Topic · 29 episodes across 11 reviews

Inside the Model: Sycophancy, Emotion, and Bias

← all reviews

Three papers looked beneath model behavior — finding a sycophancy circuit that survives alignment, emotion vectors that causally drive misbehavior, and political-bias audits that may be measuring the wrong thing entirely.

Covered in these reviews

Related concepts

Episodes and papers for this topic live on the concept pages.

AI Alignment AI Safety Mechanistic Interpretability Causal Intervention Evaluation & Benchmarks Activation Steering Reward Hacking Linear Representation CoT Faithfulness Sycophancy Residual Stream Agentic Misalignment Strategic Deception Chain of Thought