-authored but not corresponding, consistent with contributing within a team rather than leading a large network. A mid-low score here signals a compact working group; it is not a quality judgment.
Activity—
Insufficient data: no longitudinal publication timeline can be built yet. The activity axis would measure ongoing research tempo — recent output, year-over-year continuity, and gaps — but the catalog currently anchors this researcher to a single 2026 arXiv paper.
Reproducibility55/100
Reproducibility averages 55 across the one paper assessed so far, whose readiness is rated medium. That single rating fully determines the score, so it should be read as a provisional signal rather than a stable habit: the SAE-intervention study is methodologically careful, yet the evidence base is one workshop-scale arXiv paper whose generalization beyond a single model, layer, and task remains untested. Additional reviewed papers would quickly firm this axis up.
Research summary
Devesh Tiwari is a researcher working in mechanistic interpretability of large language models, the single field to which the bibliometric field distribution in this profile is entirely assigned. The Evolve catalog currently holds one reviewed paper, a 2026 arXiv preprint on which Tiwari is first author: "Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition." The study uses sparse autoencoder decomposition and causal interventions to show that the weight geometry of a linear probe does not identify the features in Gemma-2-9B that causally control a True/False truth probe's behavior. Evolve's assessment describes it as a workshop-scale but methodologically careful piece of work — convincing within its single-model, single-layer, single-task setting, with generalization explicitly untested — and rates its reproducibility readiness as medium. The profile radar reflects this narrow evidence base: focus scores 100 because the reviewed output sits entirely in one field; collaboration scores 36 on a pool of five distinct coauthors (Camille Davis, Shivank Sinha, Talia Weaver, Aditya Shah, Maheep Chaudhary); reproducibility averages 55 over one assessed paper. Output, impact, and activity dimensions lack underlying data and are reported as missing rather than estimated. "Devesh Tiwari" is treated as name-resolved from public bibliographic data; institution, country, and ORCID are not on record.
Collaboration pattern
The visible collaboration footprint is compact: five distinct coauthors — Camille Davis, Shivank Sinha, Talia Weaver, Aditya Shah, and Maheep Chaudhary — recur in the data, and the collaboration dimension scores 36. On the cataloged 2026 paper Tiwari is first author but not corresponding, suggesting a lead-contributor role inside a small, stable team rather than direction of a large, distributed network. No institutional affiliation is on record to contextualize these ties.
Research keywords
Highlights
Probe geometry is not causality
SAE-based dissociation method
Scope honestly bounded
Timeline
- 2026First-authored paper "Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition" appears on arXiv.
- 2026The study establishes that linear-probe weight geometry does not pinpoint the Gemma-2-9B features causally controlling a truth probe's behavior.
- 2026Evolve reviews the paper as methodologically careful but workshop-scale, rates its reproducibility readiness as medium, and files it under mechanistic interpretability of large language models.
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.
- Bibliometric figures and radar scores derive from OpenAlex-based public bibliometric data; no institution, country, or ORCID is on record.
- All reviewed evidence comes from a single cataloged paper (2026 arXiv preprint), so every dimensional narrative rests on a one-paper base.
- Output, impact, and activity axes lack underlying data and are reported as missing, not estimated.
- "Devesh Tiwari" is name-resolved from public bibliographic data; links to other publications or people sharing the name are unverified.