Zhengyuan Liu

Research profile

Research summary

Zhengyuan Liu is a researcher at the Agency for Science, Technology and Research (A*STAR) in Singapore whose work spans natural language processing, large language model (LLM) post-training, and multimodal AI agents. The OpenAlex bibliometric snapshot used here is heavily weighted to 2025–2026: it records 1 work in 2025 and 21 in 2026, with 22 works in the last three years and 40 distinct coauthors — a recent, rapid output acceleration rather than a long publication record. Topic annotations concentrate on Topic Modeling (15 mentions) and Natural Language Processing Techniques (10), followed by Domain Adaptation and Few-Shot Learning and Multimodal Machine Learning Applications (6 each) and AI-based Problem Solving and Planning (5), and the profile scores 100/100 on research focus, consistent with tight concentration in ML/NLP. Three 2026 arXiv papers appear in the Evolve catalog: a scaling study of near-optimal SFT–RL annotation budget allocation across model sizes, MNIST-PRO, which recasts MNIST as a glimpse-based partially observable world for agents, and CultureConverse, a multilingual multi-turn simulation harness for culturally grounded assistance in East and Southeast Asia. The cataloged work is empirical and benchmark-oriented; reviewed papers carry 'high' reproducibility ratings, and the reproducibility dimension scores 100 on that basis.

Research keywords

Highlights

• Scaling analysis shows near-optimal SFT–RL annotation budget splits transfer across model sizes

  Evolve reviewed this 2026 arXiv study as a credible empirical result: the near-optimal allocation regions between supervised fine-tuning and reinforcement learning are wide and transfer from small to larger LLMs, though key numbers rest on single seeds and experiments cap at 14B parameters. Reproducibility readiness: high.

• MNIST-PRO recasts MNIST as a partially observable world for AI agents

  The 2026 benchmark turns the classic dataset into glimpse-based partially observable search for multimodal LLM agents, probing active perception and working memory. Evolve finds the findings coherent and diagnostic, but notes the n=100 cells lack variance estimates. Reproducibility readiness: high.

• CultureConverse: a multilingual simulation harness for culturally grounded assistance

  This 2026 harness simulates multi-turn, multilingual assistance grounded in East and Southeast Asian cultural contexts. It is listed in the Evolve catalog but has not yet been reviewed — no verdict, keywords, or reproducibility rating are available.

Timeline

1. 2025First indexed work appears in the OpenAlex snapshot (1 work) — the earliest point of the countsByYear series.
2. 2026Indexed output rises to 21 works; the last-three-year total reaches 22, with 40 distinct coauthors recorded.
3. 2026arXiv release of the SFT–RL annotation budget allocation scaling study (middle author position, high reproducibility rating).
4. 2026MNIST-PRO appears on arXiv, recasting MNIST as a glimpse-based partially observable world for multimodal AI agents.
5. 2026CultureConverse, a multilingual multi-turn simulation harness for culturally grounded assistance in East and Southeast Asia, appears on arXiv.

Papers in the Evolve catalog

• 2026-09-01

  Scaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs

• 2026-08-31

  MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents

Frequent co-authors