Reproduce “Recursive Semantic Divergence for LLM Agent Consistency” 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
Harshavardhan Abichandani, Penny Chong, Atin Ghosh, Daniel Dahlmeier
The authors introduce Recursive Semantic Divergence (RSD), which estimates how trajectory-level semantic disagreement propagates across turns. It scores action differences and trains agents with that score as a reward. On ToolSandbox, RSD fine-tuning for Qwen-2.5-3B raises AUC from 0.271 to 0.340 and lowers AUC CV from 0.90 to 0.44; AUC CV falls in seven of eight settings, but consistency does not guarantee correctness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.