DeepWism/DeepWism-R2

---

1. Introduction

We introduce DeepWism® R2(Research&Report), a revolutionary next-generation AI system based on the agents framework - a Thin-Thick-Thin Crowd Entropy Dynamics System (T3CEDS) that establishes entropy reduction as the fundamental mechanism underlying crowd intelligence. This represents a paradigm shift in AI research, moving from traditional attention-based mechanisms to entropy management as the core design principle. DeepWism® R2 innovative architecture consists of three distinct yet interconnected layers:

Thin Perception Layer: Efficiently captures high-dimensional inputs while preserving essential information and reducing input entropy

Thick Processing Layer: Leverages crowd intelligence mechanisms to actively reduce entropy through structured reasoning, collaborative processing, and deep ranking

Thin Decision Layer: Distills complex representations into coherent outputs, further reducing entropy to facilitate clear, actionable decisions

DeepWism® R2 achieves 27.5% accuracy on Humanity's Last Exam (HLE), surpassing OpenAI DeepResearch's 26.6% and establishing new state-of-the-art performance. In addition, it leads on xbench-ScienceQA with 70.0% and xbench-DeepSearch with 64.0%, outperforming all other models.

These results demonstrate superior capabilities in:

Complex Problem-Solving: Enhanced reasoning through entropy reduction mechanisms

Uncertainty Management: Effective handling of high-entropy scenarios

Multi-Domain Generalization: Robust performance across diverse problem domains including science, deep retrieval, and logic

AI Explainability: Transparent decision-making processes through entropy dynamics

2. Summary

Revolutionary Architecture: Thin-Thick-Thin Crowd Entropy Dynamics System

• Entropy-Centric Design: Unlike traditional attention-based models, DeepWism® R2 is fundamentally designed around entropy reduction principles. This paradigm shift enables more effective handling of complex, uncertain problem spaces that challenge conventional AI systems.

• Crowd Intelligence Integration: The thick processing layer implements sophisticated crowd intelligence mechanisms that leverage collective reasoning patterns to systematically reduce entropy through collaborative processing and structured analysis.

• T3CEDS Framework: The three-layer architecture optimizes information flow from high-entropy inputs to low-entropy, actionable outputs, ensuring maximum coherence and decision clarity at each stage.

Performance Excellence: State-of-the-Art Results

• Humanity's Last Exam Leadership: Achieving 27.5% accuracy on HLE, DeepWism® R2 sets new benchmarks in complex reasoning tasks, demonstrating superior performance over leading models including OpenAI DeepResearch, Gemini2.5, and Claude 3.7 Sonnet.

• Multi-Benchmark Dominance: Consistent excellence across diverse evaluation metrics including AIME2025 (93.3%), GPQA-Diamond (88.0%), MMLU (86.0%), and SWE-bench Verified (72.0%), showcasing robust generalization capabilities.

• Open Research Commitment: DeepWism® R2 framework will be fully open-sourced with global API access for authenticated users, fostering collaborative advancement in entropy-based AI research.

3. Evaluation Results

Comprehensive Benchmark Performance DeepWism® R2 demonstrates exceptional performance across multiple challenging benchmarks, establishing new state-of-the-art results in complex reasoning and problem-solving tasks.

Model	xbench-ScienceQA	xbench-DeepSearch	HLE	AIME2025	GPQA-Diamond	MMLU	SWE-bench Verified
DeepWism® R2	70.0	64.0	27.5	93.3	88.0	86.0	72.0
o3-high	60.8	65.0	20.3	–	–	–	–
o3-pro	59.6	–	–	–	–	–	–
Doubao-Seed-1.6	56.6	50.0	–	–	–	–	–
OpenAI DeepResearch	–	–	26.6	–	–	–	–
Gemini 2.5 Pro	59.4	50.0	21.6	88.0	86.4	84.5	67.2
o4 mini	50.4	60.0	18.1	92.7	81.4	82.0	68.1
Claude 4 Opus	–	–	10.7	90.0	83.3	80.7	79.4
DeepSeek-R1-0528	54.6	–	14.0	87.5	81.0	84.0	57.6

🔑 Key Performance Highlights

• 🏆 Humanity's Last Exam (HLE): 27.5% accuracy, surpassing all existing models including OpenAI DeepResearch
• 🧪 xbench-ScienceQA: 70.0% accuracy, leading all models on complex scientific question answering
• 🔍 xbench-DeepSearch: 64.0% accuracy, demonstrating top-tier deep retrieval and reasoning capabilities
• 📊 AIME2025: 93.3% accuracy, showcasing exceptional mathematical reasoning capabilities
• 🔬 GPQA-Diamond: 88.0% accuracy, indicating superior performance in graduate-level science questions
• 📚 MMLU: 86.0% accuracy, reflecting strong multi-domain knowledge understanding
• 💻 SWE-bench Verified: 72.0% accuracy, proving highly effective software engineering problem-solving skills

4. Chat Website

Experience DeepWism® R2's revolutionary capabilities through our interactive platforms:

🌐 Chat Interface: i.deepwism.com

• Real-time interaction with DeepWism® R2
• Entropy visualization in reasoning processes
• Multi-domain problem-solving capabilities

5. Contact

For questions, collaborations, or support ，please contact us at: [email protected]

Website: www.deepwism.com

GitHub Issues: Report bugs or request features Twitter: @DeepWism