Mechanistic-interpretability researcher probing how language models commit to facts during recall
1
papers here
1
reviewed by Evolve
1
first-author
1
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-21) 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—
The output axis is not scored yet: this snapshot contains no paper-count or venue-volume metrics for this researcher. It would measure total and per-year publication volume across venues, and should be computed once more bibliographic records are linked to this profile.
Impact—
No impact score is available because the snapshot lacks citation and usage metrics. This axis would measure citation counts, citation velocity, and uptake of the reviewed paper, and it will be scored only after such data has been collected and verified.
Focus100/100
The focus dimension scores 100 out of 100 because every paper in the current snapshot — one of one — falls in a single field, machine learning (mechanistic interpretability of language models). The underlying field distribution lists exactly one field with one paper, so measured concentration is total. With such a small base, this perfect score reflects current depth in activation patching and factual recall rather than an established long-term publication pattern.
Collaboration0/100
The collaboration dimension scores 0 because the snapshot records zero distinct co-authors (distinctCoauthors = 0). Consistently, the single reviewed paper lists this researcher as both first and corresponding author with no co-authors named in the catalog, indicating fully solo-authored work so far. As more papers enter the catalog, repeatedly appearing co-authors will populate the frequent-collaborator list and move this score away from its floor.
Activity—
The activity axis carries no score: the snapshot provides no multi-year publication cadence data. It would measure publishing frequency and recency over time, and it requires more dated records than the single 2026 paper currently reviewed.
Reproducibility100/100
Reproducibility scores a perfect 100, the mean across the one paper Evolve has assessed so far (reproPapers = 1, reproMean = 100). Evolve's verdict calls the study a well-controlled activation-patching investigation across four language models, and its reproducibility readiness is rated high. The same review notes the absence of inferential statistics — the one flagged methodological gap in an otherwise high-readiness result, and a natural target for revision.
Research summary
Divyansh Agarwal is a researcher in machine learning, specializing in the mechanistic interpretability of language models. This is an early-stage, name-resolved profile built from a bibliometric snapshot linked to ORCID 0000-0003-0404-6877 and Evolve's review of one catalog paper. That paper, "Relation Before Entity: Deferred Commitment in Language Model Factual Recall" (arXiv, 2026), is a first-authored and corresponding-author study claiming that relation signals become generation-controlling before entity signals in four language models. Evolve's verdict characterizes it as a well-controlled activation-patching study while noting the absence of inferential statistics; its reproducibility readiness is rated high, yielding a perfect reproducibility mean of 100 across the one reviewed paper. The work sits entirely within a single field — machine learning (mechanistic interpretability of language models) — producing a perfect focus score of 100, and the snapshot records zero distinct co-authors, indicating solo-authored work so far. The output, impact, and activity axes are missing from the radar because the underlying bibliometric evidence (paper counts, citation metrics, and multi-year publication cadence) is not yet available, and no numbers are quoted for them here. As further papers and metadata accumulate, this profile should move beyond its single-paper evidence base toward a fuller picture of productivity, influence, and collaborative network.
Collaboration pattern
The available evidence shows a solo working pattern: the snapshot counts zero distinct co-authors, the frequent-collaborator list is empty, and the single catalog paper lists this researcher as both first and corresponding author. No repeated collaborators can be identified from a one-paper record. Interpreting collaboration requires care here, since the collaboration score of 0 reflects an empty evidence base as much as any confirmed preference for independent work.
Research keywords
Highlights
First and corresponding author on a 2026 interpretability study
Well-controlled patching evidence, with one flagged gap
High reproducibility readiness rating
Timeline
- 2026Publishes a first-authored and corresponding-author arXiv paper on deferred commitment in language model factual recall.
- 2026The paper enters the Evolve catalog, reviewed as a well-controlled activation-patching study spanning four language models.
- 2026The paper earns a high reproducibility readiness rating, anchoring a perfect reproducibility mean of 100.
Papers in the Evolve catalog
- first authorcorresponding
Relation Before Entity: Deferred Commitment in Language Model Factual Recall
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 OpenAlex-based bibliometric data only; no other indexing source contributes to the radar.
- The profile rests on a single reviewed paper, so output, impact, and activity are unscored and every present score has a one-paper base.
- Institution and country are not available in the input data and are left unreported.
- The name is not guaranteed unique; this profile is name-resolved from public bibliographic data and ORCID 0000-0003-0404-6877.