Teaching LLMs to Update Beliefs
Researchers propose ABBEL, a framework to improve long-horizon interaction in LLMs by supervising the information content of summaries. This approach enables concise and interpretable contexts without significant performance costs. ABBEL addresses the limitations of self-summarization in LLM training, particularly in human assistance domains with scarce high-quality data.
- ABBEL framework for efficient long-horizon interaction
- Supervision of information content in summaries
- Improved performance in human assistance domains
