01
Agentic RL
Training LLM agents with reinforcement learning over long horizons — credit assignment, reward design, and learning from experience.
Research Scientist @ ByteDance Seed · Singapore
Zhanming (Allan) Jie
I build language-model agents that reason, prove, and act — trained with reinforcement learning at scale.
Currently working on agentic RL and formal theorem proving in Lean 4 as part of the Seed-Prover team. Previously at Salesforce AI Research and ByteDance AI Lab; Ph.D. from SUTD.
- citations
- 1.7k+
- h-index
- 17
- papers
- 25
- since
- 2014
A Lean 4 snippet proving the playful theorem allan_jie: there exists a language model that reasons, proves and acts — discharged by ReFT, Seed-Prover and agentic reinforcement learning.
01/research
What I work on
02
Formal theorem proving
Lean 4 provers that search deep and broad at test time and improve through large-scale RL — the Seed-Prover line of work.
03
Reasoning in language models
From deductive math-word-problem solvers to reinforced fine-tuning (ReFT): making multi-step reasoning more reliable and more explainable.
04
Efficient inference
Cheaper long-context serving, e.g. query-driven pruning of the KV cache (ThinK).
02/highlights
Selected work
all publications →ByteDance Seed · 2025
Seed-Prover 1.5
A Lean 4 prover trained with large-scale agentic RL that keeps accumulating experience from Lean and other tools, plus a test-time workflow bridging natural-language and formal proofs.
- 88%
- PutnamBench
- 11/12
- Putnam 2025, ≤ 9 h
- 80%
- Fate-H (graduate)
ByteDance Seed · IMO 2025
Seed-Prover
Lemma-style whole-proof reasoning that iteratively refines proofs from Lean feedback, with deep and broad test-time search for olympiad-level problems.
- 30/42
- IMO 2025 · silver level
- 78.1%
- past IMO problems
- 99.6%
- miniF2F-test
Salesforce AI Research · ICLR 2025
ThinK
The KV cache is redundant along the channel dimension too. ThinK prunes the least important key channels per query, cutting long-context memory while keeping accuracy.
- >20%
- KV-cache memory saved
- Spotlight
- ICLR 2025
ByteDance Research · ACL 2024
ReFT
Reinforced fine-tuning: warm up with SFT, then run PPO over many sampled chain-of-thought paths, rewarded by answer correctness — generalising better than SFT alone on math reasoning.
- 400+
- citations
- ACL
- 2024 main
03/papers
Recent papers
Measure Twice, Locate Once: Mitigating Hallucinations in LLM-based Agents for Repository-Scale Fault Localization
Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Seed-Prover 1.5: Mastering Undergraduate-Level Theorem Proving via Learning from Experience
Seed-Prover: Deep and Broad Reasoning for Automated Theorem Proving
Solving Formal Math Problems by Decomposition and Iterative Reflection
ThinK: Thinner Key Cache by Query-Driven Pruning
04/log
News
- paper
Measure Twice, Locate Once — mitigating hallucinations in LLM agents for repository-scale fault localization — published in ACM TOSEM.
- release
The Seed2.0 Model Card is out; I’m one of the contributors.
- release
Seed-Prover 1.5: agentic RL + test-time scaling solves 88% of PutnamBench and 11/12 Putnam 2025 problems within 9 hours.
- release
Seed-Prover takes part in IMO 2025: an IMO-certified 30/42, silver-medal level. Papers: Seed-Prover and Delta Prover.
- career
Joined ByteDance Seed in Singapore to work on LLM agents and reinforcement learning.
- paper
ThinK (query-driven KV-cache pruning) is a Spotlight at ICLR 2025 in Singapore.
- paper
ReFT: Reasoning with Reinforced Fine-Tuning appears at ACL 2024, alongside a Findings paper on non-autoregressive MT as a constrained HMM.
git log --all · 6 older entries
- career
Joined Salesforce AI Research, Singapore.
- paper
Leveraging Training Data in Few-Shot Prompting for Numerical Reasoning accepted to Findings of ACL 2023.
- service
Serving as a Senior Program Committee member for AAAI 2023.
- talk
Talk at the workshop on “A Science of Certified AI” (slides).
- paper
Learning to Reason Deductively accepted to ACL 2022; talk on math word problem solving at SMT, SUTD (slides).
- award
Ph.D. from the StatNLP group at SUTD — received the Best Thesis Award (thesis & slides).
05/trajectory
Where I’ve been
full CV →-
6b801fe (HEAD → main) 2025 — now
Research Scientist @ ByteDance Seed Singapore
LLM agents and reinforcement learning. Core contributor to Seed-Prover — Lean 4 theorem-proving agents trained with large-scale agentic RL (IMO 2025 silver-level score; 88% of PutnamBench).
-
15cf2b7 2024 — 2025
Research Scientist @ Salesforce AI Research Singapore
Efficient long-context inference for LLMs — ThinK, query-driven KV-cache pruning (ICLR 2025 Spotlight).
-
63ee5f5 2020 — 2024
NLP Scientist @ ByteDance AI Lab / ByteDance Research Singapore
Reasoning with language models: ReFT (reinforced fine-tuning, ACL 2024), chain-of-thought design, few-shot prompting for numerical reasoning, and visual document understanding.
-
1f9fee5 2019 — 2020
Research Intern @ Alibaba Singapore
Knowledge-graph-to-text generation (ENT-DESC, EMNLP 2020) and named entity recognition.
-
022f388 2019
Research Intern @ Allen Institute for AI (AI2) Seattle
Hosted by Pradeep Dasigi and Ana Marasović.
-
1e3af2c (tag: phd) 2016 — 2020
Ph.D., Computer Science @ Singapore University of Technology and Design (SUTD)
StatNLP group, advised by Prof. Wei Lu. Thesis: “Leveraging Dependency Trees for Structured Prediction”. ★ Best Thesis Award