Researcher at the Chinese Academy of Sciences working on audio few-shot learning and LLM multi-agent social simulation audits.
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-14) 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 data is not yet available: the record has no verified publication count, venue list, or yearly output volume. This axis would measure how much Yi Li publishes, in which venues, and how that volume evolves over time.
Impact—
Impact data is not yet available: no citation counts, h-index, or downstream adoption signals exist in this record. This axis would measure how often and how widely the community cites and builds on Yi Li's work.
Focus0/100
The focus score is 0, and the underlying field distribution explains why: Yi Li's reviewed output splits evenly — 50% machine learning for audio signal processing and 50% AI evaluation of LLM multi-agent social simulation. Two 2026 papers sit in these separate areas, one auditing theory-of-mind behavior in negotiation agents and the other proposing channel selection for audio few-shot learning, so the record currently reads as two parallel research threads rather than one concentrated specialty.
Collaboration50/100
The collaboration score is 50 (of 100), based on 11 distinct coauthors in the record. Recurring names include Fengrui Liu, Ningxin Shen, Yiwei Fu, Feng Liu, Jiangmeng Li, Cong Li, Cheng Chen, Thomas Fung, and Alex Rossi. Authorship positions — last author on the 2026 theory-of-mind audit and middle author on the audio few-shot learning paper, with neither paper marked corresponding — suggest a collaborative contributor role rather than a consistently leading one.
Activity—
Activity data is not yet available: the record lacks a sufficient publication time series to score recent research velocity. This axis would measure how actively Yi Li is currently publishing, based on output per recent year.
Reproducibility78/100
Reproducibility is the strongest recorded axis: the mean score across the two reviewed papers is 78 (of 100), computed over 2 papers. The audio few-shot learning paper is rated high reproducibility readiness, with its 4.9–8.4 point gain measured against a matched control, though Evolve notes it is single-benchmark with no confidence intervals. The theory-of-mind audit is rated medium and carries a single-reader caveat, so the aggregate rests on only two arXiv preprints.
Research summary
Yi Li is a researcher at the Chinese Academy of Sciences whose reviewed work spans two areas in roughly equal measure — machine learning for audio signal processing and AI evaluation of LLM-based multi-agent systems (a 50/50 field distribution in the underlying bibliometric data). The Evolve catalog includes two 2026 arXiv papers. In “Mind or Message? Auditing Theory of Mind in Multi-Agent Social Simulation” (last author), the audit of multi-issue negotiation between LLM agents concludes that the agents track messages rather than minds: their utterances are fluent but driven by egocentric projection, and they remain frontier-blind to the Pareto frontier of achievable agreements; Evolve judged the audit sound while flagging a single-reader caveat. In “Sample-Conditioned Representation Selection for Audio Few-Shot Learning” (middle author), a Gumbel Top-k per-input channel selection mechanism lifts SpurAudio out-of-distribution few-shot audio accuracy by 4.9–8.4 points over a matched control — a solid result, though demonstrated on a single benchmark without confidence intervals. Across the two reviewed papers the reproducibility score averages 78 (of 100), with the audio paper rated high reproducibility readiness and the simulation audit medium. Collaboration data records 11 distinct coauthors and a collaboration score of 50, while output, impact, and activity metrics are not yet available in the data. This is an early-stage, name-resolved profile built from public bibliographic data (ORCID 0000-0003-4562-8208).
Collaboration pattern
Yi Li's reviewed papers record 11 distinct coauthors and a collaboration score of 50 (of 100). Recurring co-authors include Fengrui Liu, Ningxin Shen, Yiwei Fu, Feng Liu, Jiangmeng Li, Cong Li, Cheng Chen, Thomas Fung, and Alex Rossi, and the pool spans both research threads (audio machine learning and LLM multi-agent evaluation). Neither 2026 paper lists Yi Li as corresponding author — one last-author slot and one middle-author slot — consistent with contributing within larger teams rather than leading small groups.
Research keywords
Highlights
Auditing whether LLM negotiators track minds or just messages
Gumbel Top-k channel selection lifts OOD audio few-shot accuracy
Timeline
- 2026Published “Sample-Conditioned Representation Selection for Audio Few-Shot Learning” on arXiv (middle author), introducing Gumbel Top-k per-input channel selection for out-of-distribution few-shot audio classification.
- 2026Published “Mind or Message? Auditing Theory of Mind in Multi-Agent Social Simulation” on arXiv (last author), auditing whether LLM negotiators track minds or merely messages in multi-issue negotiation.
- 2026Both 2026 papers entered the Evolve review catalog, with a combined reproducibility mean of 78 (of 100) across the two papers and 11 distinct coauthors recorded in the underlying data.
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.
- All bibliometric scores come from a public bibliometric snapshot tied to ORCID 0000-0003-4562-8208; “Yi Li” is a common Chinese name, so identity resolution relies on that ID rather than on verified biography.
- Output, impact, and activity dimensions are not scored because the underlying data (publication counts, citations, recent cadence) is insufficient.
- Coverage is limited to two 2026 arXiv preprints reviewed by Evolve; journal papers, other preprints, and non-English output may be missing.
- arXiv preprints are not necessarily peer-reviewed, and Evolve verdicts are independent reviews rather than venue acceptance decisions.