Abliteration
Abliteration is the practice of removing refusal behavior from an open-weight large language model by editing its weights so that the model can no longer write a "refusal direction" into its residual stream.
Large Language Models
Explore language models, how they work, and the techniques used to build applications with them.
Explore articles
Browse subtopics: Interpretability
Articles that also belong to these categories. Counts cover all of Large Language Models.
All articles (566)AI Models (191)Open Source AI (173)Chinese AI (98)Natural Language Processing (58)AI Companies (55)Machine Learning (50)Mixture of Experts (49)Multimodal AI (42)Reasoning Models (42)OpenAI (38)Artificial Intelligence (37)AI Benchmarks (33)Anthropic (33)Google DeepMind (27)AI Safety (26)AI Agents (24)Developer Tools (24)AI Inference (20)Meta AI (20)AI Code Generation (19)Small Language Models (18)Training & Optimization (18)Generative AI (15)Model Architecture (14)NVIDIA (13)Prompt Engineering (13)AI Research (12)Deep Learning (12)Google (12)Model Evaluation (12)Data & Datasets (11)AI Tools & Products (9)Conversational AI (9)Microsoft (9)Transformer Models (9)Diffusion Models (8)AI Infrastructure (7)People (6)AI Alignment (5)AI Hardware (5)Enterprise AI (5)Information Retrieval (5)Interpretability (5)Research Organizations (5)Healthcare AI (4)Reinforcement Learning (4)Mathematics (3)AI History (2)AI for Science (2)AI in Cybersecurity (2)Algorithms (2)Computer Vision (2)Legal AI (2)Speech & Audio AI (2)AI Policy (1)AI Policy & Regulation (1)AI in Gaming (1)Autonomous Vehicles (1)ChatGPT (1)China AI (1)Drug Discovery (1)MLOps (1)Neural Networks (1)Robotics (1)Software Development (1)
Showing 1-5 of 5 articles
Activation steering
Activation steering is a family of inference-time techniques in mechanistic interpretability and AI safety that modify a neural network's internal activations to influence its behavior, without retraining the…
Patchscopes
Patchscopes is an interpretability framework for inspecting hidden representations of large language models by patching an internal activation from a source computation into a separate target inference whose…
Persona vectors
Persona vectors are single linear directions in the activation space of a large language model that correspond to high level character traits such as evil, sycophancy, or a propensity to hallucinate.
Refusal direction
The refusal direction is a finding from mechanistic interpretability research that the refusal behavior of safety fine-tuned chat language models is mediated by a single