University of Washington researcher in efficient LLM inference: test-time alignment, speculative sampling, watermarking
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—
No output data is available yet: the snapshot provides no publication counts, venue history, or yearly volume. This axis would measure the quantity, consistency, and venues of the researcher's output across a full publication record.
Impact—
Impact cannot be assessed: no citation counts, h-index, or downstream uptake are included in the snapshot. This axis would measure how widely the papers are cited and adopted once such bibliometric data becomes available.
Focus0/100
The radar scores focus at 0 because the field distribution splits evenly (0.5/0.5) across two labels — machine learning for LLM alignment and inference-time methods, and machine learning for LLM efficient inference and watermarking. In practice the two cataloged 2026 papers are thematically adjacent, both applying probabilistic sampling tools to efficient LLM inference, so this reads more like a labeling split than genuine topic dispersion, though only two papers back the estimate.
Collaboration36/100
Collaboration scores 36, computed from five distinct coauthors. Wan sits in the middle author positions, never as corresponding author, on both cataloged papers, indicating contributor roles inside small teams. Recurring collaborators include Yanxiao Liu, Deniz Gündüz, and Zhan Gao, though the source list duplicates Yanxiao Liu and contains Wan's own name. The small network and non-lead positions are consistent with an early-stage researcher embedded in established groups.
Activity—
Activity data is insufficient: the snapshot lists two 2026 preprints but no multi-year submission cadence. This axis would measure recent research tempo, such as papers per year and the recency of contributions.
Reproducibility55/100
Reproducibility scores 55, the mean over two papers with Evolve reproducibility ratings. Both cataloged papers are rated medium readiness: the ExpBoN paper pairs an exact finite-n tilted decomposition and geometric convergence with a compute-saving ExpGSI variant, while the Poisson-process watermarking paper is judged theoretically rigorous but flagged for minor table inconsistencies. The strong theoretical framing supports replication, though no code or artifact availability is recorded in the snapshot.
Research summary
Sicheng Wan is a machine-learning researcher at the University of Washington whose Evolve-cataloged work concentrates on sampling-based methods for efficient large language model (LLM) inference. The catalog lists two 2026 arXiv preprints: ExpBoN, which reformulates test-time LLM alignment through an exponential-noise soft best-of-n sampler with an exact finite-n tilted decomposition and geometric convergence guarantees, plus an ExpGSI variant that saves GSI compute at comparable accuracy; and a watermarkable multi-draft speculative sampling scheme built on Poisson processes that embeds an unbiased watermark without sacrificing acceptance rate. Together they span the profile's two evenly weighted field labels — LLM alignment and inference-time methods (0.5) and efficient inference with watermarking (0.5). Wan appears as a middle author on both papers, within a small circle of five distinct coauthors, including Yanxiao Liu, Deniz Gündüz, and Zhan Gao. Evolve rates reproducibility readiness as medium (mean 55 across the two papers), with theoretical rigor highlighted and minor table inconsistencies flagged on the watermarking paper. Output, impact, and activity metrics are not yet available, so this profile should be read as an early-stage, preprint-era snapshot.
Collaboration pattern
The radar scores collaboration at 36, based on a small pool of five distinct coauthors. Wan appears as a middle, non-corresponding author on both cataloged papers, suggesting a contributor role within compact research teams rather than leading an independent group. The frequent-coauthor list features Yanxiao Liu (listed twice in the source data), Deniz Gündüz, and Zhan Gao; it also includes Wan's own name, which is retained as reported. No cross-institution affiliation details are provided.
Research keywords
Highlights
ExpBoN: exponential-noise soft best-of-n for cheaper test-time LLM alignment
Unbiased LLM watermarking embedded in multi-draft speculative sampling via Poisson processes
Timeline
- 2026First Evolve-cataloged arXiv preprint: ExpBoN on efficient test-time LLM alignment (middle author, non-corresponding).
- 2026Second cataloged preprint: watermarkable multi-draft speculative sampling via Poisson processes (middle author, non-corresponding).
- 2026Profile snapshot assembled: two arXiv preprints, five distinct coauthors, reproducibility mean 55; output, impact, and activity metrics still missing.
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.
- All bibliometrics and radar scores derive from a limited OpenAlex-based snapshot covering only the two Evolve-cataloged 2026 arXiv preprints; no broader publication history is included.
- Output, impact, and activity axes have insufficient underlying data (no publication counts, citation totals, or yearly volume) and are intentionally left unscored.
- “Sicheng Wan” is a shared name; this profile is name-resolved from public bibliographic data plus Evolve-reviewed papers, and no ORCID is attached to confirm identity.
- The frequent-coauthor list contains duplicate entries and includes the profile subject; it is reproduced as reported without de-duplication assumptions.