Jie Wu

Research profile

Research summary

Jie Wu, affiliated with the China Meteorological Administration, is a machine learning researcher whose cataloged work spans three roughly equal areas: AI for weather forecasting (scientific machine learning), machine learning for code and LLM agents, and LLM post-training with reinforcement learning. The Evolve catalog holds three 2026 arXiv preprints. PCSDiff applies diffusion-based bias correction and super-resolution/downscaling to operational medium-term precipitation forecasting over China, with solid gains that Evolve notes fall short of 'consistently outperforms' on 10-day FSS and lack significance tests. Terminal-Universe, where Wu is first author, converts agent trajectories into scalable terminal training environments; ablations support the reported +11.9 TB2.1 gain, though teacher-authored verifiers and no code release warrant caution. The third paper proposes gradient-aligned (expert-CoT cosine) rewards for RLVR math reasoning with GRPO, reporting pass@k gains at under 9% overhead, small on tiny benchmarks. Bibliometrics: 32 distinct coauthors, a collaboration score of 70, a reproducibility mean of 43 over three papers, and an even three-way field split that drives a focus score of 0. Output, impact, and activity are not yet scored.

Research keywords

Highlights

• Diffusion bias correction and super resolution for operational 10-day precipitation forecasts

  PCSDiff applies a diffusion model (DDIM) to bias-correct and downscale medium-term precipitation over China, targeting practical operational use. Evolve finds solid gains but notes the 'consistently outperforms' claim fails on 10-day FSS and that no significance tests are reported; reproducibility readiness is medium.

• First-authored environment synthesis: agent trajectories become terminal training environments

  Wu is first author of Terminal-Universe, which replays agent trajectories into scalable terminal environments for agentic training and task synthesis. Ablations support the +11.9 TB2.1 gain, though Evolve cautions that verifiers are teacher-authored and no code is released, giving low reproducibility readiness.

• Gradient-aligned rewards for RLVR math reasoning at low overhead

  A middle-author paper shapes rewards by expert-CoT gradient cosine to unlock GRPO-based RLVR training, motivated via neural tangent kernel analysis and reporting pass@k gains at under 9% overhead. Evolve judges it well-motivated with theory but finds the gains small on tiny benchmarks; readiness is medium.

Timeline

1. 2026Publishes 'Gradients Know What Outcomes Don't' on arXiv (middle author), proposing gradient-aligned rewards for GRPO-based RLVR math reasoning.
2. 2026Publishes 'Terminal-Universe' on arXiv as first author, turning agent trajectories into scalable terminal training environments.
3. 2026Publishes 'PCSDiff' on arXiv (middle author), applying diffusion-based bias correction and super resolution to operational medium-term precipitation forecasting.

Papers in the Evolve catalog

• 2026-09-07

  PCSDiff: Diffusion-Based Bias Correction and Super Resolution Toward Practical Operational Medium-Term Precipitation Forecast

• 2026-09-03first author

  Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments

• 2026-09-03

  Gradients Know What Outcomes Don't: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned Rewards

Frequent co-authors