Large Language Models (LLMs) possess remarkable reasoning abilities, yet their decision-making processes are often opaque, making it challenging to understand why they behave in unexpected ways. To address this, Gemma Scope 2 has been released as a comprehensive suite of interpretability tools for the Gemma 3 model family, ranging from 270 million to 27 billion parameters. This release is the largest open-source interpretability toolkit by an AI lab, designed to help researchers trace potential risks and better understand the internal workings of AI models. With the capability to store 110 petabytes of data and manage over a trillion parameters, Gemma Scope 2 aims to assist the AI research community in auditing and debugging AI agents, ultimately enhancing safety interventions against issues like jailbreaks and hallucinations. Interpretability research is essential for creating AI that is both safe and reliable as AI systems become more advanced and complex. Gemma Scope 2 acts like a microscope for the Gemma language models, using sparse autoencoders (SAEs) and transcoders to allow researchers to explore model internals and understand how their "thoughts" are formed and connected to behavior. This deeper insight into AI behavior is crucial for studying phenomena such as jailbreaks, where a model's internal reasoning does not align with its communicated reasoning. The new version builds on its predecessor by offering more refined tools and significant upgrades, including full coverage for the entire Gemma 3 family and advanced training techniques like the Matryoshka technique, which enhances the detection of useful concepts within models. Gemma Scope 2 also introduces tools specifically designed for analyzing chatbot behaviors, such as jailbreaks and chain-of-thought faithfulness. These tools are vital for deciphering complex, multi-step behaviors and ensuring models act as intended in conversational applications. By providing a full suite of interpretability tools, Gemma Scope 2 supports ambitious research into emergent behaviors that only appear at larger scales, such as those observed in models like the 27 billion parameter C2S Scale model. As AI technology continues to progress, tools like Gemma Scope 2 are crucial for ensuring that AI systems are not only powerful but also transparent and safe, ultimately benefiting the development of more robust AI safety measures. This matters because understanding and improving AI interpretability is crucial for developing safe and reliable AI systems, which are increasingly integrated into various aspects of society.
A local interpretability toolchain has been developed to explore the coupling of hidden dimensions in small language models, specifically Llama-3.2-3B-Instruct. By focusing on deterministic decoding and stratified prompts, the toolchain reduces noise and identifies key dimensions that significantly influence model behavior. A causal test revealed that perturbing a critical dimension, DIM 1731, causes a collapse in semantic commitment while maintaining fluency, suggesting its role in decision-stability. This discovery highlights the existence of high-centrality dimensions that are crucial for model functionality and opens pathways for further exploration and replication across models. Understanding these dimensions is essential for improving the reliability and interpretability of AI models.
A local interpretability tool has been developed to visualize and intervene in the hidden-state activity of the Llama 3.2 3B model during inference, revealing a persistent hidden dimension (dim 3039) that influences the model's commitment to its generative trajectory. Systematic tests across various prompt types and intervention conditions showed that increasing intervention magnitude led to more confident responses, though not necessarily more accurate ones. This dimension acts as a global commitment gain, affecting how strongly the model adheres to its chosen path without altering which path is selected. The findings suggest that magnitude of intervention is more impactful than direction, with significant implications for understanding model behavior and improving interpretability. This matters because it sheds light on how AI models make decisions and the factors influencing their confidence, which is crucial for developing more reliable AI systems.