Shubham Santosh Pandere
Visvesvaraya National Institute of Technology
Mechanistic interpretability researcher: knowledge-conflict circuits in LLMs and 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—
This axis is unscored: the snapshot contains no publication-volume metrics. It would measure output over time — paper counts, venues and authorship roles — of which only two reviewed 2026 preprints are currently on record.
Impact—
Impact cannot be scored here: the input carries no citation counts, h-index or uptake signals. This axis would track how often the two preprints are cited and built upon; until such coverage exists, no estimate is made.
Focus0/100
The focus score is 0, driven by the raw field distribution: a 50/50 split between mechanistic interpretability of LLMs and mechanistic interpretability of vision transformers. Read qualitatively, the split is thematic rather than scattered — both reviewed 2026 papers work on circuit-level explanations, one probing knowledge-conflict circuits in language models, the other testing topographic training on vision transformers.
Collaboration44/100
The collaboration score is 44, drawn from 8 distinct coauthors across the two reviewed 2026 preprints — a small-team pattern rather than large-consortium authorship. The recurring circle includes Gautam Ranka, Ritika Varshney, Navya Deshmukh, Roushni Sareen, Roshan Kumar Singh and Aiden Dsouza. The role mix matters: first author on the knowledge-conflict study, middle author on the topographic-training paper.
Activity—
Activity is unscored because the snapshot lacks multi-year timing data. It would measure research cadence — preprints per year and gaps between outputs. Both reviewed papers are dated 2026, which alone cannot establish a productivity trajectory.
Reproducibility78/100
Reproducibility is the strongest scored axis at 78, the mean readiness across 2 reviewed papers. The first-author instruction-tuning study rates high — Evolve calls it a rigorous multi-method circuit case across three LLM families, despite 3-4B-only models and a missing chat template — while the topographic-training study rates medium, with its H2 confound unresolved.
Research summary
Shubham Santosh Pandere is a mechanistic-interpretability researcher affiliated with Visvesvaraya National Institute of Technology, profiled here from public bibliographic data. The reviewed record is early-stage but unusually coherent: two 2026 arXiv preprints, one on each half of the recorded field distribution — mechanistic interpretability of LLMs (50%) and mechanistic interpretability of vision transformers (50%). The flagship output is a first-author study, "Rewired or Gated?", arguing that instruction tuning gates rather than rewires knowledge-conflict circuits across three LLM families; Evolve's review calls it a rigorous multi-method case while noting limits — only 3-4B models and no chat template. A second, co-authored paper shows topographic training (TopoLoss) concentrates causal circuits in vision transformers (2.79x vs 1.68x) without improving neuron monosemanticity, rated a mostly convincing workshop study with an unresolved H2 confound. Reproducibility is the standout scored dimension at 78, the mean readiness across 2 papers (one high, one medium). Collaboration scores 44, built on 8 distinct coauthors with recurring partners including Gautam Ranka and Ritika Varshney. The focus score of 0 reflects the even two-way field split rather than scattered interests, since both papers sit inside mechanistic interpretability. Output, impact and activity axes lack sufficient data and are left unscored; no citation counts, country or ORCID are available in this snapshot.
Collaboration pattern
Collaboration runs on a small-team model: the radar scores it 44 from 8 distinct coauthors across the two reviewed 2026 preprints. Recurring partners include Gautam Ranka, Ritika Varshney, Navya Deshmukh, Roushni Sareen, Roshan Kumar Singh and Aiden Dsouza. The role mix is instructive — first author on the LLM circuit study, middle author on the vision-transformer study — suggesting leadership within a compact circle alongside substantive contribution to peers' projects.
Research keywords
Highlights
Instruction tuning gates, rather than rewires, knowledge-conflict circuits
TopoLoss concentrates causal circuits without monosemanticity gains
Reproducibility-first interpretability work
The reproducibility axis scores 78, the mean readiness across the 2 reviewed papers: high on the LLM knowledge-conflict study, medium on the vision-transformer topographic-training study — a strong signal for a researcher whose methods rely on circuit-level causal interventions.
Timeline
- 2026First-author arXiv preprint "Rewired or Gated?" — instruction tuning gates rather than rewires knowledge-conflict circuits across three LLM families, rated high for reproducibility readiness.
- 2026Co-authored arXiv preprint showing topographic training concentrates causal circuits in vision transformers (2.79x vs 1.68x) without improving neuron monosemanticity, rated medium for reproducibility.
- 2026Two preprints land in the same year, one on each side of the recorded 50/50 field split — mechanistic interpretability of LLMs and of vision transformers.
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.
- Radar scores are computed from a real bibliometric snapshot; output, impact and activity lack sufficient data and are not scored, and no citation counts or h-index exist in the input.
- This profile is name-resolved from public bibliographic data; no ORCID or country is recorded, so identity should be confirmed before reuse.
- Reproducibility and highlight assessments rest on only two reviewed 2026 arXiv preprints; no peer-reviewed venues or full publication history are in scope.
- The frequent-coauthor list contains duplicate and self entries; counts cited here use the de-duplicated raw figure of 8 distinct coauthors.