🤖 AI Summary
This work addresses the tension in large language model agents between stifling innovation through static safety constraints and risking unsafe behavior via unconstrained interaction. The authors propose a heterogeneous multi-agent collaboration framework comprising three specialized roles: a Disrupter that generates unconventional solutions, a Validator that enforces hard runtime checks prior to tool invocation, and a Broker that stimulates creativity through distant analogies. Failed attempts are distilled via Monte Carlo Tree Search (MCTS) into lightweight, inheritable constraint patches termed “Scars,” while a credit-based communication scoring mechanism (CAS) dynamically regulates bandwidth allocation. Experimental results in a spatial semantic sandbox demonstrate significantly enhanced exploratory capability (p<0.01), complete elimination of execution-level safety violations, a 15.1% reduction in token consumption due to Scars, and a 55.9% decrease in total communication overhead under resource constraints attributable to CAS.
📝 Abstract
LLM agents today are caught in an awkward bind. Lock them down with static safety instructions and they rarely venture beyond the obvious; give them free reign with tools and multi-agent debate, and safety violations quickly follow. Rather than forcing a single model to juggle both creativity and caution, we separate the concerns across specialized roles. A Disrupter generates unconventional proposals, a Validator enforces hard runtime checks at the tool gateway, and a Broker pulls in distant but relevant analogies. Failures are not discarded -- they are compiled, via MCTS, into compact, signed constraint patches we call Scars. These patches are cached locally and inherited by future cohorts, turning repeated failures into reusable, low-cost runtime constraints. In a spatial-semantic sandbox (N=20 runs, p<0.01), our cohort reaches remote targets where debate fails, the Validator prevents all executed breaches, and Scars reduce token consumption by 15.1% by avoiding redundant validator checks. Furthermore, credit-based Communication Allocation Scores (CAS) restrict outbound bandwidth, reducing overall token costs by 55.9% under resource constraints.