First-author researcher at Imperial College London: theory-heavy test-time LLM alignment and watermarkable inference.
2
papers here
2
reviewed by Evolve
2
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—
Output metrics are unavailable in the current snapshot, so this axis is unscored. It would measure publication volume, venue quality, and year-over-year consistency across the researcher's full record beyond the two catalogued arXiv papers.
Impact—
Impact cannot be scored yet: no citation counts, h-index, or adoption signals are available. This axis would measure how widely the test-time alignment and watermarking papers are cited, reused, or built upon by other research groups.
Focus0/100
The focus score of 0/100 reflects an evenly split field distribution: the two catalogued papers sit 50/50 across machine learning (LLM alignment, inference-time methods) and its counterpart on efficient LLM inference and watermarking. Read charitably, this is a method-driven spread — one alignment paper, one inference-security paper — rather than scattered work; a larger paper set is needed to judge whether the agenda consolidates into a single niche.
Collaboration36/100
Collaboration scores 36/100, based on 5 distinct coauthors across the two catalogued papers. The recurring circle is small and stable: Sicheng Wan, Zhan Gao, and Deniz Gündüz appear among frequent coauthors, consistent with a tight Imperial College London working group. Both papers list the researcher as first author and neither as corresponding, suggesting hands-on technical leadership within a compact team rather than large multi-institution consortia.
Activity—
Activity is unscored: the snapshot provides no publication-cadence or per-year productivity data beyond two papers dated 2026. This axis would measure research tempo — how consistently new work appears year over year — and cannot be estimated from a single-year, two-paper sample.
Reproducibility55/100
Reproducibility scores 55/100, computed as a mean over 2 reviewed papers, both individually rated medium on repro-readiness. The work is theory-dense — exact finite-n tilted decompositions, geometric convergence proofs, and Poisson-process constructions with unbiasedness guarantees — which supports faithful reimplementation, but reviewers flagged minor table inconsistencies in the watermarking paper, and code availability is not recorded in the reviewed data.
Research summary
Yanxiao Liu is a researcher at Imperial College London whose catalogued work centers on inference-time methods for large language models, spanning LLM alignment and efficient, watermarkable decoding. Both papers in the Evolve catalog are first-authored arXiv preprints dated 2026. The first, ExpBoN, introduces an exponential-noise soft variant of best-of-n sampling for efficient test-time LLM alignment, with an exact finite-n tilted decomposition, geometric convergence guarantees, and an ExpGSI variant that saves GSI compute at comparable accuracy; its keywords connect the design to noisy-mechanism ideas such as report-noisy-max and permute-and-flip. The second embeds an unbiased watermark into multi-draft speculative sampling via Poisson processes, preserving acceptance rate while adding draft invariance and communication-free coupling; Evolve's review calls the theory rigorous while noting minor table inconsistencies. Together the two papers sketch a coherent, theory-first agenda: replace expensive or vulnerable sampling procedures with probabilistic mechanisms backed by exact guarantees. The collaboration footprint is compact — 5 distinct coauthors, with Sicheng Wan, Zhan Gao, and Deniz Gündüz recurring — and both papers are rated medium on repro-readiness, averaging 55/100. Output, impact, and activity axes remain unscored because citation and productivity data are absent from the snapshot, so this reads best as an early-stage, arXiv-only profile of a focused technical line of work.
Collaboration pattern
Collaboration is compact: 5 distinct coauthors across the two catalogued first-authored papers, with Sicheng Wan, Zhan Gao, and Deniz Gündüz recurring as frequent coauthors — a pattern consistent with a small, stable Imperial College London group. Neither paper lists the researcher as corresponding author. Notably, the researcher's own name appears in the frequent-coauthor list, which is likely a name-disambiguation artifact of the underlying bibliographic data rather than evidence of a distinct collaborator.
Research keywords
Highlights
ExpBoN: exact tilted decomposition for exponential-noise best-of-n
Unbiased watermarking of multi-draft speculative sampling
A theory-first inference-time agenda
Across both catalogued papers — one on test-time alignment, one on efficient watermarkable decoding — the common thread is probabilistic mechanism design with exact guarantees: exponential noise in ExpBoN and Poisson processes in the speculative-sampling watermark, each justified by formal analysis rather than heuristics alone.
Timeline
- 2026Posts the first-authored arXiv preprint 'ExpBoN: Exponential-Noise Best-of-n for Efficient Test-Time LLM Alignment' on efficient test-time alignment.
- 2026Posts the first-authored arXiv preprint 'Watermarkable Multi-Draft Speculative Sampling via Poisson Processes' on unbiased LLM watermarking.
- 2026Both catalogued works land in the same year, establishing an inference-time efficiency and alignment agenda at Imperial College London with a compact coauthor circle.
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 bibliometric scores derive from a small snapshot covering only two reviewed arXiv papers, so they may not reflect the researcher's full publication record.
- No citation counts, h-index, output-volume, or year-by-year activity data were available; the output, impact, and activity radar axes are therefore unscored.
- 'Yanxiao Liu' is name-resolved from public bibliographic data; the researcher's own name also appears in the frequent-coauthor list, indicating possible disambiguation noise.
- Both papers are 2026 arXiv preprints; peer-review status, venue, and code availability are not recorded in the reviewed data.