Reproduce “Exploring and Controlling Diversity in LLM-Agent Conversation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
KuanChao Chu, Yi-Pei Chen, Hideki Nakayama
The University of Tokyo researchers introduced Adaptive Prompt Pruning (APP), a method to control dialogue diversity in LLM-agent simulations by dynamically removing prompt units based on their attention scores. This approach effectively increases conversational diversity, with a subsequent revision step ensuring contextual consistency, demonstrating advantages over traditional decoding parameter adjustments.
Key insights
- Agent-specific information within prompts, particularly memory, acts as a primary constraint on the diversity of LLM-agent conversations.
- Dynamically removing high-attention prompt content is an efficient way to increase dialogue diversity, indicating its disproportionate constraining effect.
- Maintaining conversational consistency while enhancing diversity is achievable through a targeted post-generation revision mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.