Simran Koul

—

Unclaimed · AI-structured profileOpenAlex

Early-career NLP researcher exposing register and dialect bias in complexity-based LLM routing and model cascades

papers here

reviewed by Evolve

first-author

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—

Not yet scored: the snapshot contains a single 2026 arXiv preprint and no publication counts or venue history. This axis would measure total research output — papers, venues, and volume over time — and cannot be estimated from one paper.

Impact—

Cannot be scored: no citation counts, h-index, or venue-level influence data are available in the snapshot. This axis would measure the citation footprint and scholarly influence of the researcher's work across the literature.

Focus100/100

Focus scores 100 because every reviewed paper sits in one field: natural language processing with an algorithmic fairness lens. The underlying field distribution contains a single entry (NLP, algorithmic fairness), so the specialization measure is maximal. That said, the base is one paper, so the score reflects an undiluted but very early agenda rather than a long-track record.

Collaboration0/100

Collaboration scores 0 because the cataloged record shows zero distinct coauthors: the single reviewed paper is solo-authored, with first and corresponding authorship. No frequent coauthors are on file. A score of 0 here indicates no co-authored work in the current snapshot, not a negative judgment, and it should rise if collaborative papers appear.

Activity—

Not yet scored: the snapshot spans a single year (2026) with one arXiv preprint, so there is no multi-year publication cadence to measure. This axis would track papers per year and submission continuity, and needs a longer record.

Reproducibility55/100

Reproducibility scores 55, the mean readiness across one assessed paper — the register-bias routing audit, rated medium. Evolve's review highlights strong paired evidence as a reproducibility asset, while noting the study tests no deployed router, which limits external replication scope. With only one paper assessed, this score is indicative rather than settled.

Research summary

Simran Koul is a researcher in natural language processing with an algorithmic fairness focus, profiled here from a public bibliographic snapshot. The Evolve catalog currently contains one paper, 'Register Bias in Complexity-Based Large Language Model Routing' (2026, arXiv), on which Koul is first and corresponding author. The paper audits complexity-based routing — systems that estimate query complexity to decide which model in a cascade should answer — and shows that length-based routing downgrades queries written in non-standard registers, including African American English. Evolve's review calls it a 'novel, honest audit' with strong paired evidence, while noting that no real deployed router was tested. The radar reflects this early, concentrated record: focus scores 100 because all reviewed work sits in a single field, reproducibility scores 55 (the mean across one paper rated medium readiness), and collaboration scores 0 because the cataloged paper lists no coauthors. Output, impact, and activity are not yet scored, since the snapshot provides no citation counts or multi-year publication data. The profile should be read as an early-stage record of a solo, fairness-driven NLP agenda.

Collaboration pattern

Every paper currently in the catalog is solo-authored: the snapshot records zero distinct coauthors and no frequent collaborators, and the collaboration dimension scores 0. On the reviewed paper Koul holds both first and corresponding authorship, controlling the full research pipeline. This rests on a single-paper base, so the collaboration pattern may change quickly as more work is indexed.

Research keywords

LLM routingalgorithmic fairnessregister biasAfrican American Englishdialect biasquery complexity estimationmodel cascadesfairness auditinglarge language modelslength-based routing

Highlights

Audit exposes register bias in complexity-based LLM routing
The 2026 first-author arXiv paper shows that length-based routing downgrades queries written in non-standard registers, including African American English; Evolve's verdict calls it a novel, honest audit with strong paired evidence.
Solo first-and-corresponding authorship, medium reproducibility readiness
The cataloged work is single-authored (zero distinct coauthors) and carries Evolve's medium repro-readiness rating (mean 55); the review notes no real deployed router was tested, leaving deployment-level validation open.
Dialect fairness framed as a routing-layer problem
The work links dialect bias and African American English to query complexity estimation and model cascades, indicating that fairness risks arise in the routing layer rather than only in individual models.

Timeline

2026First- and corresponding-author arXiv preprint 'Register Bias in Complexity-Based Large Language Model Routing' released.
2026Paper enters the Evolve catalog, classified in natural language processing (algorithmic fairness).
2026Evolve review records a 'novel, honest audit' verdict and a medium reproducibility-readiness rating (mean 55, one paper).

Papers in the Evolve catalog

2026-09-17first authorcorresponding
Register Bias in Complexity-Based Large Language Model Routing

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 dimensions are computed from a very small bibliometric base — one reviewed paper — so scores can shift substantially as more records are indexed.
No institution, country, or ORCID is available; the profile is name-resolved from public bibliographic data and may be ambiguous.
Output, impact, and activity dimensions lack underlying data — no citation counts, publication counts, or multi-year records — so they are left unscored.
Bibliometric evidence here is limited to the Evolve-reviewed catalog; the reproducibility score rests on a single paper (mean 55).