Vansh Wahi

National Society of Professional Engineers · United States

Unclaimed · AI-structured profileOpenAlex

Early-career AI-safety researcher on reward hacking and the limits of LLM-as-a-judge evaluation

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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—

Output scoring is unavailable: the OpenAlex record lists h-index 0 and i10-index 0 with only two 2026 works, too thin to compute a meaningful volume signal. This axis would measure total publications and preprint throughput over time.

Impact—

Impact data is insufficient: with h-index 0 and i10-index 0 on the current OpenAlex identity, no citation-based signal exists yet. This axis would measure citations, citations per paper, and uptake of the two reviewed preprints.

Focus0/100

The focus score is 0, mirroring a currently broad early agenda: the field distribution splits evenly (0.5 machine learning/AI safety-alignment, 0.5 AI/LLM evaluation and agent safety), and OpenAlex assigns five distinct topics — topic modeling, type systems, NLP techniques, multi-agent negotiation, and logic and reasoning — one work each. Both cataloged papers do converge on reward hacking and LLM-judge reliability, suggesting a core theme is forming rather than an established specialty.

Collaboration0/100

The collaboration score is 0 because the OpenAlex record lists zero distinct coauthors and no frequent collaborators, while both Evolve-reviewed 2026 papers position this researcher as first and corresponding author. The profile so far reflects independent authorship; co-author networks, team science, and cross-institution ties have yet to appear in the available bibliographic data.

Activity32/100

Activity scores 32, consistent with a newly active record: OpenAlex counts exactly 2 works in 2026 and 2 works over the trailing three-year window, both surfacing as arXiv preprints reviewed by Evolve. There is no earlier publication history in the data, so the score reflects a short, recent burst of output rather than a sustained multi-year publishing cadence.

Reproducibility38/100

Reproducibility scores 38 as a mean over the 2 Evolve-reviewed papers. The reward-hacking framework paper is rated medium readiness — conceptually careful but with no new experiments and self-cited evidence — while the LLM-as-a-judge paper is rated low, based on single-run, partly model-labeled production evidence. The pattern points to argument-led work where formal artifacts and released code would most raise reproducibility.

Research summary

Vansh Wahi is an early-stage, US-listed researcher working at the intersection of AI safety and LLM evaluation, with an OpenAlex record that became active in 2026. The record contains exactly 2 works, both arXiv preprints reviewed by Evolve, on which Wahi is first and corresponding author: 'Optimizing the Score, Losing Sight of the Task' (2026), a framework paper mapping reward hacking across weight updates, best-of-n selection, and prompt optimization; and 'LLM-as-a-Judge Is Not an Oracle' (2026), which argues that self-improving agents need deterministic guardrails and that LLM judges should serve as advisors rather than oracles. Field distribution splits evenly between machine learning (AI safety/alignment) and artificial intelligence (LLM evaluation and agent safety), while OpenAlex topics scatter across five areas — topic modeling, type systems, NLP techniques, multi-agent negotiation, and logic and reasoning — with one work each. Quantitative signals are thin: h-index and i10-index are 0, no citation-based output or impact score is available, and the reproducibility mean over 2 papers is 38, with one paper rated medium and one low readiness. The consistent thread is Goodhart-style failure of proxy rewards and the case for deterministic verification around model-based judgment; the evidence so far is conceptual and production-anecdotal rather than experimental.

Collaboration pattern

Available data shows an independent working pattern: OpenAlex records 0 distinct coauthors, the frequent-coauthor list is empty, and both 2026 papers are first-authored and corresponding-authored. No institutional co-author network can be inferred; the listed affiliation is the National Society of Professional Engineers (US). These collaboration figures are early-record zeros rather than evidence of an isolated agenda — any team-based work would immediately change this axis.

Research keywords

reward hackingreward model overoptimizationbest-of-n selectionprompt optimizationLLM-as-a-judgeself-improving agentsdeterministic verificationGoodhart's lawproxy rewardLLM evaluationagent safety

Highlights

Mapping reward hacking across weights, selection, and prompts
The 2026 framework paper compares how reward hacking arises in weight updates, best-of-n selection, and prompt optimization, with LLM-as-a-judge scoring as a recurring amplifier. Evolve's verdict: conceptually careful but formally thin, with no new experiments and self-cited evidence; reproducibility readiness is rated medium.
Demoting LLM judges behind deterministic guardrails
The second 2026 preprint argues that self-improving agents cannot treat LLM-as-a-judge as an oracle: model-based scores should sit behind deterministic verification gates. Evolve notes vivid production evidence, but single-run, partly model-labeled, and non-reproducible; readiness is rated low.
A solo, first-author start to the record
Both cataloged papers list Wahi as first and corresponding author, and the OpenAlex record shows 0 distinct coauthors with an empty frequent-coauthor list — an independent, argument-led entry into AI safety with a combined reproducibility mean of 38 over 2 papers.

Timeline

2026First Evolve-cataloged preprint appears: 'Optimizing the Score, Losing Sight of the Task', a framework paper on reward hacking across weights, selection, and prompts, authored first and corresponding on arXiv.
2026Second preprint, 'LLM-as-a-Judge Is Not an Oracle', argues self-improving agents need deterministic guardrails, again first and corresponding author on arXiv.
2026OpenAlex year-by-year counts place both works in 2026 (countsByYear 2026: 2; recent3yWorks 2), with h-index and i10-index at 0 — an identity record just beginning to accumulate signal.

Papers in the Evolve catalog

2026-09-22first authorcorresponding
Optimizing the Score, Losing Sight of the Task: Reward Hacking Across Weights, Selection, and Prompts
2026-09-02first authorcorresponding
LLM-as-a-Judge Is Not an Oracle: Why Self-Improving Agents Need Deterministic Guardrails

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.

All bibliometrics come from the OpenAlex record only; no ORCID is linked, so this profile is name-resolved from public bibliographic data and identity merges cannot be ruled out.
Citation-based output and impact are unscored: h-index and i10-index are both 0 and no citation counts exist yet, so these reflect a very new record, not measured impact.
The reviewed corpus is limited to 2 arXiv preprints from 2026; no peer-reviewed venue, dataset release, or code release appears in the available data.
The affiliation shown ('National Society of Professional Engineers', US) is taken from the input record and unverified, and the reproducibility score averages only 2 papers.