Evolve.Pub
Sign inScan your paper
SG

Shivam Gupta

—

Unclaimed · AI-structured profileOpenAlex

Early-career ML researcher in LLM agent evaluation and agent post-training theory (GRPO, partial identification)

2

papers here

2

reviewed by Evolve

2

first-author

1

corresponding

—

works · OpenAlex

—

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-02) 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 dataOutputImpact: insufficient dataImpactFocus: 0/100FocusCollaboration: 14/100CollaborationActivity: insufficient dataActivityReproducibility: 100/100Reproducibility

Output—

No output data is available yet for this profile. This axis would measure publication volume — total papers, yearly counts, and authorship positions across venues — none of which can be computed from a snapshot holding only the two reviewed papers.

Impact—

Impact cannot be scored yet: no citation counts, h-index, or venue-level influence data are present. This axis would measure how often the work is cited and taken up by other researchers, once such bibliometric data accumulates.

Focus0/100

The focus score is 0, the least concentrated end of the radar, and the raw field distribution explains why: the two cataloged papers split evenly (0.5/0.5) between machine learning / AI agent evaluation (statistical identification) and machine learning (reinforcement learning for LLM agent post-training). With only two papers, this even split is as much a small-sample artifact as a genuine scatter, so the axis should be re-read as more papers accumulate.

Collaboration14/100

The collaboration score is 14, and the raw metric behind it is stark: exactly 1 distinct coauthor appears across the cataloged papers, and the frequent-coauthor list resolves only to Shivam Gupta themselves, which reads as a name-disambiguation artifact. Both 2026 papers are first-authored — one also corresponding — consistent with a solo or near-solo working pattern rather than a documented team footprint.

Activity—

Activity is unmeasurable in this snapshot: no per-year publication cadence or recent-output signal is available. The axis would track publishing rhythm over time — whether output is sustained, accelerating, or dormant — once more dated records exist.

Reproducibility100/100

Reproducibility is the strongest axis, scoring 100 as the mean across the Evolve-reviewed papers (reproMean 100 over reproPapers 2). Each paper carries a high reproducibility readiness rating: the delegation-audit paper is a computational audit and the tool-cache paper is exact mathematics, and Evolve's verdicts in both cases explicitly note honest scoping of limitations, which supports the rating.

Research summary

Shivam Gupta is an early-stage machine learning researcher whose Evolve-reviewed work sits at the intersection of LLM agent evaluation and the statistical theory of agent post-training. The catalog contains two 2026 arXiv papers, both first-authored. The first, 'The Delegation Blind Spot,' is a computational audit showing that coarse agent logs leave downstream product decisions statistically unidentified — a partial-identification argument drawing on label shift, measurement error, and decision theory. The second, 'Marginally Correct Tool Caches Can Reverse Group-Normalized Policy Updates,' is an exact theory result: stochastic tool-result caches that preserve per-rollout reward marginals can nonetheless reverse group-normalized (GRPO-style) policy updates, exposing a failure mode in advantage estimation for agent training. Evolve's verdicts characterize both as honest and rigorous but limited in scope: the audit is fully synthetic with largely known theory, and the cache result is a toy model. Both papers carry high reproducibility readiness, and the reproducibility dimension scores 100 across the two reviewed papers. Field identification splits evenly (0.5/0.5) between AI agent evaluation and reinforcement learning for LLM agent post-training. Coauthorship data is thin — a single distinct coauthor is recorded — and output, impact, and activity metrics are not yet available, so this profile should be read as a narrow, early snapshot rather than a career summary.

Collaboration pattern

Coauthorship data for this profile is minimal: only 1 distinct coauthor is recorded across the cataloged papers, and the frequent-coauthor list resolves only to Shivam Gupta themselves — likely a name-disambiguation artifact rather than a real collaborator. Both 2026 papers are first-authored (one also corresponding), consistent with a solo or near-solo working mode. The collaboration dimension therefore scores low (14), and any conclusion about long-term collaboration patterns would be premature.

Research keywords

partial identificationLLM agentslabel shiftdecision theorymeasurement erroraudit methodologyGRPOgroup normalizationtool-result cachingLLM agent trainingpolicy gradient estimatorsadvantage estimation

Highlights

Timeline

  1. 2026First-authored arXiv paper 'The Delegation Blind Spot: Auditing Product Decisions from Agent Choices' appears, framing product-decision audits of LLM agents as a partial-identification problem.
  2. 2026First-authored theory paper 'Marginally Correct Tool Caches Can Reverse Group-Normalized Policy Updates' appears on arXiv, targeting GRPO-style agent post-training.
  3. 2026Both 2026 papers enter the Evolve review catalog with high reproducibility readiness, yielding a reproducibility mean of 100 across the 2 reviewed papers.

Papers in the Evolve catalog

Frequent co-authors

Shivam Gupta×1

Is this you? Claim this profile

Enter the email you published with. We'll send a one-time claim link; after claiming you can edit the introduction or hide this page. Requests from a mismatched address go to manual review.

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.

  • Dimension scores are computed from public bibliometric data only; no institutional records, citation counts, or venue-level metrics beyond this snapshot are available.
  • 'Shivam Gupta' is a common name; this profile is name-resolved from public bibliographic data, and institution, country, and ORCID remain unverified.
  • Only the two Evolve-reviewed 2026 arXiv papers inform this profile; it is not a complete publication record.
  • The frequent-coauthor list containing the researcher's own name suggests a disambiguation artifact in the coauthorship data.
Shivam Gupta — Researcher Profile | Evolve.Pub