Xiao Li

University of Chinese Academy of Sciences

Unclaimed · AI-structured profileOpenAlex

Xiao Li — UCAS researcher spanning VLA robot-learning inference efficiency and efficient language modeling.

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

—

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—

Not yet scored: this snapshot lacks publication counts, venue records, and authorship-position totals. Once a fuller, disambiguated bibliographic record is linked, this axis would quantify total and recent scholarly output.

Impact—

Not yet scored: no citation or influence data is available for this profile. This axis would measure citation volume, per-paper influence, and uptake of the two reviewed papers once citation records are linked.

Focus0/100

The focus axis scores 0, and the raw field distribution explains why: output splits exactly 50/50 between robot learning (vision-language-action models) and natural language processing / efficient language modeling. Rather than one concentrated line, the reviewed record shows two parallel tracks — flow-matching inference for VLA policies on one side, streaming attention and long-context evaluation on the other — so field concentration is currently minimal.

Collaboration51/100

Collaboration scores 51 of 100, based on 12 distinct coauthors recorded in the bibliometric snapshot, with recurring partners including Zhipeng Tang, Xinda Chen, Xiaofang Zhao, and Jie Tang. Both catalog papers list Xiao Li as a middle author rather than first or corresponding, which fits a contributor-style, team-embedded pattern; the mid-range score suggests a moderate network rather than an expansive one.

Activity—

Not yet scored: the snapshot has no per-year publication cadence. This axis would measure recent productivity and publishing rhythm; with both reviewed papers dated 2026, no multi-year trend can yet be established.

Reproducibility38/100

Reproducibility scores 38, the mean across 2 papers with reproducibility assessments. The VLA paper is rated medium: strong simulation evidence and honestly stated limitations, but simulation-only without fast-sampler baselines. The streaming-language-model paper is rated low: sound framing but thin reproduction and statistics details. Overall the signal is early-stage — methods are described, while released artifacts and repro detail lag behind.

Research summary

Xiao Li is a researcher at the University of Chinese Academy of Sciences whose reviewed work splits evenly across two areas: robot learning for vision-language-action (VLA) models, and natural language processing focused on efficient language modeling. The Evolve catalog holds two 2026 arXiv papers, both listing Xiao Li as a middle-position, non-corresponding author. The first, 'Fewer Steps, Better Actions,' introduces Coda, a single learned endpoint correction applied to frozen flow-matching VLA policies; Evolve judges the simulation evidence strong and the limitations honestly stated, while noting the work is simulation-only and lacks fast-sampler baselines. The second, 'Separating Stream Stability from Long-Term Recall,' proposes the ThreeH evaluation contract, showing that attention sinks provide stability rather than long-term memory in streaming language models. Radar scores place collaboration at 51 of 100 (12 distinct coauthors), focus at 0 (an exact 50/50 field split), and reproducibility at 38 (the mean across 2 assessed papers, rated medium and low respectively); output, impact, and activity remain unscored for lack of data. This is an early-stage, name-resolved profile built solely from public bibliographic data and Evolve's reviews.

Collaboration pattern

The record shows a team-embedded, middle-author pattern: both reviewed papers list Xiao Li outside the first and corresponding positions. Twelve distinct coauthors appear in the snapshot, matching the twelve frequent collaborators, including Zhipeng Tang, Xinda Chen, Weining Rao, Wenting Tan, Xiaofang Zhao, Peipei Cao, Xin Zhang, Jie Tang, and Qing Pei. The collaboration radar score of 51/100 sits mid-range, consistent with steady group work rather than a large or leader-centric network.

Research keywords

vision-language-action policiesflow matchinginference accelerationresidual correctionrobot manipulationRoboTwin benchmarkstreaming language modelsattention sinkKV cachelong-context evaluationlong-term memoryThreeH

Highlights

Coda: endpoint correction cuts flow-matching steps for VLA policies
The 2026 arXiv paper adds a single learned endpoint correction to frozen flow-matching VLA policies, reducing inference steps for robot manipulation; Evolve finds strong simulation evidence on RoboTwin and honestly stated limits, but flags sim-only results and missing fast-sampler baselines (repro readiness: medium).
ThreeH: separating stream stability from long-term recall
This 2026 paper frames stability versus access versus utility horizons for streaming language models and proposes the ThreeH evaluation contract, showing attention sinks provide stability rather than memory; the framing is judged sound, though reproduction and statistics details are thin (repro readiness: low).
One efficiency agenda, two modalities
Together the two reviewed papers pursue model efficiency from complementary angles — fewer denoising steps for VLA action generation and clearer evaluation horizons for streaming context — suggesting a coherent interest in making large learned models cheaper to run and more honestly measured.

Timeline

2026Publishes 'Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies' on arXiv, introducing Coda for frozen flow-matching VLA policies.
2026Publishes 'Separating Stream Stability from Long-Term Recall in Language Models' on arXiv, proposing the ThreeH evaluation contract.
2026Both papers enter the Evolve reviewed catalog, with reproducibility readiness rated medium (VLA) and low (language models).

Papers in the Evolve catalog

Frequent co-authors

Zhipeng Tang×1 Xinda Chen×1 Weining Rao×1 Wenting Tan×1 Yuning Wang×1 Xiao Shi×1 Xiaofang Zhao×1 Peipei Cao×1 Xin Zhang×1 Jie Tang×1 Siying Li×1 Qing Pei×1

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.

'Xiao Li' is a common name and no ORCID is provided; this profile is name-resolved from public bibliographic data and may be affected by homonym conflation.
Radar scores (collaboration 51, focus 0, reproducibility 38) come from the platform's bibliometric snapshot only; the output, impact, and activity axes are missing, and no citation counts are available.
The evidence base is two 2026 arXiv preprints reviewed by Evolve; neither is venue-confirmed as peer-reviewed, and country/affiliation metadata is incomplete.
Collaboration and timeline statements derive solely from the reviewed catalog and the provided coauthor list, not from a full publication history.