Imperial College London researcher working on inference-time LLM alignment and watermarking for efficient decoding.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) 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 data: the current snapshot provides no publication-volume, venue, or year-by-year output figures, so this axis is unscored. It would measure the size, venue spread, and cadence of the researcher's publication record.
Impact—
Insufficient data: no citation counts, h-index, or downstream-adoption figures are available in this snapshot, so the axis is unscored. It would measure scholarly influence, including uptake of the test-time alignment and watermarking methods.
Focus0/100
Scored 0 on focus: the field distribution is an even 0.5/0.5 split between two machine-learning directions — LLM alignment with inference-time methods, and LLM efficient inference with watermarking. The two cataloged 2026 papers occupy one thread each: exponential-noise best-of-n test-time alignment on the alignment side, and Poisson-process watermarkable speculative sampling on the efficient-inference side. The portfolio is therefore balanced across two adjacent subfields rather than concentrated in a single direction.
Collaboration36/100
Scored 36: the snapshot records 5 distinct coauthors, with Yanxiao Liu and Sicheng Wan each recurring across the cataloged work, alongside Zhan Gao. On both 2026 arXiv papers Gündüz holds the last-author position without being corresponding author, consistent with a senior supervisory role within a small, stable core team rather than a broad collaboration network.
Activity—
Insufficient data: year-by-year publication counts are absent, so recent research tempo cannot be quantified and the axis is unscored. It would measure publishing frequency and ongoing engagement in the most recent years.
Reproducibility55/100
Scored 55, the mean reproducibility-readiness across the 2 reviewed papers, each rated medium. ExpBoN is backed by an exact finite-n tilted decomposition and geometric convergence, which aids faithful reimplementation, while the Poisson-process watermarking paper is theoretically rigorous but carries minor table inconsistencies that a reproducer would need to reconcile against the text.
Research summary
Deniz Gündüz is a researcher at Imperial College London whose recent output, as indexed in the Evolve catalog, centers on machine learning for large language models — specifically inference-time alignment methods and efficient-decoding techniques coupled with watermarking. The catalog records two 2026 arXiv papers on which Gündüz appears as last author (though not corresponding author). The first, ExpBoN, introduces an exponential-noise soft best-of-n scheme for test-time LLM alignment, with an exact finite-n tilted decomposition and geometric convergence guarantees; its ExpGSI variant reportedly saves compute relative to GSI at comparable accuracy. The second applies Poisson-process machinery to multi-draft speculative sampling, embedding an unbiased watermark without sacrificing speculative-acceptance rates, with rigorous theory marred only by minor table inconsistencies. Evolve's bibliometric radar splits the profile evenly (0.5/0.5) across LLM alignment/inference-time methods and LLM efficient inference/watermarking, records 5 distinct coauthors (collaboration score 36), and rates reproducibility at a mean of 55 across the 2 papers, both classed as medium reproducibility readiness. Output, impact, and activity axes lack sufficient underlying data and are therefore not scored.
Collaboration pattern
The Evolve snapshot records 5 distinct coauthors, with Yanxiao Liu and Sicheng Wan each recurring across the cataloged work, joined by Zhan Gao. On both 2026 arXiv papers Gündüz sits in the last-author position without corresponding-author designation, a pattern consistent with senior supervision of a small, stable core team rather than membership in a large rotating collaboration network; the radar's collaboration score of 36 reflects this compact co-author footprint.
Research keywords
Highlights
ExpBoN: exponential-noise soft best-of-n for test-time alignment
ExpGSI variant saves GSI compute at comparable accuracy
Unbiased watermark embedded in multi-draft speculative sampling
Timeline
- 2026Releases ExpBoN on arXiv, proposing exponential-noise soft best-of-n sampling for efficient test-time LLM alignment (last author).
- 2026Publishes watermarkable multi-draft speculative sampling via Poisson processes on arXiv, embedding an unbiased watermark without acceptance-rate loss (last author).
- 2026Both papers enter the Evolve review catalog with medium reproducibility readiness, contributing to a mean reproducibility score of 55.
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.
- Bibliometric dimensions come from Evolve's limited snapshot; output, impact, and activity axes have insufficient data and are unscored.
- The catalog covers only 2 papers, both 2026 arXiv preprints; no peer-reviewed venue or citation information is included.
- Profile is name-resolved from public bibliographic data via ORCID 0000-0002-7725-395X and the Imperial College London affiliation; verify identity when citing.
- The reproducibility score (mean 55) is based on only the 2 reviewed papers and Evolve's own readiness ratings.