Your Transformer Can Hold Two Thoughts at Once: Evidence of Linear Superposition in LLMs
Parts-of-Speech as Emergent Categories in SAE Latent Space
Order-Invariant Answers, Order-Sensitive Representations in Mathematical Reasoning
Capable yet Parsimonious: Extracting and Characterizing Hidden Chain-of-Thought in Frontier Models
The Linear Representation Hypothesis Needs a Group Action
Linguistic Features for Interpretable Textual Entailment
Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms
DiaVLo: Diagnosing Behaviours of Vision-Language Models
A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
Visible chains of thought are a safety advantage for AI, but that transparency is slipping away
1
A Lie Detector Test for Language Models: Reading Knowledge a Model Won't Reveal
Semantic Action Graph: A Shared Representation for Agent Grounding and Human Interpretation of Sports Highlights
Base Labs launches an open-weight AI safety partnership with Hugging Face and Goodfire
2 min
NeuSOGA3D: A Neuro-Symbolic Framework for Explainable 3D Geometric Reconstruction
Flag Game: A Toy Model for Mechanistic Swarm Interpretability
Monitoring and Discovering Reward Hacking with Internal Representations during LLM Evaluations
1
Probabilistic Linear Explanations
Large Language Models Develop Belief State Geometry In-Context
1
A Chosen Future Can Still Be Rewritten: Causal Writability in Video Models
Disentangling Representation Evolution in Transformers through Directional Decomposition
1
MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup
2
Disentangling Representation Evolution in Transformers through Directional Decomposition
AI models' written reasoning steps correspond to distinct internal patterns, a new study finds
4 min2