Early-career ML researcher spanning learnable sparse coding and mechanistic analysis of multimodal LLM hallucination
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-16) 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 metrics are available yet. This axis measures total publication volume and venue composition — annual work counts and the preprint versus peer-reviewed split — but the snapshot lacks a work-count series sufficient to score it, so this profile reports no output figure.
Impact—
No citation data is available yet. This axis measures uptake and influence — citations, h-index, and how often others build on the work — but the snapshot shows an h-index and i10-index of 0 with no citation series, so it is left unscored rather than estimated.
Focus0/100
Focus scores 0/100, and the raw data supports that: the field distribution splits evenly — 0.5 computer vision / machine learning and 0.5 multimodal machine learning — and each of the five indexed topics carries a single work, from convolutional sparse coding to natural language processing, sparse sensing, and denoising. With only two cataloged papers, this reads as method-spanning breadth rather than a settled specialty; whether the breadth is deliberate or an early-stage record still forming remains open.
Collaboration41/100
Collaboration scores 41/100 on a modest base: 7 distinct coauthors across the two 2026 works, including recurrent collaborators such as Meng'en Qin and Song Wang. Authorship sits in the last position on the sparse-coding paper and a middle position on the MLLM paper, without corresponding-author credit — a contributor role within a compact team rather than visible independent group leadership.
Activity32/100
Activity scores 32/100, reflecting a thin recent record: 2 works in the last three-year window, both surfacing in 2026 according to the counts-by-year series, and both arXiv preprints. With no earlier years represented in the snapshot, output cadence cannot yet be separated from a genuinely recent start or from prior work that was simply not indexed.
Reproducibility55/100
Reproducibility scores 55/100, Evolve's mean across the 2 reviewed papers, both rated medium readiness. Both works report strong gains but carry evaluation caveats — the sparse-coding paper is single-run and evaluated only under noise, while the hallucination study involves per-benchmark tuning and thin causal evidence — so reimplementation is plausible, but independent confirmation is not trivial.
Research summary
Youlu Xing, affiliated with University Town of Shenzhen (China), is an early-stage machine learning researcher whose profile is name-resolved from OpenAlex bibliographic data; no ORCID is linked. The Evolve catalog reviews two 2026 arXiv preprints involving this author, spanning two complementary strands. The first, Adaptive Convolutional Sparse Coding via Information Bottleneck for Robust Visual Signal Representation, unrolls FISTA into a learnable network whose sparsity coefficient adapts label-free at test time for robust visual signal representation; Evolve finds the reported gains strong but tempered by single-run, noise-only evaluation. The second, HEAL (MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads), attributes multimodal hallucination to information distribution drift inside synergy heads, combining attention-head analysis, partial information decomposition, causal intervention, and difference-in-differences before calibrating at inference time; gains are broad, but per-benchmark tuning and causal evidence remain thin. A shared instinct — lightweight test-time or inference-time correction instead of retraining — links the two papers. Raw metrics are correspondingly thin: 2 works in the recent three-year window, 7 distinct coauthors, h-index and i10-index of 0, and an even 50/50 split between computer vision / machine learning and multimodal machine learning. Evolve rates reproducibility at 55/100 across the two medium-readiness papers, while output and impact axes remain unscored for lack of data.
Collaboration pattern
The snapshot shows 7 distinct coauthors across the two 2026 arXiv works, with a recurring core — Meng'en Qin, Yinchen Liu, Mingxuan Cui, Junye Chen, Jucheng Liu, Song Wang, and Ruize Han. Youlu Xing appears as last author on the sparse-coding paper and middle author on the MLLM paper, without corresponding-author credit — a contributor position within a compact team, consistent with a collaboration score of 41/100.
Research keywords
Highlights
Information-bottleneck-guided adaptive convolutional sparse coding
Tracing MLLM hallucination to drift in synergy heads
A recurring instinct: inference-time correction over retraining
Across both reviewed papers the pattern is the same — label-free λ adaptation at test time in the sparse-coding work, inference-time calibration in the hallucination work — lightweight fixes that avoid retraining the underlying model.
Timeline
- 2026Posts 'Adaptive Convolutional Sparse Coding via Information Bottleneck for Robust Visual Signal Representation' to arXiv as last author.
- 2026Posts 'MLLMs Hallucinate when Information Distribution Drifts in Synergy Heads' (HEAL) to arXiv as a middle author, on multimodal hallucination.
- 2026Both cataloged works rate medium reproducibility readiness, giving a 55/100 Evolve reproducibility mean across the 2 papers.
- 2026Snapshot records 2 works in the recent three-year window with 7 distinct coauthors; no earlier years appear in the index.
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 snapshot is sourced from OpenAlex only; no ORCID is linked, so identity is name-resolved from public bibliographic data and homonyms cannot be fully excluded.
- Only the two arXiv preprints reviewed by Evolve are in scope; no peer-reviewed venue, funding, or award information is included.
- Output and impact axes are unscored because the snapshot reports h-index 0 and i10-index 0 with no citation history.
- The counts-by-year series contains only 2026 (2 works), so earlier productivity cannot be assessed from this snapshot.