Amazon researcher spanning multimodal vision-language, speech generation, LLM agents, alignment, and inference serving
5
papers here
5
reviewed by Evolve
1
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—
Output is not yet scored: aggregated publication counts for this OpenAlex identity are incomplete in the current snapshot. This axis would measure total works, first- and corresponding-author share, and yearly output volume.
Impact—
Impact is not yet scored: citation counts, h-index, and per-paper influence are insufficiently populated for this identity in the current snapshot. This axis would measure citations, h-index, and uptake of the 2026 preprints.
Focus0/100
OpenAlex field weights split perfectly evenly — five fields at 0.20 each (speech/audio machine learning, LLM inference and serving, multimodal vision-language, LLM alignment, LLM agents and reinforcement learning) — which produces the lowest focus score. The 2026 catalog mirrors this breadth with one reviewed paper in each area. Such range fits project-based industrial lab work, but no single field yet dominates the record.
Collaboration83/100
Collaboration scores 83 from real bibliometric data: 64 distinct co-authors appear in the recent record, around a recurring core of 12 collaborators including Ziyang Ma, Zhikang Niu, and Wenming Tu. All five 2026 reviewed papers are multi-authored, spanning speech, vision-language, agents, serving, and alignment teams — a team-based industrial research pattern, with one first-author and one last-author placement and no corresponding role.
Activity77/100
Activity scores 77: the record lists 13 works over the last three years, with yearly counts rising from 2 in 2025 to 11 in 2026. All five Evolve-reviewed papers appeared as 2026 arXiv preprints, indicating a sharply accelerating and currently intense publication cadence within a single year.
Reproducibility55/100
Reproducibility averages 55 across 5 reviewed papers, each rated medium readiness. Verdicts credit strong internal evidence — Elastic Horizon's thorough ablations, RedKnot-MLA's explicit evidence labeling — while repeatedly flagging missing variance reports, significance tests, human MOS, or system baselines. AuK's open-source release helps; statistical rigor is the recurring gap.
Research summary
Yang Liu is a US-based researcher at Amazon whose 2026 output spans five machine learning areas in near-equal measure: multimodal vision-language modeling, speech generation, LLM agents and reinforcement learning, LLM inference and serving systems, and LLM alignment. OpenAlex field weights assign 0.20 to each of these five fields, and the Evolve catalog mirrors that spread with one reviewed paper per area. The recent record lists 13 works over three years, accelerating from 2 in 2025 to 11 in 2026, with 64 distinct co-authors and a recurring core of 12 collaborators — a pattern typical of team-based industrial research. Five reviewed papers define the profile: RedKnot-MLA (first author) achieves 2.02-3.84x TTFT gains via exact head-level KV reuse for DeepSeek-V4 long-context serving; AuK is an open-source 1.5B-parameter unified speech generation and editing model; Elastic Horizon adds a closed-loop interaction-horizon controller to agentic RL; AlignDiff filters DPO preference data using model-intrinsic reward margins; and a tokenizer study shows task-specific validation loss judges image tokenizers better than reconstruction metrics. Reproducibility averages 55 (medium readiness throughout): ablations are often strong, but variance reporting, significance tests, human evaluation, and system baselines are recurring gaps. Citation-based output and impact scores are withheld pending sufficient data.
Collaboration pattern
Collaboration is the strongest scored axis (83): 64 distinct co-authors appear in the recent record, with a recurring core of 12 — Ziyang Ma, Zhikang Niu, Wenming Tu, Tianrui Wang, Ruiqi Yan, Junxi Liu, Yanru Huo, Nickk Huang, Qicong Xie, Zeyu Xie, Hui Wang, and Haitao Li. All five reviewed papers are multi-authored; Yang Liu is first author once, last author once, and never the corresponding author — a contributor-within-large-teams pattern typical of industrial research.
Research keywords
Highlights
First-author RedKnot-MLA: exact head-level KV reuse for DeepSeek-V4 long-context serving
AuK: open-source 1.5B-parameter unified speech generation and editing
Elastic Horizon: closed-loop control of interaction horizons in agentic RL
AlignDiff: model-intrinsic signals guide DPO preference data selection
Studying image tokenizers as visual languages in unified multimodal models
Timeline
- 2025Earliest works in the current OpenAlex record appear (2 works), preceding an intense 2026 phase.
- 2026Records 11 works in the year; five Evolve-reviewed arXiv papers span speech, vision-language, agents, serving, and alignment.
- 2026First-author RedKnot-MLA reports 2.02-3.84x TTFT gains from exact head-level KV reuse for DeepSeek-V4 long-context serving.
- 2026Co-authors the AuK technical report, an open-source 1.5B-parameter unified speech generation and editing model.
- 2026Elastic Horizon's closed-loop horizon controller, AlignDiff's DPO data selection, and the image-tokenizer study complete the year's reviewed catalog.
Papers in the Evolve catalog
Studying Image Tokenizers as Visual Languages in Unified Multimodal Models
AuK Technical Report: An Open-Source Foundational Model for Speech Generation and Editing
Elastic Horizon: Discovering the Effective Interaction Frontier in Agentic Reinforcement Learning
- first author
RedKnot-MLA: Multi-Head Offline-Online Reuse for DeepSeek-V4 Long-Context Serving
AlignDiff: Exploiting Model-Intrinsic Information for Better Preference Data Selection
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.
- Bibliometrics come from OpenAlex only; the snapshot registers h-index and i10-index as 0, and the output and impact axes are withheld as insufficient.
- All five reviewed papers are 2026 arXiv preprints; peer-review status and citation impact are not yet observable.
- "Yang Liu" is a very common name; this profile is resolved from public bibliographic data (OpenAlex identity, ORCID 0000-0001-7300-9215) and may include homonymous records.
- The reproducibility score covers only the 5 Evolve-reviewed papers; the remaining recent works are unreviewed.