Lingbo Wang

Lingbo Wang

M.S. student at Soochow University · VLM · RL · Agent Memory

About Me

I am Lingbo Wang, a master’s student in the Natural Language Processing Lab at Soochow University, advised by Prof. Yu Hong. I received my B.Eng. in Intelligent Science and Technology from Qilu University of Technology.

My work focuses on vision-language model (VLM) pre-training and post-training, reinforcement learning (MPO, GSPO, ROPD, and Agentic RL), and memory for long-horizon LLM agents. I have experience building and training large-scale multimodal models with Megatron, DeepSpeed, VeRL, and vLLM.

Download my résumé (PDF) · GitHub · Technical Blog

🔥 News

  • 2026: EffectMem: Effectiveness-Driven Memory Reuse for Long-Horizon LLM Agents was accepted by EMNLP 2026.
  • 2026: Anti-Exception Action Generation for Automatic Planning was accepted by ICASSP 2026.
  • 2025.09 - 2026.03: Joined Baidu’s Application Model Unit as a research intern, working on VLM pre-training, post-training, and reinforcement learning.

📝 Publications

  1. EffectMem: Effectiveness-Driven Memory Reuse for Long-Horizon LLM Agents
    Jingjing Yin*, Lingbo Wang*, Dongyan Cao, Yu Hong. EMNLP 2026, CCF-B.
    * Equal contribution.
    We introduce an effectiveness-driven memory reuse framework that represents successful trajectories as executable contracts and validates their applicability before reuse. EffectMem outperforms ReAct, ExpeL, AutoGuide, and CFGM on ALFWorld, ScienceWorld, and WebShop.

  2. Anti-Exception Action Generation for Automatic Planning
    Lingbo Wang, Dongyan Cao*, Yu Hong*. ICASSP 2026, CCF-B.
    * Corresponding author.

  3. RubricFlow-Mem: Rubric-Guided Self-Evolving Curriculum for Long-Term Memory Agents
    Lingbo Wang (first author). Under review at ICLR 2027.

💻 Experience

Baidu · Application Model Unit (AMU), Multimodal Perception and Understanding

Research Intern · 2025.09 - 2026.03

Wenxin-Lightning VLM foundation model

  • Contributed to data construction, architecture adaptation, and multi-stage pre-training/post-training of a multimodal MoE model for AI Search using Megatron.
  • Built large-scale MapReduce data-cleaning and quality-evaluation pipelines; adapted InternViT/RiceViT and Lightning LLM/Qwen3 LLM; implemented bidirectional Megatron-Hugging Face checkpoint conversion.
  • Helped establish an H200 thousand-GPU-scale training pipeline and an EvalScope evaluation workflow, then iterated SFT data mixtures and training strategies based on benchmark results.
  • Constructed MPO preference data and conducted MPO post-training with DeepSpeed to improve mathematical reasoning.
  • The resulting model reached 105.19% of Qwen3-VL-30B-A3B’s aggregate score on general VLM benchmarks and 124.11% in STEM, while reducing average chain-of-thought length by 18.7% and improving inference speed by 23.1%.

Wenxin Assistant photo-based K-12 problem solving

  • Built a data flywheel and MPO preference samples from real user queries; adapted the 30B-A3B MoE model in VeRL and established a hundred-GPU GSPO training pipeline.
  • Developed offline evaluation to refine rewards and constraints and mitigate reward hacking; optimized LLM-as-a-Judge scheduling and vLLM image processing to reduce training bubbles, visual tokens, and inference cost.
  • Improved offline accuracy from 79% for the online baseline to 81% with MPO and 85% with GSPO, matching Seed1.6. After launch, page views increased by 1.37% and search sessions by 1.12%.

🧩 Selected Project

World-Model-Driven Lookahead Planning for LLM Agents

Agentic RL open-source project · 2025.06 - 2025.09

  • Identified and fixed an Agentic RL training bug in the VeRL main repository; the pull request was merged upstream.
  • Extended verl-agent with a world-model-driven lookahead planner. At each rollout state, the planner samples candidate actions, predicts their next observations in batches, scores predicted states with rule-based or learned rewards, and executes the action with the highest expected return.
  • On ALFWorld with Qwen2.5-1.5B Base, the success rate reached 50.78% after 40 co-training steps and peaked at 93.84% after convergence. World-model loss fell from 0.824 to 0.155, while average task steps decreased by 70%.

🎓 Education

  • Soochow University, M.S. in Computer Science and Technology, Natural Language Processing Lab
    2024.09 - Present
  • Qilu University of Technology, B.Eng. in Intelligent Science and Technology, School of Mathematics and Statistics
    2020.09 - 2024.07

🛠 Skills

  • Research: VLM pre-training and post-training, MPO, GSPO, ROPD, Agentic RL, agent memory, automatic planning
  • Training and inference: Megatron, DeepSpeed, VeRL, vLLM, EvalScope
  • Model architectures: multimodal MoE, InternViT, RiceViT, Lightning LLM, Qwen3 LLM
  • Data and infrastructure: MapReduce, large-scale distributed training, SFT and preference-data construction, LLM-as-a-Judge