Recursive Semantic Divergence for LLM Agent Consistency
Source
arxiv.org
Author
Harshavardhan Abichandani, Penny Chong, Atin Ghosh, Daniel Dahlmeier
Date
Why it matters
Task-success scores miss agents that behave differently on identical runs. RSD measures cumulative semantic disagreement between two rollouts from the same state, giving a way to quantify reliability before deployment.
Key takeaways · AI-distilled
Recursive Semantic Divergence (RSD) measures AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → consistency as the expected cumulative semantic disagreement between two independent rollouts from the same state, using natural language inferenceRunning a trained model to get answers — the phase where AI is actually used, as opposed to trained.Full definition → to detect contradictions.
RSD is adapted from bisimulation metrics in reinforcement learning and approximated by a trained network, so it can score consistency at each turn without waiting for a full trajectory to finish.
Used as a fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → reward, RSD cut variation in task progress from 132% to 91% on tau2-bench and from 90% to 44% on ToolSandbox, and matched supervised fine-tuning with lower variance and no ground-truth labels, per the abstract.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
inference — Running a trained model to get answers — the phase where AI is actually used, as opposed to trained.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
