Workflow
Pythia
icon
Search documents
AI解数学题只靠最后一个token
量子位· 2025-09-14 05:05
Core Insights - The research indicates that in mental arithmetic tasks, the majority of calculations are concentrated on the last token, rather than being distributed across all tokens, suggesting that global information access is not necessary for specific tasks like mental arithmetic [1][11]. Group 1: Research Methodology - Researchers employed Context-Aware Mean Ablation (CAMA) and attention-based peeking techniques to conduct a series of ablation experiments on models like Llama-3-8B [2][22]. - The experiments aimed to identify the "minimum computation" required for models to perform well by systematically removing or altering parts of the model [3]. - A sparse subgraph termed "All-for-One" (AF1) was identified, which allows efficient computation with minimal layers and limited information transfer [4][5]. Group 2: Model Structure and Functionality - In the AF1 structure, initial layers (L_wait) do not perform calculations related to their own values but instead focus on general preparatory tasks [7]. - Information is transferred to the last token through intermediate layers (L_transfer), which then independently performs the final calculations [8][9]. - This separation of general computation and input-specific computation highlights the model's efficiency in handling arithmetic tasks [10]. Group 3: Experimental Findings - The experiments revealed that Llama-3-8B requires only the first 14 layers for general computation, followed by 2 layers for information transfer, with the remaining layers dedicated to the last token's self-computation [24][26]. - AF1_llama demonstrated high fidelity across eight tasks, maintaining performance levels close to the original model [28][29]. - The importance of specific attention heads in arithmetic calculations was confirmed, with the model retaining approximately 95% accuracy even after removing nearly 60 heads, indicating redundancy in attention heads [30]. Group 4: Generalization and Limitations - AF1_llama was tested for its ability to generalize to other arithmetic forms, showing high accuracy in direct arithmetic tasks but failing in tasks requiring semantic understanding, such as word problems and Python code [32][34]. - Similar AF1-like subgraphs were found in Pythia and GPT-J models, although these models exhibited shorter waiting periods and less clear performance boundaries compared to Llama [35][36]. Group 5: Contributions and Innovations - This research contributes to the understanding of arithmetic reasoning and cross-token computation mechanisms in large language models [37]. - The methodologies introduced, CAMA and ABP, offer innovative approaches that could extend beyond arithmetic tasks to broader applications [37].
“基因魔剪”携手AI提升DNA编辑精度
Ke Ji Ri Bao· 2025-08-13 10:00
Core Insights - Scientists from Swiss Federal Institute of Technology Zurich, University of Zurich, and Ghent University have developed an innovative gene editing method that combines advanced "gene scissors" technology with artificial intelligence (AI), significantly enhancing the precision of DNA editing [1][2] - The new AI tool named "Pythia" allows scientists to predict the outcomes of gene editing with unprecedented accuracy, particularly in how cells will repair DNA breaks caused by techniques like CRISPR/Cas9 [1][2] Group 1 - The technology aims to ensure that "gene scissors" do not cause unintended genetic mutations while maintaining overall genomic stability, which is crucial for safe treatments [1] - The AI-guided repair templates have been validated in human cell culture systems, achieving efficient and precise gene editing and foreign gene integration [2] - The method has been tested in various biological models, including African clawed frogs and live mice, successfully editing DNA in non-dividing tissues such as the brain [2] Group 2 - Pythia utilizes identifiable and learnable repair patterns, indicating that DNA repair processes are not random but follow specific rules [2] - The technology can modify single bases or insert foreign genes and can also be used to tag specific proteins, enabling direct observation of protein behavior in healthy and diseased tissues [2] - Its versatility allows application in various cell types, including non-dividing cells like neurons, providing new hope for treating challenging conditions such as neurological diseases [2]
“基因魔剪”携手AI提升DNA编辑精度 为未来精准基因疗法发展奠定基础
Ke Ji Ri Bao· 2025-08-13 00:09
Core Insights - Scientists from Swiss Federal Institute of Technology Zurich, University of Zurich, and Ghent University have developed an innovative gene editing method that combines advanced "gene scissors" technology with artificial intelligence (AI), significantly enhancing the precision of DNA editing [1][2] - The new AI tool named "Pythia" allows scientists to predict the outcomes of gene editing with unprecedented accuracy, particularly in how cells will repair DNA breaks caused by techniques like CRISPR/Cas9 [1][2] Group 1 - The technology aims to ensure that "gene scissors" do not cause unintended genetic mutations while maintaining overall genomic stability, which is crucial for safe treatments [1] - Pythia has been validated in human cell culture systems, achieving efficient and precise gene editing and foreign gene integration [2] - The method has been tested in various biological models, including African clawed frogs and live mice, successfully editing DNA in non-dividing tissues such as the brain [2] Group 2 - Pythia utilizes identifiable and learnable repair patterns, indicating that DNA repair processes are not random but follow specific rules [2] - The technology can modify single bases, insert foreign genes, and label specific proteins, enabling direct observation of protein behavior in healthy and diseased tissues [2] - Its versatility allows application in various cell types, including non-dividing cells like neurons, providing new hope for targeting difficult-to-treat neurological diseases [2] Group 3 - The predictive capability of Pythia is likened to meteorologists using AI to forecast weather, emphasizing its importance for safe, reliable, and clinically applicable gene editing [2]
X @CoinMarketCap
CoinMarketCap· 2025-07-22 14:00
Top Token Movers - Chainbasehq (C) rockets +134.6% [1] - DIA (DIA) surges +133% [1] - Pythia (PYTHIA) gains +80.9% [1] - Qubic (QUBIC) climbs +57% [1] Market Drivers - Strong fundamentals and fresh partnerships underpin these rallies [1]