Yiming Li

Nanyang Technological University · Singapore

Unclaimed · AI-structured profileOpenAlex

NTU Singapore researcher in LLM post-training systems, world-model security, and vision-language-action robot learning

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—

No output score yet: only 4 works appear in the recent three-year window (all indexed in 2026), with h- and i10-index recorded as 0. This axis would measure cumulative publication volume, venue mix, and yearly output growth as coverage improves.

Impact—

Impact is not yet scorable: the OpenAlex snapshot shows h- and i10-index of 0 with no citation accumulation, likely reflecting thin coverage for a common name. The axis would measure citations per paper, citation velocity, and uptake by follow-up work.

Focus0/100

Focus scores 0: the field distribution splits evenly three ways — LLM training infrastructure, world-model backdoor security, and vision-language-action robot learning each hold one-third weight — and OpenAlex topics span five clusters. This is a broad systems-to-security-to-robotics portfolio rather than a single-track agenda.

Collaboration68/100

Collaboration scores 68, grounded in 30 distinct coauthors within the recent window. The frequent list mixes individual researchers (Tao Lin, Yusheng Su, Kangrui Du, Mao Cheng, Yanbin Jiang) with an entity-style entry (RadixArk). Across all three reviewed papers the researcher sits in a middle authorship position without corresponding role, indicating a strong contributor inside larger teams.

Activity47/100

Activity scores 47: the recent three-year window contains 4 works, and countsByYear shows a single active year, 2026. On the current record the publication cadence is nascent; a one-year signal may partly reflect indexing coverage for a common name rather than actual output rhythm.

Reproducibility55/100

Reproducibility scores 55 as the mean across 3 reviewed papers, each rated medium readiness. Evolve's verdicts flag recurring gaps: Miles v0.1 rests on a single self-run case study without baselines; TrojanWorld has broad simulation evidence but a single target action and thin defense test; LayerRoute reaches +7.2pp on LIBERO-Long with small margins lacking variance statistics.

Research summary

Yiming Li is a machine-learning researcher at Nanyang Technological University, Singapore, whose OpenAlex-resolved record centers on reinforcement learning in robotics (4 topic mentions), with further mentions in robot manipulation and learning (2), natural language processing techniques, robotic locomotion and control, and security and verification in computing. The Evolve-reviewed catalog holds three 2026 arXiv papers, each with a middle authorship position and none as corresponding author, spread evenly across three fields: LLM training infrastructure (Miles v0.1, a production-level RL post-training system built around SGLang, mixture-of-experts, and fully asynchronous RL with weight synchronization), AI security (TrojanWorld, backdooring world-model agents via imagination steering), and robot learning (LayerRoute, action-conditioned mixture-of-layers routing for vision-language-action policies with reported gains up to +7.2pp on LIBERO-Long). Bibliometric coverage is thin: 4 works in the recent three-year window, all indexed in 2026, 30 distinct coauthors, and no citation-impact scores yet (h- and i10-index recorded as 0). Evolve rates reproducibility at a mean of 55 across the 3 reviewed papers (all medium readiness), with recurring evaluation caveats: a single self-run case study without baselines, small evaluations, and margins lacking variance statistics.

Collaboration pattern

The recent record lists 30 distinct coauthors, with recurring names including Tao Lin, Yusheng Su, Kangrui Du, Mao Cheng, Shi Dong, and Yanbin Jiang, plus an entity-style entry (RadixArk) consistent with the systems-oriented Miles v0.1 work. Across all three Evolve-reviewed papers, Yiming Li appears in a middle authorship position and never as corresponding author, indicating collaboration within larger engineering or lab teams rather than independent group leadership; the collaboration score of 68 reflects breadth of contacts rather than visible seniority.

Research keywords

reinforcement learning post-trainingLLM training systemsmixture-of-expertsfully asynchronous RLreinforcement learning in roboticsvision-language-action policieslayer routingrobot manipulationworld-model backdoor attacksimagination steeringLIBERO benchmark

Highlights

Miles v0.1: a production-level RL post-training stack
An arXiv systems report covering fully asynchronous RL, mixture-of-experts support, SGLang serving, and weight synchronization. Evolve calls it unusually honest in self-reporting, but the evidence rests on a single self-run case study without baselines.
TrojanWorld: backdooring world-model agents via imagination steering
The paper claims a world-model backdoor driven by a physical trigger that persistently steers agent behavior. Evolve notes broad simulation evidence, but only a single target action, small evaluations, and a thin defense test.
LayerRoute: action-conditioned mixture-of-layers routing for VLAs
Adds action-conditioned VLM-layer routing and an action-state reread mechanism to vision-language-action policies, with multi-benchmark gains up to +7.2pp on LIBERO-Long — though Evolve flags that small margins lack variance statistics.

Timeline

2026Posts LayerRoute on arXiv: action-conditioned mixture-of-layers routing for vision-language-action policies, with reported gains up to +7.2pp on LIBERO-Long.
2026Posts TrojanWorld on arXiv: a backdoor attack steering world-model agents via imagination.
2026Contributes to Miles v0.1 on arXiv: a production-level RL post-training system with fully asynchronous training, SGLang serving, and weight synchronization.
2026OpenAlex indexes 4 works for this profile in the year, with 30 distinct coauthors across the recent window.

Papers in the Evolve catalog

Frequent co-authors

RadixArk×1 :×1 Tom Chen×1 Mao Cheng×1 Shi Dong×1 Kangrui Du×1 Yanbin Jiang×1 Jiajun Li×1 Tao Lin×1 Yusheng Su×1 Andy Ye×1 Yueming Yuan×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.

Profile is name-resolved from public OpenAlex bibliographic data; 'Yiming Li' is a common name and no ORCID is available, so coverage may mix or omit same-name authors.
Bibliometrics derive from the OpenAlex snapshot only; citation-impact scores are unavailable (h-index and i10-index recorded as 0).
Paper-level judgments rest on the 3 Evolve-reviewed 2026 arXiv preprints; peer-review venue status is unknown.
Authorship roles (middle position, non-corresponding) come from Evolve review metadata and may not reflect contributions on uncatalogued work.