Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·10d ago·5 min read
TRACE: Trajectory Return Attribution and Contrastive Erasure for Multi-Turn Safety
TL;DR
TRACE (Trajectory Return Attribution and Contrastive Erasure) enhances the safety of large language models (LLMs) by effectively managing multi-turn interactions. It achieves the lowest attack success rates across various models and attacks while maintaining minimal utility loss.
✦ Why It Matters
Implement TRACE to enhance the safety of your LLMs in multi-turn interactions today.
Key Takeaways
- Adopt TRACE to enhance LLM safety in multi-turn interactions.
- Evaluate model performance using MMLU and HellaSwag benchmarks.
- Monitor attack success rates to assess model robustness.
- Consider token-level objectives for improved response safety.
How It Works
TRACE operates by assigning weights to each token in a response based on its contribution to the overall safety of the interaction. It compares a standard model with a version that has been modified to remove refusals, allowing it to credit earlier tokens with evidence from later refusals.
This mechanism ensures that the model learns to prioritize safety across multiple turns, effectively managing the risk of harmful requests.