Bag of Dims: Training-Free Mechanistic Interpretability via Dimension-Level Sign Patterns
We show the standard basis of transformer hidden states already provides a training-free, architecture-general feature basis. Individual dimensions encode semantic content via their signs (+/-1) and confidence via their magnitudes, acting as independent binary registers; a feature is a subset of dimensions with a consistent sign pattern, read by counting sign agreements with no learned rotation. We validate this Bag of Dims framework across seven models spanning language (Qwen 3.5-4B, Gemma 3-4B, Mistral 7B, Qwen3-32B), vision (DINOv2, ViT-Base), and audio (AST).
Signs alone carry predictive content: unit-magnitude sign patterns preserve 60-93% top-5 next-token accuracy through the LM head, and decoder-free Hamming scoring reaches 80-90% top-4096. From a single-token cache (one forward pass per token, no context, no labels), we detect 175 categories at AUC 0.97-0.99 by sign agreement; a trained probe adds only +0.018 AUC and converges to axis-aligned weights. These features are causally operative: they survive the K/V attention projections, trace to the FFN neuron coalitions that write them (random-weight controls never reproduce this), and flipping a feature's signs during the live forward pass suppresses its concept across four language models, magnitude-matched and concept-specific. Dimensions stay independent throughout (pairwise mutual information below 0.006 bits).
The structure is not specific to language: the same per-dimension signs appear in self-supervised vision (DINOv2, 9/12 ImageNet superclasses), supervised vision (ViT-Base, 11/12), and audio (AST, 50/50 ESC-50 categories), so it reflects transformer training in general, not the language-modeling objective. The standard basis already suffices for feature reading at one forward pass, no optimization, no GPU-days. The open problem shifts from finding the right rotation to cataloging what each dimension encodes.
Mistral is announcing an expanded global strategic partnership with @Microsoft to give enterprises and regulated industries frontier AI they can control. As Mistral is expanding its AI compute capacit
Mistral is announcing an expanded global strategic partnership with @Microsoft to give enterprises and regulated industries frontier AI they can control. As Mistral is expanding its AI compute capacity in Europe, the companies are expanding their strategic partnership with https://t.co/tfK6pt3Ixz
X/Twitter@MistralAIMistral AIannouncementgeneral1w ago
Mistral is announcing an expanded global strategic partnership with @Microsoft to give enterprises and regulated industries frontier AI they can control. As Mistral is expanding its AI compute capacit
Mistral is announcing an expanded global strategic partnership with @Microsoft to give enterprises and regulated industries frontier AI they can control. As Mistral is expanding its AI compute capacity in Europe, the companies are expanding their strategic partnership with https://t.co/tfK6pt3Ixz
X/Twitter@MistralAIMistral AIannouncementgeneral3w ago
The Solutions team @MistralAI is hiring globally! If you’re entrepreneurial, hands-on, and want to shape how enterprises adopt AI, let’s talk. Apply: https://t.co/MV0YiejYgb https://t.co/hgLP8rFTYE
Product Your Prompts and Skills need a system of record. Studio gives AI prompts & skills a system of record—versioned, owned, and traceable. July 9, 2026 By Mistral
X/Twitter@MistralAIMistral AIannouncementgeneral3w ago
Customers are not an abstraction for us: we exist to help enterprises, public institutions, and industries build their own intelligence, so the value created from their data, workflows, feedback, and
Customers are not an abstraction for us: we exist to help enterprises, public institutions, and industries build their own intelligence, so the value created from their data, workflows, feedback, and models accrues to them rather than to model providers.
X/Twitter@MistralAIMistral AIannouncementgeneral1mo ago
Mistral officially has 1,000 team members around the world! Thank you to our incredible talent for believing in the mission and contributing to what we are building each and every day. You too can be
Mistral officially has 1,000 team members around the world! Thank you to our incredible talent for believing in the mission and contributing to what we are building each and every day. You too can be a part of the next chapter in our journey as we continue to grow and serve our https://t.co/nkxE0IQJi2
Bag of Dims: Training-Free Mechanistic Interpretability via Dimension-Level Sign Patterns
We show the standard basis of transformer hidden states already provides a training-free, architecture-general feature basis. Individual dimensions encode semantic content via their signs (+/-1) and confidence via their magnitudes, acting as independent binary registers; a feature is a subset of dimensions with a consistent sign pattern, read by counting sign agreements with no learned rotation. We validate this Bag of Dims framework across seven models spanning language (Qwen 3.5-4B, Gemma 3-4B, Mistral 7B, Qwen3-32B), vision (DINOv2, ViT-Base), and audio (AST).
Signs alone carry predictive content: unit-magnitude sign patterns preserve 60-93% top-5 next-token accuracy through the LM head, and decoder-free Hamming scoring reaches 80-90% top-4096. From a single-token cache (one forward pass per token, no context, no labels), we detect 175 categories at AUC 0.97-0.99 by sign agreement; a trained probe adds only +0.018 AUC and converges to axis-aligned weights. These features are causally operative: they survive the K/V attention projections, trace to the FFN neuron coalitions that write them (random-weight controls never reproduce this), and flipping a feature's signs during the live forward pass suppresses its concept across four language models, magnitude-matched and concept-specific. Dimensions stay independent throughout (pairwise mutual information below 0.006 bits).
The structure is not specific to language: the same per-dimension signs appear in self-supervised vision (DINOv2, 9/12 ImageNet superclasses), supervised vision (ViT-Base, 11/12), and audio (AST, 50/50 ESC-50 categories), so it reflects transformer training in general, not the language-modeling objective. The standard basis already suffices for feature reading at one forward pass, no optimization, no GPU-days. The open problem shifts from finding the right rotation to cataloging what each dimension encodes.
