Exact margin and divergence-timing theory for depth-recurrent LMs and KV-cache eviction in LLM 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—
Unscored: the data contains no publication counts or venue record for this researcher beyond two 2026 preprints. This axis would measure cumulative research output — papers, venues, and authorship roles — once a fuller bibliographic record is linked.
Impact—
Unscored: no citation counts, h-index, or downstream-usage signals are available, so influence cannot be estimated. This axis would measure how far the two cataloged papers propagate — citations, referencing venues, and uptake — once citation data is attached.
Focus0/100
The focus score is 0, computed from a field distribution that splits exactly evenly: 0.5 machine learning (LLM inference and evaluation) and 0.5 machine learning (language model interpretability), one cataloged paper on each side. Both subfields sit inside machine learning and share a theory-driven evaluation style, so the split reflects two adjacent angles on one theme rather than scattered interests. More single-subfield output would lift this score.
Collaboration28/100
The collaboration score of 28 is grounded in a raw count of 3 distinct co-authors across the two cataloged papers, with Fei Yu the only name appearing repeatedly in the frequent-coauthor list. Both papers are first-authored without a corresponding-author flag, consistent with a small, hands-on working mode. The list also contains the researcher's own name, an apparent disambiguation artifact that inflates apparent ties.
Activity—
Unscored: no publication-volume or cadence metrics over time are provided, only two 2026-dated papers. This axis would measure recent research tempo — outputs per period and recency of work — and cannot be scored from a single-year snapshot.
Reproducibility55/100
The reproducibility score of 55 is the mean over 2 papers, both rated medium readiness. Evolve's verdicts explain the ceiling: the KV-cache paper releases no code and tests only 2 models at H=128, while the depth-recurrent paper is retrospective with small-N. Exact, closed-form derivations support replication, but the absence of artifacts keeps independent verification nontrivial.
Research summary
Xinyue Luo is a machine learning researcher whose cataloged work concentrates on theoretical analysis of large language model behavior at inference time. The Evolve catalog holds two 2026 arXiv preprints, both first-authored. 'Prediction Dynamics in Depth-Recurrent Language Models' gives an exact margin decomposition showing that direction and pairing — not translation alone — keep depth-recurrent latent-reasoning answers stable under test-time compute and early exit; Evolve's verdict calls the work rigorous but retrospective, with small-N evidence. 'Divergence Timing and Cumulative Disagreement under KV-Cache Eviction' derives an exact divergence-timing-versus-exposure decomposition using maximal coupling, total variation distance, and conditional Monte Carlo; Evolve judges the theory sharp and the statistics honest while noting only two models, H=128, and no released code. The field distribution splits evenly between LLM inference and evaluation and language model interpretability (0.5 each), and the reproducibility mean is 55 across two papers rated medium readiness. Collaboration is compact — 3 distinct co-authors, with Fei Yu the recurring name — and both papers carry first-author rather than corresponding-author flags. Output, impact, and activity axes remain unscored because citation and volume data are absent, and no institution, country, or ORCID is attached, so this profile is name-resolved from public bibliographic data.
Collaboration pattern
Collaboration is compact: 3 distinct co-authors across two first-author papers, with Fei Yu the only repeatedly listed collaborator. Neither paper carries a corresponding-author flag, suggesting a hands-on, small-team mode rather than large-lab coordination. The frequent-coauthor list oddly includes the researcher's own name — likely an entity-resolution artifact — so effective external ties may be even fewer. No institutional affiliation is on record, which limits assessment of broader networks.
Research keywords
Highlights
Exact margin decomposition keeps depth-recurrent answers stable
Sharp divergence-timing theory for KV-cache eviction
Timeline
- 2026First cataloged paper: 'Prediction Dynamics in Depth-Recurrent Language Models', first-authored on arXiv, in language model interpretability.
- 2026Second first-author arXiv preprint, 'Divergence Timing and Cumulative Disagreement under KV-Cache Eviction', in LLM inference and evaluation.
- 2026Both 2026 papers together define an even split between interpretability and inference/evaluation, with a reproducibility mean of 55 and 3 distinct co-authors.
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 signals come from a sparse snapshot: only two 2026 arXiv preprints inform this profile; no venue, citation, or volume data is included.
- No institution, country, or ORCID is attached; the profile is name-resolved from public bibliographic data, and 'Xinyue Luo' may match multiple researchers.
- Output, impact, and activity radar axes are unscored because the underlying data is missing; no values were estimated.
- The frequent-coauthor list repeats the researcher's own name, an apparent disambiguation artifact; collaboration counts should be read with that in mind.