Embodied AI and efficient-LLM researcher at Tongji University spanning ego-robot pretraining and KV-cache compression
2
papers here
2
reviewed by Evolve
1
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—
Publication-volume metrics are not yet available for this profile, so the axis is left unscored. Once a fuller bibliographic record is linked, it would measure total output, yearly paper counts, and venue quality.
Impact—
Citation-based impact metrics (citations, h-index, highly cited papers) are absent from the current snapshot, so this axis is unscored. It would measure how widely Di Wu's work is cited and taken up by other researchers.
Focus0/100
Scores 0, the lowest focus band, because the field distribution splits exactly 50/50 between two distinct areas — efficient large language model architecture (machine learning) and robotics/embodied AI (robot learning). The two 2026 papers sit in unrelated problem spaces: one pretrains embodied foundation models from ego-robot data, the other compresses KV cache in a 552B-parameter mixture-of-experts model. Topical concentration in the current record is therefore minimal.
Collaboration86/100
Scores 86 on collaboration, grounded in 72 distinct coauthors across the reviewed record — a large network for a two-paper snapshot. The pattern mixes industrial-scale group authorship (the DeepSeek-AI collective on the KV-cache work) with a long academic roster on AtomEgo (Anyi Xu, B. Li, Bing Xue, and Bochao Wu among the listed frequent coauthors). This suggests comfort operating inside big, distributed teams rather than small two-person collaborations.
Activity—
Recent-activity metrics (papers per recent year, preprint cadence) are not yet computed, so this axis is unscored. It would measure how actively Di Wu is publishing right now, beyond the two 2026 preprints already reviewed.
Reproducibility38/100
Reproducibility averages 38 across 2 scored papers. AtomEgo is rated medium: Evolve calls its three-paradigm comparison valuable but notes the gains lack statistical tests and step-matching is inconsistent. DeepSeek-V4.1-Flash is rated low: results are self-evaluated with no ablations and mixed Table 1 results. Neither paper currently meets a high reproducibility bar, largely due to missing statistical rigor and internal evaluation rather than absent artifacts.
Research summary
Di Wu is a researcher at Tongji University whose 2026 record spans two demanding corners of modern AI: embodied robot learning and efficient large language model architecture. As first author, Di Wu led AtomEgo, a systematic three-paradigm study of ego-robot co-training for embodied foundation model pretraining built on 2,659 hours of data; Evolve judges the comparison valuable while noting that reported gains lack statistical tests and that step-matching is inconsistent. In parallel, Di Wu contributed as a middle author to DeepSeek-V4.1-Flash, where the DeepSeek-AI team describes a 552B-parameter mixture-of-experts model said to match frontier agents with one-quarter to one-eighth the KV cache — engineering-strong in Evolve's view, but self-evaluated and without ablations. The bibliometric field distribution splits exactly 50/50 between these two areas, yielding a focus score of 0, while a network of 72 distinct coauthors drives a collaboration score of 86. Reproducibility averages 38 across the two scored papers, with the low readiness of the self-evaluated DeepSeek work weighing against AtomEgo's medium rating. Output, impact, and activity metrics are not yet available, so this early-stage profile should be read as a two-paper snapshot rather than a settled career record.
Collaboration pattern
Di Wu's collaboration footprint is wide: 72 distinct coauthors, driving a collaboration score of 86. The record pairs a huge industrial group author (DeepSeek-AI) on the efficient-LLM paper with an extended academic roster on AtomEgo, where frequent coauthors include Anyi Xu, B. Li, BingCheng Xian, Bochao Wu, and Boyi Deng. The pattern points to large-team, resource-intensive projects rather than small-lab work.
Research keywords
Highlights
AtomEgo: a systematic three-paradigm study of ego-robot co-training
Contributing to DeepSeek-V4.1-Flash's KV cache compression push
Timeline
- 2026Posted the first-author preprint AtomEgo on arXiv, exploring ego-robot data integration for embodied foundation model pretraining.
- 2026Appeared as a middle author on the arXiv paper DeepSeek-V4.1-Flash, pushing KV cache compression in a 552B-parameter MoE.
- 2026Both catalog entries landed in the same year, splitting output evenly (50/50) between robot learning and efficient LLM architecture.
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.
- "Di Wu" is a common name; this profile is name-resolved from public bibliographic data, and no ORCID is linked.
- Bibliometrics come from an OpenAlex-derived snapshot only; country is not recorded and the institution is as listed in the input.
- Output, impact, and activity axes are unscored — no citation counts, h-index, or publication-volume data are present in the input.
- Only two preprints (both 2026, arXiv source) have been reviewed by Evolve; no peer-reviewed venue record is reflected here.