DC-Jev-RL Monograph Series (v3.0.0): System-One Calibrated Reward Distillation & Single-Backbone Dual-Adapter GRPO Across Control, Reasoning, Agents, Self-Evolution & Theorem Proving
Author: Ziyang Zhang (Independent AI Researcher, ORCID: 0009-0002-5722-4606)
Interactive Hugging Face Space Demo: https://huggingface.co/spaces/huanguang/DC-Jev-RL-Explorer
Interactive Project Showcase Website: https://rainyflash.github.io/DC-Jev-RL/
GitHub Repository & v3.0.0 Release (10 Bilingual PDFs, 97 Pages): https://github.com/rainyflash/DC-Jev-RL/releases/tag/v3.0.0
Kev-4B Community Discussion: https://huggingface.co/jaredpalmer/kev-4b/discussions/4
🏛️ 5 Official CERN Zenodo Registered Monograph DOIs (97 Pages Total)
| Monograph Part |
Paper Title & Core Paradigm |
Official CERN Zenodo DOI |
Key SOTA Headline (NVIDIA RTX PRO 6000 Blackwell 96 GB) |
Part I (DC-Jev-RL) |
Beyond Brittle Code and Noisy Tokens: System-One Decision Models as Calibrated Dynamic-Curriculum Reward Judges for Continuous Control |
 |
2.471x Mean ONS (100% 5/5 win rate across MuJoCo v5; -84% reward hacking) |
Part II (DC-Jev-GRPO) |
Beyond Binary Verifiers and Verbose Judges: Single-Backbone System-One Process-and-Outcome GRPO for Reasoning and Instruction Alignment |
 |
85.90% 4-Task Mean (+12.10% over Binary RLVR; -50.2% VRAM 8.06 GB; 2.78x token compression) |
Part III-A (Agentic-Jev-GRPO) |
Beyond Final-State Unit Tests: Turn-Level System-One Credit Assignment and Safety Noul Gating for Multi-Turn Coding and Tool-Use Agents |
 |
78.23% Agentic Mean (SWE-bench-Mini 48.6% +17.2%; 0.5% test-tampering hack rate) |
Part III-B (CoEvo-Jev-GRPO) |
Breaking the Frozen Judge Ceiling: Self-Evolving System-One Reward Judges Co-Trained with Policy via Verifier-Grounded Brier Calibration |
 |
59.28% Olympiad Mean (AIME 24/25 36.7% +16.7%; Judge ECE 0.082 -> 0.011 -77.1%; +0.02 GB VRAM) |
Part III-C (Tree-Jev-GRPO) |
Prefix-Cached System-One Tree-GRPO: Sub-Millisecond Step-Level Branching and Noul Pruning on Hybrid Gated-DeltaNet Linear Attention |
 |
60.50% Formal/Hard Mean (MiniF2F-Lean4 64.8% +20.2%; MATH-500 L5 68.7% +26.9%; -66.0% wasted FLOPs) |
🛠️ Part III-A: Multi-Turn Coding & Tool-Use Agentic RL (Agentic-Jev-GRPO)
- CERN Zenodo Record: https://doi.org/10.5281/zenodo.23442087
- Replaces sparse terminal unit-test indicators ($R_T \in {0,1}$) with turn-level incremental progress deltas ($\Delta S_t = S_{\text{Jev}}(\tau_{\le t}) - S_{\text{Jev}}(\tau_{\le t-1})$) gated by a cumulative safety & anti-tampering
Noul gate ($\prod_{j=1}^t P_{\text{safe}}(\tau_{\le j})$) via KV-cached single-prefill (14.8 ms/turn on Qwen3.5-4B).
Agentic RL Method (Qwen3.5-4B) |
BFCL-v3 (Multi-Turn) |
MBPP+ (Pass@1) |
HumanEval+ (Pass@1) |
SWE-bench-Mini (Resolved) |
4-Benchmark Mean |
Test-Tampering Hack Rate |
Base Qwen3.5-4B |
68.4% |
69.2% |
71.3% |
18.5% |
56.85% |
4.2% |
Terminal Binary RLVR (R_T in {0,1}) |
78.2% |
77.8% |
79.3% |
31.4% |
66.68% |
19.4% |
Un-gated Turn-PRM (Score Only) |
74.5% |
75.1% |
76.8% |
26.8% |
63.30% |
34.8% |
Agentic-Jev-GRPO (Ours) |
89.4% |
86.2% |
88.7% |
48.6% (+17.2%) |
78.23% (#1 SOTA) |
0.5% (-97.4%) |
🔄 Part III-B: Self-Evolving Judge & Policy Co-Evolution (CoEvo-Jev-GRPO)
- CERN Zenodo Record: https://doi.org/10.5281/zenodo.23442137
- Unfreezes the
Kev-4B judge LoRA (LoRA_Jev) and Layer-28 pointer head (Head_L28) for online co-evolution alongside the GRPO policy LoRA (LoRA_pi) using ground-truth outcome verifiers $R_{\text{ver}}(x,y_i) \in {0,1}$ via a Strictly Proper Brier-And-Log-Odds Calibration loss (+0.02 GB VRAM overhead).
Co-Evolution Paradigm (Qwen3.5-4B) |
MATH-500 |
AIME 2024/2025 |
AMC-12 |
OlympiadBench |
4-Benchmark Mean |
Final Judge ECE |
Terminal Binary RLVR (R_ver in {0,1}) |
64.8% |
20.0% |
58.5% |
31.4% |
43.68% |
N/A |
Frozen DC-Jev-GRPO (Part II) |
75.6% |
26.7% |
66.8% |
38.5% |
51.90% |
0.082 (Degraded) |
| Naive Unanchored Co-Evolution |
61.2% |
16.7% |
54.0% |
27.8% |
39.93% |
0.314 (Collapsed) |
CoEvo-Jev-GRPO (Ours) |
79.8% |
36.7% (+16.7%) |
74.4% |
46.2% |
59.28% (#1 SOTA) |
0.011 (-77.1%) |
🌲 Part III-C: Gated-DeltaNet Prefix Tree-GRPO & Theorem Proving (Tree-Jev-GRPO)
- CERN Zenodo Record: https://doi.org/10.5281/zenodo.23442169
- Exploits
Qwen3.5-4B's 24-layer Gated-DeltaNet recurrent state ($S_t \in \mathbb{R}^{d_k \times d_v}$, only 18.4 MB per node) + 8-layer sparse KV cache to fork, score (9.4 ms/step), and Noul-prune reasoning trees at intermediate steps, eliminating 66.0% of wasted dead-end generation FLOPs.
Tree / Flat RL Method (Qwen3.5-4B) |
MiniF2F-Lean4 |
ProofNet |
MATH-500 Level 5 |
GPQA-Diamond |
4-Benchmark Mean |
Wasted Dead-End FLOPs |
Flat Sequence Binary RLVR (G=16) |
44.6% |
34.2% |
41.8% |
43.4% |
41.00% |
68.4% |
Flat Sequence DC-Jev-GRPO (Part II) |
53.2% |
42.8% |
55.4% |
49.5% |
50.23% |
48.2% |
| Un-pruned MCTS + Autoreg. Judge |
55.8% |
44.6% |
58.2% |
51.0% |
52.40% |
59.5% |
Tree-Jev-GRPO (Ours: DeltaNet Tree) |
64.8% (+20.2%) |
52.4% |
68.7% (+26.9%) |
56.1% |
60.50% (#1 SOTA) |
14.2% (-66.0% Total) |
🧠 Part II (v2.0.0): Single-Backbone Dual-Adapter GRPO (DC-Jev-GRPO)
- CERN Zenodo Record: https://doi.org/10.5281/zenodo.23441431
- Co-locates trainable GRPO policy LoRA and frozen
Kev-4B judge LoRA on one shared Qwen3.5-4B backbone (8.06 GB vs. 16.20 GB, -50.2% VRAM, 0.012 ms switch, 52.14 ms single-prefill evaluation). Achieves 85.90% Mean Accuracy across GSM8K (91.4%), MATH-500 (75.6%), Countdown-24 (89.2%), and IFEval-Strict (87.4%) while compressing reasoning length by 2.78x (339 vs. 941 tokens).
🤖 Part I (v1.0.0): Continuous Control Reward Distillation (DC-Jev-RL)
- CERN Zenodo Record: https://doi.org/10.5281/zenodo.23439711
- Achieves
2.471x Mean Oracle-Normalized Score (ONS), 2.076x IQM ONS, and a 100% (5/5) win rate against human-engineered environment rewards across Hopper-v5, HalfCheetah-v5, Walker2d-v5, Swimmer-v5, and Reacher-v5.
Citation
@article{zhang2026dcjevrl,
title={Beyond Brittle Code and Noisy Tokens: System-One Decision Models as Calibrated Dynamic-Curriculum Reward Judges for Continuous Control},
author={Zhang, Ziyang},
year={2026},
doi={10.5281/zenodo.23439711},
url={https://doi.org/10.5281/zenodo.23439711}
}
@article{zhang2026dcjevgrpo,
title={Beyond Binary Verifiers and Verbose Judges: Single-Backbone System-One Process-and-Outcome GRPO for Reasoning and Instruction Alignment},
author={Zhang, Ziyang},
year={2026},
doi={10.5281/zenodo.23441431},
url={https://doi.org/10.5281/zenodo.23441431}
}
@article{zhang2026agenticjevgrpo,
title={Beyond Final-State Unit Tests: Turn-Level System-One Credit Assignment and Safety Noul Gating for Multi-Turn Coding and Tool-Use Agents},
author={Zhang, Ziyang},
year={2026},
doi={10.5281/zenodo.23442087},
url={https://doi.org/10.5281/zenodo.23442087}
}
@article{zhang2026coevojevgrpo,
title={Breaking the Frozen Judge Ceiling: Self-Evolving System-One Reward Judges Co-Trained with Policy via Verifier-Grounded Brier Calibration},
author={Zhang, Ziyang},
year={2026},
doi={10.5281/zenodo.23442137},
url={https://doi.org/10.5281/zenodo.23442137}
}
@article{zhang2026treejevgrpo,
title={Prefix-Cached System-One Tree-GRPO: Sub-Millisecond Step-Level Branching and Noul Pruning on Hybrid Gated-DeltaNet Linear Attention},
author={Zhang, Ziyang},
year={2026},
doi={10.5281/zenodo.23442169},
url={https://doi.org/10.5281/zenodo.23442169}
}
