Mechanistic-interpretability researcher: knowledge-conflict circuits in LLMs, causal circuits in vision transformers.
2
papers here
2
reviewed by Evolve
1
first-author
0
corresponding
—
works · OpenAlex
—
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 publication-volume data is available yet. This axis would measure cumulative and per-year paper counts, authorship positions, and venue spread. With only two reviewed preprints on file, productivity cannot be scored honestly, so the axis is left unscored rather than estimated.
Impact—
No citation-based data is available yet. This axis would measure citations, h-index-style aggregates, and influence on follow-up interpretability work. Because both reviewed papers are 2026 preprints, meaningful citation evidence has not accumulated, so no impact score is assigned.
Focus0/100
The focus axis scores 0 because the indexed field distribution splits evenly: 0.5 for 'machine learning — mechanistic interpretability of llms' and 0.5 for 'machine learning — mechanistic interpretability of vision transformers'. In practice both halves sit inside one narrow specialty — circuit-level mechanistic interpretability — so the zero reflects label-level dispersion across two catalog fields rather than genuine topic switching, and topical concentration is likely understated by this score.
Collaboration44/100
Collaboration scores 44. The raw data lists 8 distinct coauthors across the two reviewed 2026 papers, with Shubham Santosh Pandere the most recurring partner, alongside Ritika Varshney, Navya Deshmukh, Roushni Sareen, Roshan Kumar Singh, and Aiden Dsouza. Team sizes are small and roles are mixed — first author on one paper, middle author on the other — with no corresponding authorship on record, consistent with a mid-range collaboration score.
Activity—
No recent-output cadence data is available yet. This axis would measure publication frequency and recency over trailing one- to five-year windows. Since both reviewed papers carry the same 2026 date, year-over-year activity cannot be computed and the axis is left unscored.
Reproducibility78/100
Reproducibility scores 78, computed as the mean over 2 papers (reproPapers = 2). The knowledge-conflict study rates high readiness — Evolve calls it a rigorous multi-method case — while the topographic-training study rates medium, with the H2 confound unresolved. Both rely on activation patching and circuit analysis, methods with established tooling, supporting above-average verifiability for preprint-stage interpretability work.
Research summary
Gautam Ranka is a mechanistic interpretability researcher affiliated with Visvesvaraya National Institute of Technology whose reviewed work concentrates on how transformer models organize computation into circuits. The Evolve catalog holds two 2026 arXiv papers, splitting the indexed field distribution evenly (0.5 each) between mechanistic interpretability of LLMs and of vision transformers. As first author, Ranka showed that topographic training (TopoLoss) concentrates causal circuits (2.79x vs 1.68x) without improving neuron monosemanticity — a workshop-scale study Evolve judged mostly convincing, with one confound unresolved. As a middle author, Ranka contributed to a multi-method analysis of knowledge-conflict circuits across three LLM families, arguing that instruction tuning gates rather than rewires these circuits; Evolve flagged the rigor but noted it covers only 3-4B models and omits a chat template. Both papers score well on reproducibility (mean 78 across 2 papers; high and medium readiness), reflecting activation-patching and circuit-analysis methods with established tooling. With 8 distinct coauthors and a collaboration score of 44, Ranka works in small teams and has no corresponding authorship on record yet. Output, impact, and activity metrics are not available, making this an early-stage profile built from a narrow bibliographic slice.
Collaboration pattern
Across the two reviewed papers, Ranka works in small teams: 8 distinct coauthors give a collaboration score of 44. Shubham Santosh Pandere recurs most often, alongside Ritika Varshney, Navya Deshmukh, Roushni Sareen, Roshan Kumar Singh, and Aiden Dsouza. Authorship roles are mixed — first author on the topographic-training study, middle author on the knowledge-conflict paper — with no corresponding authorship yet recorded, a pattern typical of a junior researcher inside one close-knit interpretability group.
Research keywords
Highlights
TopoLoss concentrates causal circuits without monosemanticity gains
Instruction tuning gates knowledge-conflict circuits rather than rewiring them
A shared activation-patching toolkit across LLMs and vision transformers
Both reviewed papers apply circuit analysis and activation patching, earning a reproducibility mean of 78 across 2 papers (one high, one medium readiness) — an unusually verifiable methodological signature for preprint-stage interpretability work.
Timeline
- 2026Published first-author arXiv study showing topographic training (TopoLoss) concentrates causal circuits (2.79x vs 1.68x) without improving neuron monosemanticity; Evolve rates medium reproducibility readiness.
- 2026Co-authored (middle position) arXiv study concluding that instruction tuning gates rather than rewires knowledge-conflict circuits in three LLM families; Evolve rates high reproducibility readiness.
- 2026Both reviewed outputs land in mechanistic interpretability — split evenly (0.5/0.5) between LLMs and vision transformers — with 8 distinct coauthors recorded in the same year.
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 scores are computed from a small snapshot; output, impact, and activity axes have no underlying data and are intentionally unscored.
- No ORCID or country was provided; the profile is name-resolved from public bibliographic data, and this name is common enough that identity confidence is moderate.
- Only two 2026 arXiv preprints are in the reviewed catalog; peer-reviewed venues, citations, and h-index are not represented.
- The reproducibility score (78) averages just 2 papers, so treat it as provisional.