Function Vectors Reproduction: Exploration of Functional Representation Mechanisms Inside Large Language Models

This project partially reproduces the paper 'Function Vectors in Large Language Models', exploring the concept vectors inside large language models that represent specific functions and the mechanisms of guiding model behavior through vector manipulation. Core keywords include: large language models, explainable AI, Function Vectors, model editing, Transformer, neural mechanisms, causal intervention.

Although large language models excel in various tasks, their internal working mechanisms have long been considered a 'black box'. In recent years, researchers have explored interpretable semantic structures within models, and the Function Vectors theory has attracted significant attention — it posits that there exist directions corresponding to specific functions or behavioral patterns in the activation space of the model's Transformer layers. This project is a practical validation of this cutting-edge theory.

Core hypothesis of the Function Vectors theory: When a large language model performs a specific task, the activation states of its intermediate layers form stable functional representations (e.g., activation of different functional modules such as arithmetic operations, translation, reasoning). By analyzing activation patterns, specific function vectors can be extracted, which can then be used to manipulate or enhance model behavior, opening up new possibilities for model editing, capability enhancement, and safety research.

Technical Path of Reproduction

1. Design controlled experimental scenarios and collect activation data from the model's intermediate layers under specific tasks;
2. Identify activation patterns related to target functions through comparative analysis and dimensionality reduction techniques;
3. Verify the impact of adding/subtracting function vectors on model output behavior (causal intervention is key).

Vector Extraction and Manipulation Methods

• Extraction strategies: Comparative sample difference method, gradient importance analysis method, PCA dimensionality reduction method (each has applicable scenarios and limitations);
• Manipulation mechanism: Inject function vectors during forward propagation; effects can be tested without modifying model weights.

Reproduction findings:

1. Function Vectors do exist and can be detected in models of different scales and architectures;
2. Vectors have certain transferability and can be shared between similar tasks;
3. The effect of function vectors is closely related to the depth of the injection layer; specific functions are easier to manipulate in specific layers. These findings provide a new perspective for understanding the internal organization of large models.

The research results of Function Vectors have broad application potential:

• Model safety: Identify and suppress harmful function vectors to enhance safety;
• Model customization: Strengthen specific function vectors to improve performance in vertical tasks;
• Model compression: Guide efficient pruning strategies;
• Theoretical foundation: Provide support for the development of neuro-symbolic AI and hybrid intelligent systems.

Current research limitations:

1. Limited effectiveness in handling complex multi-step reasoning tasks;
2. The interference problem between function vectors has not been fully resolved;
3. Automated discovery and classification of function vectors and extension to multi-modal models still need to be explored. This open-source reproduction project lays the foundation for subsequent community research.

Function Vectors research represents an important trend in explainable AI from 'observation' to 'intervention'. It not only satisfies academic curiosity but also provides an actionable technical path for practical applications. This project's reproduction verifies the feasibility of this direction, exposes the limitations of current methods, points out improvement directions for subsequent research, and is an open-source project worth attention in the field of large model interpretability and controllability.