Salesforce researcher at the intersection of efficient LLM inference and LLM-agent memory systems
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-02) 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 contains no publication-count, venue, or authorship-share data beyond two 2026 preprints, so research output cannot be measured. The axis would capture total papers, venue quality, and first/corresponding-author share.
Impact—
Not yet scored: citation counts and downstream adoption are absent from the current data, so impact cannot be assessed. This axis would measure citations, references from other papers, and real-world uptake of the methods.
Focus0/100
The focus score is 0 because the field distribution splits evenly, 50/50, between 'machine learning (LLM architecture and efficient inference)' and 'LLM-based agents and agent memory systems'. Read cautiously: both halves sit inside modern LLM research and look like complementary halves of one systems-plus-agents agenda rather than scattered work.
Collaboration54/100
Collaboration scores 54, mid-range. The snapshot counts 14 distinct coauthors, and 12 recur frequently, including Zhiheng Hu, Yixun Wei, Jian Zhou, Xiangyu Zhang, Daxin Jiang, and Yibo Zhu. On both cataloged papers Yang Li sits in a middle author position, neither first nor corresponding, consistent with contributing within larger team-based efforts rather than leading them.
Activity—
Not yet scored: the snapshot indexes no multi-year publishing record — only two 2026 arXiv entries — so activity cadence cannot be measured. The axis would track yearly output volume and publishing rhythm once a longer record is available.
Reproducibility38/100
Reproducibility averages 38 across the 2 papers Evolve reviewed. JITMEM rates medium: good ablations but no significance tests for its ALFWorld/WebShop/τ²-bench results. KITE rates low: reported 67B-versus-47B/63B FLOPs-matched gains rest on single runs, analytic cost estimates, and no ablations. Together they pull the mean down to 38.
Research summary
Yang Li is a researcher at Salesforce (United States) working across two tightly linked areas of large language model research: LLM architecture and efficient inference, and LLM-based agents with memory systems. Evolve's catalog contains two 2026 arXiv papers listing Yang Li as a middle author, neither corresponding. The first, KITE, presents a KV-invariant transformer expansion for efficient agentic LLM scaling; Evolve notes that KITE/SST 67B upcycling reportedly beats from-scratch 47B/63B models at equal FLOPs with cheaper prefill, while flagging solid controls alongside single-run results, analytic cost comparisons, and no ablations. The second, JITMEM, defers memory curation to read time and trains the curator with GRPO, reporting large gains on ALFWorld and WebShop with only modest gains on tau2-bench; Evolve credits good ablations but no significance tests. The bibliometric record splits evenly, 50/50, between the two subfields, and 14 distinct coauthors underpin a mid-range collaboration score of 54. Reproducibility averages 38 over the two reviewed papers — medium readiness for JITMEM, low for KITE. Output, impact, and activity axes remain unscored in this snapshot.
Collaboration pattern
The cataloged record involves 14 distinct coauthors, and 12 collaborators recur frequently, including Zhiheng Hu, Yixun Wei, Jian Zhou, Xiangyu Zhang, Daxin Jiang, and Yibo Zhu. On both reviewed papers Yang Li appears in a middle author position — neither first nor corresponding — consistent with contributing inside larger team-based efforts at Salesforce rather than leading them. The collaboration score of 54 is mid-range, reflecting moderately broad but not yet extensive co-authorship in this snapshot.
Research keywords
Highlights
KITE: KV-invariant transformer expansion — reported 67B upcycling beats from-scratch 47B/63B at equal FLOPs
JITMEM: deferring agent memory curation to read time, trained with GRPO
Timeline
- 2026KITE, a KV-invariant transformer expansion for efficient agentic LLM scaling, appears on arXiv with Yang Li as a middle author.
- 2026JITMEM, on task-adaptive memory curation for LLM agents, appears on arXiv, also with a middle author position.
- 2026Both papers enter Evolve's catalog in the same year, splitting the record evenly between efficient inference and agent memory.
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.
- Profile is name-resolved from public bibliographic data; 'Yang Li' is a common name, so identity continuity beyond the two cataloged papers is uncertain.
- Only two 2026 arXiv preprints were reviewed; no peer-reviewed venue, citation, or publication-count data are included, so output, impact, and activity dimensions are unscored.
- The reproducibility score (mean 38) rests on Evolve's review of just 2 papers and may shift as more work is reviewed.
- Collaboration and focus scores come from a limited bibliometric snapshot (14 distinct coauthors; an even 50/50 field split) and may not reflect the full career record.