Early-career researcher in LLM-based agents and multimodal post-training, with closed-loop agentic RL pipelines
2
papers here
2
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Output is unscored: the snapshot lacks sufficient publication-volume data. This axis would measure total works, venue quality, and output cadence over time; currently only 2 OpenAlex-recorded works exist, both 2026 preprints.
Impact—
Impact is unscored: citation data is insufficient. This axis would measure citations, h-index trajectory, and community uptake; the snapshot records h-index 0 and i10-index 0 on works too recent (2026) to have accumulated citations.
Focus0/100
The focus score is 0, and the raw field split explains why: the snapshot divides this profile evenly, 50% artificial intelligence (multimodal large language models) and 50% machine learning — LLM-based agents, while OpenAlex topics scatter across five clusters (reinforcement learning in robotics with count 2, topic modeling, problem solving and planning, games, multimodal applications). With only 2 recorded works, concentration has not yet stabilized.
Collaboration76/100
Collaboration scores 76, the strongest axis in the snapshot: OpenAlex records 45 distinct coauthors, and 12 recurring collaborators appear in the Evolve data (including Heyun Chen, Qi She, Fei Yu, and Yiheng Lin). Both catalog papers place Sihan Yang in a middle authorship position without corresponding-author role, a pattern consistent with contributing inside a large, engineering-heavy team rather than leading solo projects.
Activity32/100
Activity scores 32. The OpenAlex countsByYear series contains a single entry — 2 works in 2026 — and recent3yWorks is likewise 2, both arXiv preprints (Qwen-Planner-Agent and the Pistis Technical Report). This indicates a very recent, low-volume record: research output has only just begun to register in bibliographic databases, so the activity level remains modest.
Reproducibility20/100
Reproducibility scores 20 (reproMean 20 across 2 reviewed papers), matching Evolve's 'low' repro-readiness ratings on both. Qwen-Planner-Agent reports 77.05% on the self-built MobilePA-Bench but presents gains without significance tests; the Pistis Technical Report releases no artifacts. The current evidence therefore does not straightforwardly support independent replication.
Research summary
Sihan Yang is an early-stage researcher working at the intersection of LLM-based agents and multimodal large language models — the OpenAlex field split is exactly 50/50 between 'artificial intelligence (multimodal large language models)' and 'machine learning — LLM-based agents'. The profile is name-resolved from public bibliographic data (OpenAlex identity), with no institution, country, or ORCID attached. Two 2026 arXiv preprints are reviewed in the Evolve catalog, with Sihan Yang in middle authorship positions on both: Qwen-Planner-Agent, a closed-loop AI-for-AI framework that trains a 27B mobile planner via agentic reinforcement learning (GRPO) and tops MobilePA-Bench at 77.05%, and the Pistis Technical Report, which stabilizes multimodal post-training through interleaved on-policy distillation and reinforcement learning with verifiable rewards. OpenAlex topics — reinforcement learning in robotics (2), topic modeling, AI-based problem solving and planning, AI in games, and multimodal applications — reinforce an agentic-RL orientation. The collaboration dimension is the strongest (76): 45 distinct coauthors and 12 recurring collaborators. Citation metrics (h-index 0, i10-index 0) and low repro-readiness ratings on both papers (reproMean 20) mark this as an emerging profile whose impact and output are not yet scoreable.
Collaboration pattern
Collaboration is the strongest dimension (76): OpenAlex records 45 distinct coauthors, with 12 recurring collaborators in the Evolve data — including Heyun Chen, Xiaohan Lan, Jiaxi Li, Qi She, Fei Yu, and Yiheng Lin. On both catalog papers Sihan Yang sits in a middle authorship position without the corresponding-author role, consistent with contributing within a sizable research or engineering team rather than independently leading projects.
Research keywords
Highlights
Closed-loop AI-for-AI training lifts a 27B mobile planner to 77.05% on MobilePA-Bench
Interleaved on-policy distillation and RL stabilize multimodal post-training
A 50/50 portfolio: LLM-based agents and multimodal LLM post-training
OpenAlex splits this profile evenly (50% artificial intelligence — multimodal large language models; 50% machine learning — LLM-based agents), and five topics — reinforcement learning in robotics (2), topic modeling, AI-based problem solving and planning, AI in games, and multimodal applications — point to a reinforcement-learning-centered agentic agenda shared by both 2026 papers.
Timeline
- 2026Pistis Technical Report posted on arXiv: interleaved on-policy distillation and reinforcement learning with verifiable rewards stabilize multimodal post-training over Qwen bases, with search gains but no released artifacts.
- 2026Qwen-Planner-Agent posted on arXiv: a closed-loop AI-for-AI framework trains a 27B mobile planner via GRPO-based agentic RL, reaching 77.05% on MobilePA-Bench.
- 2026OpenAlex snapshot records this profile's only 2 works to date, both in 2026, with h-index 0 and i10-index 0 — an emerging, name-resolved profile.
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.
- Bibliometrics come from OpenAlex only; no institution, country, or ORCID is available, and this common-name profile is name-resolved from public bibliographic data.
- Citation impact cannot be scored: h-index and i10-index are recorded as 0, and both works are 2026 arXiv preprints too recent to have accumulated citations.
- The profile rests on only 2 Evolve-reviewed papers; the 'output' and 'impact' radar axes are marked missing rather than estimated.
- Reproducibility (reproMean 20 over 2 papers) reflects Evolve's review assessments, not independent replication attempts.