RIKEN researcher bridging efficient LLM inference and text watermarking, from sparse attention to diffusion LMs.
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—
Output metrics — publication volume and its trajectory over time — are not available in the current snapshot: only the two Evolve-reviewed 2026 papers exist as evidence, so this axis is left unscored until a fuller bibliometric record is ingested.
Impact—
Impact measures such as citation counts, h-index, and downstream uptake are absent from the provided data. Nothing in the snapshot indicates how the two 2026 papers have been received, so this axis is reported as insufficient rather than estimated.
Focus0/100
The focus score sits at 0 on the 0–100 scale, and the underlying field distribution explains why: work is split exactly 50/50 between natural language processing / text watermarking and machine learning for efficient LLM inference. Rather than converging on one specialty, current output alternates between watermarking diffusion language models and KV-cache efficiency, making this an intentionally two-track profile with low measured concentration.
Collaboration59/100
The collaboration score is 59, computed from 18 distinct coauthors in the snapshot. Twelve recurring partners are listed, including Masashi Sugiyama, Takashi Ishida, and Dong-Dong Wu, alongside Jerry Huang, Qiuhao Zeng, and Gezheng Xu. On both reviewed 2026 papers Gang Niu appears in a middle position rather than as corresponding author, indicating team-based contribution within larger collaborations.
Activity—
Activity metrics — publication cadence, recency of output, and year-over-year consistency — are not part of the provided snapshot. The only dated evidence is two arXiv papers in 2026, so this axis cannot be scored and is flagged as data-insufficient.
Reproducibility55/100
Reproducibility is covered for 2 papers with a mean readiness score of 55, and both the ARM and DenMark reviews rate reproReadiness as medium. Evolve notes concrete caveats: ARM's efficiency gains are real-looking but baseline confounds weaken attribution, while DenMark's strongest claim carries a 12x generation cost and weak expansion TPR — conditions that temper straightforward replication.
Research summary
Gang Niu is a researcher at RIKEN whose recent work, as indexed in the Evolve catalog, spans two equally weighted directions: natural language processing / text watermarking and machine learning for efficient LLM inference. The bibliometric field distribution assigns 0.5 to each, and the radar focus score sits at 0 on the 0–100 scale, consistent with a deliberately two-track profile rather than a single specialty. Two 2026 arXiv papers anchor the catalog. DenMark proposes what Evolve summarizes as the first semantic watermark for diffusion language models, reporting the best result in all 48 test cells, though at a 12x generation cost and with weaker expansion TPR. ARM introduces a differentiable routed-memory KV cache with gated writes and adaptive retrieval; Evolve judges the efficiency gains plausible but notes baseline confounds that weaken attribution. Both papers list Gang Niu in a middle author position without corresponding-author status, so this profile reflects a contributor role on current output. Collaboration data shows 18 distinct coauthors (collaboration score 59), with frequent partners including Masashi Sugiyama, Takashi Ishida, and Dong-Dong Wu. Reproducibility coverage spans 2 papers with a mean readiness of 55, both rated medium. Output, impact, and activity metrics are not yet available and are flagged as missing rather than estimated.
Collaboration pattern
The snapshot records 18 distinct coauthors with a collaboration score of 59, and 12 recurring partners: Masashi Sugiyama, Takashi Ishida, Dong-Dong Wu, Farshid Nooshi, Naoto Yokoya, Jerry Huang, Qiuhao Zeng, Gezheng Xu, Tianhao Ma, Weihao Xuan, Peng Lu, and Ruiyi Fang. On both reviewed papers Gang Niu is a middle-listed, non-corresponding author, suggesting contribution within larger team-led projects rather than solo-led groups.
Research keywords
Highlights
First semantic watermark for diffusion language models (per Evolve review)
Differentiable routed-memory KV cache for learnable sparse control
A deliberately two-track 2026 output across inference and watermarking
Both cataloged papers appeared in 2026, one in each of the profile's two equally weighted fields, pairing a watermarking contribution for diffusion language models with a KV-cache efficiency contribution — consistent with the 0.5/0.5 field distribution and the low focus score.
Timeline
- 2026DenMark: Robust Semantic Watermarking for Diffusion Language Models released on arXiv; Evolve rates its reproducibility readiness medium.
- 2026ARM: Attention with Routed-Memory for Learnable Sparse Control released on arXiv, entering the efficient LLM inference track.
- 2026Both papers enter the Evolve catalog with Gang Niu in a middle, non-corresponding author position, giving a 0.5/0.5 split across the profile's two fields.
- 2026Reproducibility snapshot covers both 2026 papers with a mean readiness score of 55.
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.
- Name is resolved from public bibliographic data; 'Gang Niu' carries common-name risk, so identity matching relies on ORCID 0000-0002-7353-5079.
- Bibliometric scores come from the provided OpenAlex-derived radar data only; citation counts, h-index, and output volumes are unavailable (dims marked missing).
- Only the 2 Evolve-reviewed papers are covered here; other publications by this author fall outside the current data scope.
- Both papers are arXiv-sourced preprints from 2026 with middle-author, non-corresponding positions, so authorship role may change in published versions.