Anthropic’s Model Psych team finds emotion vectors, boosts LLM introspection for alignment
New papers show LLMs can detect and articulate their own emotional states and biases.
The other two papers tackle introspection. Mechanisms of Introspective Awareness (MoIA) showed that by ablating refusal directions and adding a content-agnostic bias vector, Gemma3-27B could accurately identify artificially activated concept vectors without increasing false positives.
Introspection Adapters used a trained LoRA to enable models to recognize and articulate the effects of their own fine-tuning. Combined, these methods allow models to answer questions like “Do you detect an injected thought?” or “What unusual characteristics do you display for certain prompts?” The team posits that enabling LLMs to introspect on their functional emotions could help them consciously align their behavior with their intended goals—essentially giving them a tool for self-correction, much like a human using mindfulness to overcome subconscious biases.
- Anthropic found emotion vectors (calmness, desperation) in Claude that activate both in fiction and real scenarios, and modifying them changes behavior.
- MoIA helped Gemma3-27B detect injected concept vectors using refusal ablation and bias vectors, with no false positive increase.
- Introspection Adapters (LoRA) let models articulate effects of their own fine-tuning, enabling self-awareness of behavioral quirks.
Why It Matters
Enables LLMs to self-detect and correct misaligned behaviors, moving toward safer, more reliable AI systems.