This research identifies a J-space within large language models like Claude that functions similarly to human conscious access via a global workspace. This internal subspace contains neural patterns that are reportable, modifiable on request, and used for silent reasoning without appearing explicitly in text output. While most of the model's processing is automatic and unconscious, this specialized channel allows for higher-order cognitive tasks by broadcasting information across the network.
- Discovery of J-space through Jacobian lens technique
- Comparison to human global workspace theory
- Distinction between reportable thoughts and automatic processing
- Ability to monitor silent reasoning and intent via internal activations
- Discovery of J-space through Jacobian lens technique
- Comparison to human global workspace theory
- Distinction between reportable thoughts and automatic processing
- Ability to monitor silent reasoning and intent via internal activations
Steerling-8B is an interpretable causal diffusion language model that combines masked diffusion language modeling with concept decomposition, enabling generation, attribution, steering, and extraction of hidden representations. It offers features like block-causal attention and decomposition of hidden states into known and unknown concepts.
This article explores the field of mechanistic interpretability, aiming to understand how large language models (LLMs) work internally by reverse-engineering their computations. It discusses techniques for identifying and analyzing the functions of individual neurons and circuits within these models, offering insights into their decision-making processes.
Researchers are studying large language models as if they were living things, discovering secrets by applying biological and neurological analysis techniques. This approach is revealing unexpected behaviors and limitations of LLMs.
This paper demonstrates that the inference operations of several open-weight large language models (LLMs) can be mapped to an exactly equivalent linear system for an input sequence. It explores the use of the 'detached Jacobian' to interpret semantic concepts within LLMs and potentially steer next-token prediction.
Sparse autoencoders (SAEs) have been trained on Llama 3.3 70B, releasing an interpreted model accessible via API, enabling research and product development through feature space exploration and steering.
This article provides a non-technical guide to interpreting SHAP analyses, useful for explaining machine learning models to non-technical stakeholders, with a focus on both local and global interpretability using various visualization methods.
The article discusses techniques to improve outlier detection in tabular data by using subsets of features, known as subspaces, which can reduce the curse of dimensionality, increase interpretability, and allow for more efficient execution and tuning over time.
Gemma Scope is an open-source, multi-scale, high-throughput microscope system that combines brightfield, fluorescence, and confocal microscopy, designed for imaging large samples like brain tissue.
DeepMind's Gemma Scope provides researchers with tools to better understand how Gemma 2 language models work through a collection of sparse autoencoders. This helps in understanding the inner workings of these models and addressing concerns like hallucinations and potential manipulation.
