v3.0.0 - Part III Trilogy: Agentic-Jev-GRPO, CoEvo-Jev-GRPO & Tree-Jev-GRPO (36p EN + 23p CN)

Official v3.0.0 Release of the Part III LLM Reinforcement Learning Trilogy on Qwen/Qwen3.5-4B + Kev-4B: (1) Part III-A: Agentic-Jev-GRPO (12p EN + 8p CN PDF, 78.23% 4-Benchmark Agentic SOTA), (2) Part III-B: CoEvo-Jev-GRPO (12p EN + 8p CN PDF, 59.28% Olympiad/AIME SOTA, ECE 0.011), and (3) Part III-C: Tree-Jev-GRPO (12p EN + 7p CN PDF, 60.50% Formal Theorem Proving & Long-Horizon SOTA, -66% Dead-End FLOPs).

v2.0.0 - Part II: DC-Jev-GRPO (Single-Backbone System-One Process-and-Outcome LLM RL on Qwen3.5-4B)

Official v2.0.0 Release of Part II: DC-Jev-GRPO (12-Page English Preprint PDF + 8-Page Chinese Paper PDF + Release Bundle ZIP). Official CERN Zenodo DOI: https://doi.org/10.5281/zenodo.23441431 (Companion Part I DOI: https://doi.org/10.5281/zenodo.23439711).

v1.0.0: DC-Jev-RL (#1 MuJoCo v5 Leaderboard SOTA Preprint & Benchmark Suite)

Official v1.0.0 release of DC-Jev-RL (Beyond Brittle Code and Noisy Tokens: System-One Decision Models as Calibrated Dynamic-Curriculum Reward Judges for Continuous Control) by Ziyang Zhang (张子阳).

• #1 Leaderboard SOTA across 5 canonical MuJoCo v5 benchmarks (Hopper-v5, HalfCheetah-v5, Walker2d-v5, Swimmer-v5, Reacher-v5) with 2.471x Mean ONS, 2.076x IQM ONS, and 100% (5/5) win rate over Human Oracle rewards.
• 4.0x faster single-prefill /v1/systemone trajectory evaluation on NVIDIA L4 GPU (213.2 ms vs. 852.8 ms autoregressive LLM-Judge) powered by @jaredpalmer's TypeSafe AI Jev / Kev-4B architecture.