Fudan University RL researcher: policy distillation and critic learning for LLM post-training
2
papers here
1
reviewed by Evolve
0
first-author
0
corresponding
—
works · OpenAlex
10
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-09-28) 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 scoring is unavailable in this snapshot: no aggregated works or venue-normalized productivity composite is on record, and only a thin h-index of 10 (i10 index 10) with per-year counts exists — too sparse to score this axis reliably.
Impact—
Impact cannot be scored yet: field-normalized citation aggregates are missing from the snapshot. With an h-index of 10, an i10-index of 10, and no per-paper citation counts on record, influence relative to field peers remains unassessed.
Focus100/100
Scores 100 — the maximum. The reviewed output is tightly clustered: both 2026 papers sit in machine learning (reinforcement learning for large language models), matching the fieldDist entry. OpenAlex topics add battery and supercapacitor materials lines (topic counts 7, 5, 4) alongside RL in robotics (3) and multimodal ML (3), so the index blends research lines under one name.
Collaboration74/100
Scores 74. The record lists 12 works with 39 distinct coauthors over the recent three years, plus 12 recurring collaborators including Shuai Dong, Wenqi Shao, Congcong Wang, and Siyuan Wang. Both Evolve-reviewed 2026 papers place Tian in a middle author position without corresponding-author role, consistent with contributing inside larger machine-learning teams rather than leading them.
Activity75/100
Scores 75. Recorded works per year run 6 (2021), 2 (2022), 2 (2023), 1 (2024), 5 (2025), and 6 (2026), with 12 works inside the recent three-year window. After a 2022-2024 trough that bottomed out at one work, 2025-2026 output has rebounded to its highest recorded level, an active and accelerating phase.
Reproducibility55/100
Scores 55 (reproMean over 1 assessed paper). Both Evolve-reviewed 2026 arXiv papers carry medium reproducibility readiness: verdicts praise strong diagnostics and ablations, but flag single-seed runs — and terminal-reward-only evaluation in the PPO work — as limiting confidence in small reported margins. Independent replication should treat fine-grained deltas cautiously.
Research summary
Kanghui Tian is a researcher at Fudan University whose Evolve-reviewed work centers on reinforcement learning for large language model post-training. Evolve has reviewed two 2026 arXiv papers on which Tian is a middle author (non-corresponding): one showing that on-policy distillation preparation yields better RL starts than Base or SFT initialization, with the reverse/forward KL ranking flipping after RL, and one identifying "value flattening" in PPO critics, mitigated by SP3O's three-anchor sparse supervision. Both carry medium reproducibility readiness — strong diagnostics and ablations, but single-seed evidence. The bibliometric record (OpenAlex) lists an h-index of 10 and an i10-index of 10, 12 works over the recent three years, and 39 distinct coauthors; yearly output moved from 6 works in 2021 down to 1 in 2024, then rebounded to 5 in 2025 and 6 in 2026. Notably, the OpenAlex topic profile for this name is dominated by battery and supercapacitor materials topics alongside reinforcement learning and multimodal machine learning, so this profile is name-resolved from public bibliographic data and the research lines under this name may not fully overlap. Within RL for LLMs, the record so far is that of an early-stage contributor embedded in larger teams, with output concentrated in 2025-2026.
Collaboration pattern
The OpenAlex record lists 39 distinct coauthors over the last three years across 12 works, with 12 recurring collaborators including Shuai Dong, Wenqi Shao, Congcong Wang, Siyuan Wang, and Ji Zhao. Both Evolve-reviewed 2026 papers place Tian as a middle author without corresponding-author status, indicating a contributor role within larger teams rather than independent group leadership. Collaboration dimension scores 74.
Research keywords
Highlights
On-policy distillation as a better starting point for RL
Diagnosing "value flattening" in PPO critics
Timeline
- 2021The bibliometric record opens with 6 works — the highest single-year count until 2026.
- 2022Output cools to 2 works, beginning a multi-year lower-output phase.
- 2024The trough of the record: only 1 work logged this year.
- 2025Rebound to 5 works, signaling renewed research activity.
- 20266 works to date, including both Evolve-reviewed arXiv papers: on-policy distillation for RL starts, and value flattening in PPO critics.
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 figures come from OpenAlex only; no citation-based output or impact composite is available, so those two radar axes are left unscored.
- This profile is name-resolved from public bibliographic data: the OpenAlex record for this name mixes battery/supercapacitor materials topics with the RL/LLM papers reviewed here, so work attribution should be treated with care.
- Reproducibility scoring covers only 1 assessed paper (reproMean 55), and Evolve verdicts flag single-seed runs in both reviewed papers.
- Both reviewed papers are 2026 arXiv preprints with middle-author, non-corresponding positions; peer-reviewed venue status is not on record.