A new arXiv paper presents Murano, an open-source framework for designing, running, and reproducing mechanistic interpretability studies of large language models. It targets researchers across disciplines whose workflows span loading, recording, attribution, intervention, and evaluation — areas typically covered by separate libraries. Murano represents these operations as composable steps, reducing the need to adapt outputs between tools.
Researchers propose a Universal Activation Bus that maps language models' hidden activations into a shared dense space using one lightweight linear encoder-decoder adapter pair per model. This lets activation-based tools like probes and sparse autoencoders work across models without being rebuilt for each one. The interface is frozen after source training; new models join by fitting only their adapter pair on unlabeled matched data.
Researchers propose a Universal Activation Bus that maps language models' hidden activations into a shared dense space using one lightweight linear encoder-decoder adapter pair per model. This lets activation-based tools like probes and sparse autoencoders work across models without being rebuilt for each one. The interface is frozen after source training; new models join by fitting only their adapter pair on unlabeled matched data.
Nanointerpret is a minimal open-source playground for LLM interpretability, offering sparse autoencoder (SAE) training, automatic feature interpretation, and a GUI for visualizing features and running interventions. It helps researchers and enthusiasts explore how LLMs work internally by providing hands-on tools in a simple, accessible package.
Nanointerpret is a minimal open-source playground for LLM interpretability, offering sparse autoencoder (SAE) training, automatic feature interpretation, and a GUI for visualizing features and running interventions. It helps researchers and enthusiasts explore how LLMs work internally by providing hands-on tools in a simple, accessible package.
LumiXAI is a new modular full-stack framework designed to unify fragmented feature attribution tools. It addresses specialization in single modalities, APIs, or GUIs and improves accessibility for non-expert users. This consolidation aims to make model interpretability more practical and extensible for practitioners.
LumiXAI is a new modular full-stack framework designed to unify fragmented feature attribution tools. It addresses specialization in single modalities, APIs, or GUIs and improves accessibility for non-expert users. This consolidation aims to make model interpretability more practical and extensible for practitioners.
A new arXiv paper introduces DMDIntel, a method that applies dynamic mode decomposition to LLM hidden states to rank input tokens by importance for classification predictions. Experiments across three datasets and three model families show it outperforms state-of-the-art attribution techniques.
A new arXiv paper introduces DMDIntel, a method that applies dynamic mode decomposition to LLM hidden states to rank input tokens by importance for classification predictions. Experiments across three datasets and three model families show it outperforms state-of-the-art attribution techniques.
A new arXiv paper proposes a graphical notation for designing interpretable AI architectures, adapted from Penrose tensor notation. It aims to provide a global view of an architecture while describing actual tensor manipulations, addressing shortcomings of symbolic equations and probabilistic graphical models.
A new arXiv paper proposes a graphical notation for designing interpretable AI architectures, adapted from Penrose tensor notation. It aims to provide a global view of an architecture while describing actual tensor manipulations, addressing shortcomings of symbolic equations and probabilistic graphical models.