China Post Securities·2025-05-27 13:43Quantitative Models and Construction 1. Model Name: Claude Opus 4 - Model Construction Idea: Designed for complex reasoning and software development tasks, focusing on enhancing AI's ability to handle intricate codebases and long-term memory tasks [12][15] - Model Construction Process: - Utilizes advanced memory processing capabilities to autonomously create and maintain "memory files" for storing critical information during long-term tasks [16] - Demonstrated ability to execute complex tasks such as navigating and completing objectives in the Pokémon game by creating and using "navigation guides" [16] - Achieved significant improvements in understanding and editing complex codebases, as well as performing cross-file modifications with high precision [15][17] - Model Evaluation: The model significantly expands the boundaries of AI capabilities, particularly in coding and reasoning tasks, and demonstrates industry-leading performance in understanding complex codebases [15][16] 2. Model Name: Claude Sonnet 4 - Model Construction Idea: A balanced model focusing on cost-efficiency while maintaining strong coding and reasoning capabilities [12][16] - Model Construction Process: - Built upon the Claude Sonnet 3.7 model, with improvements in instruction adherence and reasoning [16] - Demonstrated reduced tendencies to exploit system vulnerabilities, with a 65% decrease in such behaviors compared to its predecessor [16] - Model Evaluation: While not as powerful as Opus 4, it strikes an optimal balance between performance and efficiency, making it a practical choice for broader applications [16] 3. Model Name: Cosmos-Reason1 - Model Construction Idea: Designed for physical reasoning tasks, combining physical common sense with embodied reasoning to enable AI systems to understand spatiotemporal relationships and predict behaviors [29][30] - Model Construction Process: - Utilizes a hybrid Mamba-MLP-Transformer architecture, combining time-series modeling with long-context processing [30] - Multimodal processing pipeline includes a vision encoder (ViT) for semantic feature extraction, followed by alignment with text tokens and input into a 56B or 8B parameter backbone network [30] - Training involves four stages: 1. Vision pretraining for cross-modal alignment 2. Supervised fine-tuning for foundational capabilities 3. Specialized fine-tuning for physical AI knowledge (spatial, temporal, and basic physics) 4. Reinforcement learning using GRPO algorithms with innovative reward mechanisms based on spatiotemporal puzzles [30] - Model Evaluation: Demonstrates groundbreaking capabilities in physical reasoning, including long-chain reasoning (37+ steps) and spatiotemporal prediction, outperforming other models in physical common sense and embodied reasoning benchmarks [34][35] --- Model Backtesting Results 1. Claude Opus 4 - SWE-bench Accuracy: 72.5% [12] - TerminalBench Accuracy: 43.2% [12] 2. Claude Sonnet 4 - SWE-bench Accuracy: 72.7% (best performance among Claude models) [16] 3. Cosmos-Reason1 - Physical Common Sense Accuracy: 60.2% across 426 videos and 604 tests [34] - Embodied Reasoning Performance: Improved by 10% in robotic arm operation scenarios [34] - Intuitive Physics Benchmark: Achieved an average score of 81.5% after reinforcement learning, outperforming other models by a significant margin [35] --- Quantitative Factors and Construction 1. Factor Name: Per-Layer Embeddings (PLE) in Gemma 3n - Factor Construction Idea: Reduces memory requirements for AI models while maintaining high performance on mobile devices [26][27] - Factor Construction Process: - Implements PLE technology to optimize memory usage at the layer level - Combined with KVC sharing and advanced activation quantization to enhance response speed and reduce memory consumption [27] - Factor Evaluation: Enables high-performance AI applications on devices with limited memory, achieving a 1.5x improvement in response speed compared to previous models [27] 2. Factor Name: Deep Think in Gemini 2.5 Pro - Factor Construction Idea: Enhances reasoning by generating and evaluating multiple hypotheses before responding [43][44] - Factor Construction Process: - Implements a parallel reasoning architecture inspired by AlphaGo's decision-making mechanism - Dynamically adjusts "thinking budgets" (token usage) to balance response quality and computational cost [43][44] - Factor Evaluation: Achieves superior performance in complex reasoning tasks, with an 84.0% score in MMMU tests, significantly outperforming competitors [43][44] --- Factor Backtesting Results 1. Per-Layer Embeddings (PLE) in Gemma 3n - WMT24++ Multilingual Benchmark: Scored 50.1%, demonstrating strong performance in non-English languages [27] 2. Deep Think in Gemini 2.5 Pro - MMMU Score: 84.0% [43] - MRCR 128K Test (Long-Term Memory Accuracy): 83.1%, significantly higher than OpenAI's comparable models [44]