Peng Lai
Southern University of Science and Technology · China
SUSTech researcher in LLM agent memory, reward modeling, and preference alignment (early-stage profile)
3
papers here
3
reviewed by Evolve
2
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—
No reliable output metrics are available: OpenAlex attributes an h-index and i10-index of 0 to this name, and no cumulative publication count is provided. This axis would measure publication volume and consistency over time; it remains unscored.
Impact—
Citation-based impact cannot be assessed: no citation counts or field-weighted influence figures are in the record, and the reported h-index of 0 reflects incomplete name attribution. This axis would measure citations and downstream uptake of the work.
Focus0/100
The focus score is 0, and the underlying distribution explains why: the declared field split is exactly even — 33.3% each for LLM agent memory systems, LLM alignment, and reward modeling. OpenAlex topics similarly span topic modeling (5), NLP techniques (3), information retrieval and search behavior (2), AI-based problem solving and planning (2), and reinforcement learning in robotics (1). Work is broad, though still clustered within LLM post-training and agent memory.
Collaboration61/100
Scoring 61, the collaboration profile rests on 20 distinct co-authors across the 6 works recorded in the last three years. The recurring list includes Guanhua Chen, Jianjie Zheng, Sijie Cheng, Jiehui Zhao, and Lei Yang, alongside Dongdong Zhang, Yun Chen, Peng Li, and Furu Wei. Peng Lai is first author on two of the three catalog papers and a middle author on ROAM, indicating active hands-on collaboration.
Activity57/100
The activity score of 57 reflects a concentrated recent burst: per the counts-by-year record, all 6 works in the three-year window land in 2026, and all three Evolve-catalog papers are 2026 arXiv preprints. This is a high-tempo but short-history signal — output is current rather than cumulative, consistent with a visible record that is just ramping up.
Reproducibility55/100
Reproducibility scores 55 (reproMean) across the 3 catalog papers, each assessed at medium readiness. Evolve's verdicts flag concrete gaps: UniRRM reports no significance tests and carries an LLM-verified data circularity risk, while AlignDiff shows strong gains on LLaMA and Qwen without variance or significance testing. ROAM's trade-offs are quantified (+29.8pp under interference, 9x call overhead), but natural-condition gains are modest.
Research summary
Peng Lai is an NLP and machine-learning researcher at Southern University of Science and Technology (SUSTech), China, whose recent work clusters around the memory and post-training side of large language model agents. Three 2026 arXiv preprints in the Evolve catalog define the current profile: UniRRM (first author) unifies reasoning reward models across 103 languages and pairwise, listwise, and point-wise judging using the 64.5K-preference MixReward dataset; AlignDiff (first author) proposes a two-stage preference-data selection method for DPO based on inverse and positive reward margins plus an NLL-gap difficulty signal; and ROAM (middle author) reorganizes agent memory into semantically related atomic units, reporting +29.8pp robustness under interference at a 9x tool-call overhead. OpenAlex topic data mirrors this spread — topic modeling (5), NLP techniques (3), information retrieval and search behavior (2), AI-based problem solving and planning (2), and reinforcement learning in robotics (1) — while the declared field distribution splits evenly across agent memory, alignment, and reward modeling. The record shows 6 works in the recent three-year window with 20 distinct co-authors. Output and impact are not yet scorable (h-index and i10-index both recorded as 0), so this is an early-stage profile.
Collaboration pattern
Across 6 works in the last three years, Peng Lai records 20 distinct co-authors. The recurring list is anchored by Guanhua Chen (listed twice), with Jianjie Zheng, Sijie Cheng, Jiehui Zhao, Lei Yang, He Zhu, and Zhiwen Ruan, plus Dongdong Zhang, Yun Chen, Peng Li, and Furu Wei. First author on two of three catalog papers and middle author on the third, suggesting a hands-on contributor embedded in a multi-group NLP network rather than a solo researcher.
Research keywords
Highlights
Unified reasoning reward models across 103 languages
Two-stage preference data selection for DPO
Robust atomic memory organization for LLM agents
Timeline
- 2026First-author arXiv preprint UniRRM unifies reasoning reward models across languages and evaluation paradigms.
- 2026First-author arXiv preprint AlignDiff introduces model-intrinsic two-stage selection of DPO preference data.
- 2026Co-authors ROAM on robust organization of atomic memories for LLM agents, with quantified robustness gains.
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 signals come from OpenAlex name-based attribution only; no ORCID is linked, so 'Peng Lai' is name-resolved from public bibliographic data.
- h-index and i10-index are recorded as 0 and no citation counts are available, so output and impact are unscored rather than genuinely low.
- All three catalog papers are 2026 arXiv preprints; no peer-reviewed venue, acceptance, or award information is in scope.
- The frequent-coauthor list contains a duplicated entry (Guanhua Chen appears twice), so named-collaboration counts are indicative only.