| --- |
| license: other |
| license_name: research-only |
| license_link: LICENSE |
| language: |
| - en |
| - zh |
| tags: |
| - reflexive-intelligence |
| - multi-reward-grpo |
| - cognitive-architecture |
| - financial-reasoning |
| - observer-depth |
| - phase-transition |
| - ouroboros |
| - mixture-of-experts |
| pipeline_tag: text-generation |
| library_name: transformers |
| --- |
| |
| # Ouroboros V24: Cognitive Architecture for Reflexive Financial Reasoning |
|
|
| **Ouroboros V24** is the latest iteration of a cognitive architecture designed for autonomous financial decision-making. Built on a 35B-parameter Mixture-of-Experts (MoE) base model with ~3B active parameters, trained through **24 iterative rounds** of multi-reward GRPO with a **54-dimensional cognitive reward topology**. |
|
|
| > ⚠️ **Weights are not publicly released.** This model card documents the architecture and training methodology. For research collaboration inquiries, contact the author. |
|
|
| ## Architecture |
|
|
| ### Base Model |
| - **Type**: Mixture-of-Experts (MoE) |
| - **Total Parameters**: ~35B |
| - **Active Parameters**: ~3B per token |
| - **Context Window**: 32K tokens |
|
|
| ### Training Methodology |
| - **Algorithm**: R-GRPO (Reflexive Group Relative Policy Optimization) |
| - **Training Rounds**: 24 iterative cycles (V1 → V24) |
| - **Adapter Strategy**: 20-layer sequential LoRA merge chain |
| - **Reward Architecture**: SCRGNDWMT (9-tier, 54 sub-dimensions) |
|
|
| ### 9-Tier Reward Topology (SCRGNDWMT) |
|
|
| | Tier | Name | Sub-dimensions | Description | |
| |------|------|----------------|-------------| |
| | **S** | Structure | 6 | XML formatting, JSON decision blocks | |
| | **C** | Content | 7 | Domain expertise, data fidelity, causal depth | |
| | **R** | Reasoning | 5 | Temporal-causal chains, counterfactual depth | |
| | **G** | Game Theory | 5 | K-level thinking, deception detection, coalition | |
| | **N** | Narrative | 4 | Scenario construction, debate, arc coherence | |
| | **D** | Data Fidelity | 3 | Numerical accuracy, source attribution | |
| | **W** | World Model | 6 | Regime detection, cross-market transmission, macro | |
| | **M** | Metacognition | 7 | Self-awareness, Bayesian confidence, falsification | |
| | **T** | Temporal-Causal | 5 | Causal chains, temporal depth, granularity | |
|
|
| ### V24 Upgrades (from V22) |
| - **C7 (CausalChainDepthV2)**: Multi-step causal chains with time-lag annotations |
| - **M7 (BayesianConfidence)**: Calibrated confidence field in JSON decisions |
| - **W3 (CrossMarketPath)**: Structural contagion paths (Market A → Mechanism → Market B) |
| - **M5 (FalsificationV2)**: Quantitative, price-based invalidation conditions |
|
|
| ### Key Training Parameters |
|
|
| | Parameter | Value | |
| |-----------|-------| |
| | Learning rate | 5 × 10⁻⁷ | |
| | Group size | 12 | |
| | Max completion tokens | 1000 | |
| | Temperature | 1.15 | |
| | β-annealing | Stable (β=0.05) ↔ Break-up (β=0.03) | |
| | LoRA rank | ≥ 10 | |
|
|
| ## Key Results |
|
|
| ### Reflexive Intelligence Emergence |
| During V17 training, reflexive reasoning emerged through a **discontinuous phase transition** at Step 153 — after 150+ steps of zero reflexivity scores, the capability appeared spontaneously and sustained. This is documented in Papers 1-3 of the research program. |
|
|
| ### V24 Training (ongoing) |
| - **54-dimensional reward** actively guiding cognitive development |
| - **Bayesian confidence calibration** observed from Step 18 |
| - **Cross-market causal reasoning** emerging by Step 25 |
| - **Zero gradient failures** through 55+ steps |
|
|
| ## Research Program |
|
|
| This model is part of a six-paper research program: |
|
|
| | Paper | Title | DOI | |
| |-------|-------|-----| |
| | P1 | Reflexive Intelligence in LLMs | [10.5281/zenodo.19557261](https://doi.org/10.5281/zenodo.19557261) | |
| | P2 | Observer Depth (ReflexBench) | [10.5281/zenodo.19627242](https://doi.org/10.5281/zenodo.19627242) | |
| | P3 | When Rewards Collide (Multi-Reward GRPO) | [10.5281/zenodo.19665969](https://doi.org/10.5281/zenodo.19665969) | |
| | P4 | Ouroboros V22 Architecture | [10.5281/zenodo.19666786](https://doi.org/10.5281/zenodo.19666786) | |
| | P5 | The Cognitive Lifecycle | [10.5281/zenodo.19666806](https://doi.org/10.5281/zenodo.19666806) | |
| | P6 | Cognitive Reward Topology | [10.5281/zenodo.19666829](https://doi.org/10.5281/zenodo.19666829) | |
|
|
| ## Related Resources |
|
|
| | Resource | Link | |
| |----------|------| |
| | **ReflexBench Dataset** | [MMJBDS/reflexbench](https://huggingface.co/datasets/MMJBDS/reflexbench) | |
| | **ReflexBench Eval Results** | [MMJBDS/reflexbench-eval](https://huggingface.co/datasets/MMJBDS/reflexbench-eval) | |
| | **Papers Repository** | [github.com/mmjbds/ouroboros-papers](https://github.com/mmjbds/ouroboros-papers) | |
| | **Evaluation Code** | [github.com/mmjbds/reflexbench](https://github.com/mmjbds/reflexbench) | |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{zhang2026ouroborosv22, |
| title={Ouroboros V22: Bayesian Scenario Simulation and Recurrent Depth Cognition}, |
| author={Zhang, Mian}, |
| year={2026}, |
| doi={10.5281/zenodo.19666786} |
| } |
| |
| @article{zhang2026topology, |
| title={Cognitive Reward Topology: A Nine-Tier Architecture for Multi-Reward GRPO}, |
| author={Zhang, Mian}, |
| year={2026}, |
| doi={10.5281/zenodo.19666829} |
| } |
| ``` |
|
|
| ## Author |
|
|
| - **Mian Zhang** — Independent AI Researcher |
| - **ORCID**: [0009-0001-9556-3839](https://orcid.org/0009-0001-9556-3839) |
| - **Email**: 373743743@qq.com |
| - **GitHub**: [@mmjbds](https://github.com/mmjbds) |
| - **Twitter/X**: [@Henry_Avery666](https://x.com/Henry_Avery666) |
| - **LinkedIn**: [henryavery-mianzhang](https://linkedin.com/in/henryavery-mianzhang) |
|
|
| ## License |
|
|
| This model card is released under CC BY 4.0. Model weights are not publicly available. |
|
|
