Siqi Li

Xiaomi (China) · China

Unclaimed · AI-structured profileOpenAlex

Industry ML researcher at Xiaomi working on efficient LLM inference and self-evolving embodied agents

papers here

reviewed by Evolve

first-author

corresponding

—

works · OpenAlex

h-index · OpenAlex

Bibliometric figures come from OpenAlex (2026-10-17) 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 measures publication volume and productivity over time. It is not yet scored for this profile: only 2025–2026 records exist (14 works total), and stable output indicators are unavailable this early. No estimate is made.

Impact—

Impact captures citation-based influence (citations, h-index, i10-index). This profile reports h-index 0 and i10-index 0, which carry no signal for 2026-era arXiv work, so the axis is left unscored rather than estimated.

Focus100/100

Focus scores 100/100, the maximum. The field distribution attributes the labeled record to a single field — natural language processing (efficient LLM inference) — indicating a tightly concentrated recent line of work on inference acceleration and, per the two reviewed 2026 papers, LLM-driven embodied agents. Topic counts add adjacent breadth (neural network applications 8, visual saliency 4, multimodal learning 4), but the core specialization is unambiguous.

Collaboration46/100

Collaboration scores 46/100, a mid-range value drawn from 9 distinct coauthors in the raw record. Recurring partners include Zuchao Li, Lefei Zhang, Ziyang Ma, and Börje F. Karlsson. Both reviewed 2026 papers place Li in middle authorship on multi-author arXiv teams, without corresponding-author role — a contributor pattern within larger group efforts rather than small-team or solo leadership.

Activity79/100

Activity scores 79/100, backed by raw counts of 1 work in 2025 and 13 in 2026 — 14 works within the recent three-year window. The steep 2026 ramp-up coincides with the two reviewed arXiv papers on efficient LLM inference and self-evolving embodied agents, indicating a fast-moving research tempo concentrated in the present year.

Reproducibility55/100

Reproducibility scores 55/100, computed from a single reviewed paper (reproPapers = 1, reproMean 55). Both catalog entries rate medium reproReadiness. Evolve's verdicts flag concrete issues: ECHO's broad 2.4–2.9x speedup results read as credible, but its 'training-free' claim conflicts with fine-tuning, while PRACTICE shows internal number conflicts and no variance tests. The evidence is real but thin — one paper carries the whole score.

Research summary

Siqi Li is a machine learning researcher at Xiaomi (China) whose recent work centers on efficient large language model inference and learning-based autonomous agents. The OpenAlex-resolved profile records 14 works in the last three years — 1 in 2025 and 13 in 2026 — driving an activity score of 79/100, with a fully concentrated field profile (focus 100/100) in natural language processing (efficient LLM inference). Topic counts show wider coverage: advanced neural network applications (8), visual attention and saliency detection (4), multimodal machine learning (4), 3D shape modeling (3), and topic modeling (2). Evolve has reviewed two 2026 arXiv papers: ECHO, which claims 2.4–2.9x lossless speedup via dual-loop early-layer speculative decoding, and PRACTICE, a self-evolving embodied agent whose skill learner improves a frozen executor by +9.7/+2.6 points over experience baselines. Both rate medium on reproducibility readiness, and the reproducibility dimension scores 55/100 from one reviewed paper. Collaboration scores 46/100 across 9 distinct coauthors, with Li in middle authorship positions on both reviewed papers. Output and impact are not yet scored: h-index and i10-index are both 0, reflecting how recent the record is rather than measured performance.

Collaboration pattern

Li works within multi-author industry-research teams: the raw record lists 9 distinct coauthors (collaboration score 46/100), with recurring partners including Zuchao Li, Lefei Zhang, Ziyang Ma, Zihong Zhang, and Börje F. Karlsson. On both reviewed 2026 arXiv papers Li holds a middle authorship position and is not corresponding author, suggesting a specialist contributor role — likely bringing inference or learning expertise — within larger Xiaomi-affiliated and academic groups rather than leading small, self-driven projects.

Research keywords

speculative decodingLLM inference accelerationdraft-model-free decodingearly exittree attentionbonus logitsself-evolving agentsembodied task planningskill libraryexperience learningknowledge distillationvisual saliency detection

Highlights

ECHO: dual-loop early-layer speculative decoding claiming 2.4–2.9x lossless speedup
This 2026 arXiv paper orchestrates early layers collaboratively and adds bonus logits in a draft-model-free speculative decoding scheme. Evolve judges the broad speedup results credible, while noting that the 'training-free' claim conflicts with the fine-tuning involved. Reproducibility readiness: medium.
PRACTICE: skill learner turns experience into expertise for self-evolving embodied agents
This 2026 arXiv paper evolves a frozen executor's skill library, reporting +9.7/+2.6 points over experience baselines. Evolve rates the design solid but flags internal number conflicts and the absence of variance tests; reproducibility readiness is medium.
A sharp 2026 acceleration: 13 of 14 recent works land in one year
The OpenAlex record shows 1 work in 2025 and 13 in 2026 (14 within the recent three-year window), an unusually steep ramp spanning efficient LLM inference, embodied agents, and adjacent topics such as visual saliency detection and multimodal learning.

Timeline

2025First recorded work appears in the OpenAlex profile (1 publication).
2026Output accelerates sharply: 13 works recorded in a single year, bringing the recent three-year total to 14.
2026ECHO posted to arXiv — early-layer collaborative speculative decoding claiming 2.4–2.9x lossless speedup.
2026PRACTICE posted to arXiv — skill learner for self-evolving embodied agents, +9.7/+2.6 points over experience baselines.

Papers in the Evolve catalog

Frequent co-authors

Ziyang Ma×1 Zihong Zhang×1 Zuchao Li×1 Lefei Zhang×1 Baoyuan Qi×1 Simin Yu×1 Ziyi Bai×1 Tinglei Huang×1 Börje F. Karlsson×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.

Siqi Li is a common Chinese name; this profile is name-resolved from public OpenAlex bibliographic data, without ORCID or author-disambiguation verification.
All bibliometrics (work counts, topics, coauthors, field labels) come from OpenAlex only; no publisher, grant, or institutional data is included.
Citation impact is uninformative here: h-index and i10-index are both 0 because the recorded works date to 2025–2026, so the output and impact axes are unscored.
The reproducibility score (55/100) rests on a single reviewed paper, and only 2 of the ~14 recorded works carry Evolve reviews; verdicts note internal number conflicts (PRACTICE) and a 'training-free' claim conflicting with fine-tuning (ECHO).