Abstract
Current large language model (LLM) agents are caught in an awkward bind. When constrained by static safety instructions, they rarely venture beyond the obvious; when given free rein with tools and multi-agent debate, safety violations quickly follow. To address this, we propose separating concerns across specialized roles instead of forcing a single model to balance creativity and caution.
In our framework, a Disrupter generates unconventional proposals, a Validator enforces strict runtime checks at the tool gateway, and a Broker pulls in relevant distant analogies. Failures are not discarded; instead, they are compiled via Monte Carlo Tree Search (MCTS) into compact, signed constraint patches we call Scars. These patches are cached locally and inherited by future cohorts, transforming repeated failures into reusable, low-cost runtime constraints.
We validated this approach in a spatial-semantic sandbox (N=20 runs, p=0.05), demonstrating that agents can explore creatively while maintaining safety.
Blogger's Review: This research effectively addresses the conflict between creativity and safety in LLMs through role separation, enhancing the exploration capabilities of agents with runtime constraint memory. Its theoretical and practical implications are significant.