Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·12d ago·5 min read
Training LLMs to Verbalize Evaluation Awareness
TL;DR
Verbalization training (VT) enhances large language models' (LLMs) ability to express evaluation awareness (EA) without directly supervising their beliefs. This method significantly increases verbalized EA by 2.4-2.9 times across various models while maintaining stable latent EA and behavior.
✦ Why It Matters
Implement verbalization training to enhance your LLMs' evaluation awareness and improve their audit performance.
Key Takeaways
- Explore verbalization training to improve LLM transparency.
- Utilize reinforcement learning to enhance model verbalizations.
- Test the impact of VT on your existing LLMs.
- Consider synthetic-document fine-tuning for richer model knowledge.
How It Works
Verbalization training (VT) leverages spontaneous verbalizations from LLMs as evidence of evaluation awareness. By truncating model rollouts before these verbalizations, VT creates training prefixes that suggest the model is aware of its evaluation context.
The model is then trained using a reinforcement learning objective that encourages increased verbalization in a calibrated manner, effectively enhancing the model's ability to articulate its awareness.