ML researcher deriving exact theoretical decompositions of LLM inference and depth-recurrent latent reasoning.
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—
The output axis is unscored: the input provides no publication counts, venue list, or per-year volume beyond the two reviewed 2026 arXiv papers, so overall research productivity cannot yet be measured for this profile.
Impact—
Impact is also unscored: no citation counts, h-index, or downstream-usage signals were supplied, so this axis — which would measure citations, influence, and uptake of the two reviewed papers — cannot be estimated honestly.
Focus0/100
The focus score is 0 because the field distribution is an even 0.5/0.5 split between two adjacent subfields: machine learning for LLM inference and evaluation, and machine learning for language model interpretability. Neither dominates the two-paper catalog — one paper sits in each — so the profile reads as balanced across inference-time LLM analysis rather than concentrated in a single line of work.
Collaboration28/100
The collaboration score is 28, drawn from just 3 distinct coauthors in the underlying record. Xinyue Luo recurs across the frequent-coauthor list, while Fei Yu holds the last author position on both reviewed papers without being flagged as corresponding author — a compact, small-team pattern typical of theory-heavy analysis work rather than large-lab empirical projects.
Activity—
The activity axis is unscored: both catalog papers carry the same 2026 year, so there is no multi-year cadence to measure publishing rhythm or sustained output over time from the data provided.
Reproducibility55/100
Reproducibility scores 55, a mean taken over 2 reviewed papers, both rated medium readiness. The KV-cache paper releases no code and tests only two models at H = 128, while the depth-recurrent paper is retrospective with small-N; exact, fully stated analyses partially offset these gaps.
Research summary
Fei Yu is a machine learning researcher whose reviewed work centers on exact, theory-first analysis of how large language models behave at inference time. The Evolve catalog contains two 2026 arXiv papers, both with Fei Yu in the last author position and neither flagged as corresponding author. The first, 'Prediction Dynamics in Depth-Recurrent Language Models', gives an exact margin decomposition showing that depth-recurrent language models keep answers stable through direction and pairing rather than simple translation; Evolve calls it rigorous but retrospective and small-N. The second, 'Divergence Timing and Cumulative Disagreement under KV-Cache Eviction', derives an exact divergence-timing versus exposure decomposition of KV-eviction disagreement using maximal coupling, total variation distance, and conditional Monte Carlo; Evolve notes sharp theory and honest statistics but only two models, H = 128, and no released code. The catalog splits evenly, at 0.5 and 0.5, between LLM inference and evaluation and language model interpretability, and the three distinct coauthors — with Xinyue Luo recurring — point to a compact, theory-oriented collaboration pattern.
Collaboration pattern
The underlying record lists only 3 distinct coauthors, and Xinyue Luo is the recurring partner in the frequent-coauthor data. Fei Yu sits in the last author position on both reviewed 2026 papers without a corresponding-author flag, a placement consistent with a senior or supervisory contribution. The overall pattern is a compact, theory-oriented collaboration footprint rather than a large multi-institution network; institutional affiliation is not available in the input data.
Research keywords
Highlights
Exact margin decomposition for depth-recurrent language models
Divergence timing theory for KV-cache eviction
Timeline
- 2026Published 'Prediction Dynamics in Depth-Recurrent Language Models' on arXiv, in the language model interpretability subfield, in the last author position.
- 2026Published 'Divergence Timing and Cumulative Disagreement under KV-Cache Eviction' on arXiv, in the LLM inference and evaluation subfield.
- 2026Concentrated both reviewed papers within the same year, splitting evenly between the two subfields and working with a small coauthor pool that includes Xinyue Luo.
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 scores come from a limited bibliometric snapshot of real records, not a full citation database; only two reviewed papers underpin this profile.
- 'Fei Yu' is a common name; this profile is name-resolved from public bibliographic data, and no ORCID, institution, or country was available to confirm identity.
- The output, impact, and activity axes are unscored because publication counts, citation metrics, and multi-year cadence data were absent from the input.
- The frequent-coauthor list contains a 'Fei Yu' entry, which likely reflects same-name disambiguation noise in the underlying author records.