Xiaomi researcher advancing LLM post-training: RL credit assignment, teacher-gated distillation, search agents
3
papers here
3
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
0
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—
Output metrics are not yet scored: the OpenAlex-linked record has no established career-level publication profile beyond the 2026 activity cluster, so cumulative output cannot be computed. This axis would measure total works and annual productivity.
Impact—
Impact cannot be scored yet: citation-based signals (h-index and i10-index recorded as 0) have not accumulated, and all catalogued papers are 2026 arXiv preprints. This axis would measure citations and influence on follow-up work.
Focus0/100
Focus scores 0 because the field distribution splits evenly across three adjacent sub-fields — reinforcement learning for LLM post-training, LLM post-training, and NLP/LLM agents (about 33% each). The dispersion lowers the score, yet the strands converge on one theme: improving post-trained LLMs. Topic data (Reinforcement Learning in Robotics 7, Topic Modeling 6, Domain Adaptation and Few-Shot Learning 5) also shows breadth beyond LLM post-training.
Collaboration62/100
Collaboration scores 62: the record lists 22 distinct coauthors, with a recurring core of 12 frequent collaborators (e.g., Zhaokai Luo, Yao Hu, Dongyan Zhao, Yong Zhang). On all three catalogued papers Mu Chuan is the last author but not corresponding — a pattern consistent with senior or supervisory contribution within team-based industrial research.
Activity70/100
Activity scores 70, driven by a concentrated recent burst: OpenAlex logs 10 works in 2026 (recent-3-year works: 10), and all three Evolve-reviewed papers appeared on arXiv in 2026. The record currently shows no works in earlier years, so activity rests almost entirely on this single-year cluster.
Reproducibility43/100
Reproducibility averages 43 across the 3 papers assessed (reproMean 43, reproPapers 3). Readiness is medium for PACT and Iris, low for Verify Before You Distill. Evolve's verdicts converge on one weakness: single-run evaluations with no error bars or significance tests, which caps confidence in otherwise strong reported gains.
Research summary
Mu Chuan is a researcher at Xiaomi (China) whose recent work concentrates on the post-training of large language models, spanning reinforcement-learning-based credit assignment, on-policy distillation, and LLM-driven search agents. The OpenAlex-linked record shows a burst of recent productivity: 10 works in 2026 (activity score 70) produced with 22 distinct coauthors (collaboration score 62). Evolve has reviewed three 2026 arXiv papers, on all of which Mu Chuan is the last author (non-corresponding) — a position typically associated with senior or supervisory contributions. PACT derives token-level credit assignment for actor-critic LLM post-training from three axioms, unifying PPO/GAE- and GRPO/RLOO-style objectives; Iris reports open-source search agents reaching BrowseComp 82.2/88.6 via synthetic multi-hop data and SFT-RL training; Verify Before You Distill introduces prompt-level teacher gating for on-policy distillation using verifier-scored probes and reverse KL objectives. The field distribution splits evenly (about 33% each) across reinforcement learning for LLM post-training, LLM post-training more broadly, and NLP/LLM agents, yielding a low focus score of 0 — though the three strands share one theme: improving post-trained LLMs. Reproducibility averages 43 across the three assessed papers (readiness: medium, medium, low), and Evolve's verdicts repeatedly flag single-run evaluations without error bars or significance tests as the main caveat.
Collaboration pattern
Mu Chuan works in sizeable teams: 22 distinct coauthors appear in the record, with 12 recurring collaborators including Zhaokai Luo, Yao Hu, Jiayan Fu, Hang Xu, Yong Zhang, and Dongyan Zhao. On every catalogued paper the position is last author without corresponding-author status, suggesting a senior contributor or advisor role within Xiaomi's LLM post-training efforts rather than day-to-day lead authorship.
Research keywords
Highlights
PACT grounds token-level credit assignment in three axioms
Iris claims open-source state-of-the-art search agents on BrowseComp
Verify Before You Distill gates teacher supervision per prompt
Timeline
- 2026Publishes PACT on axiom-grounded token-level credit assignment for actor-critic LLM post-training (last author, arXiv).
- 2026Releases Iris, claiming open-source state-of-the-art search agents on BrowseComp via synthetic multi-hop data and SFT-RL (last author, arXiv).
- 2026Introduces prompt-level teacher gating for on-policy distillation in Verify Before You Distill (last author, arXiv).
- 2026OpenAlex logs 10 works in this single year, the only publication year currently present in the record.
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.
- Bibliometric scores derive from OpenAlex only; no ORCID is linked, so identity is resolved by name from public bibliographic data.
- All catalogued papers are 2026 arXiv preprints; citation-based impact has not had time to accumulate (h-index and i10-index are recorded as 0).
- The reproducibility score (43) covers only the 3 Evolve-assessed papers, not the full output.
- "Mu Chuan" is a common Chinese name; the profile may merge or miss records, and the OpenAlex record shows no pre-2026 works.