Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study introduces a novel approach to safe open-ended exploration by employing heterogeneous agent cohorts that specialize in distinct roles: a Disrupter for generating creative proposals, a Validator for enforcing runtime safety checks, and a Broker for contextual analogy retrieval. By compiling and caching failures into reusable constraint patches called Scars, the method not only enhances safety but also optimizes resource usage, leading to significant reductions in token consumption and costs. In experiments, this cohort achieved remote target exploration effectively while maintaining strict safety protocols, demonstrating a 15.1% reduction in redundant checks and a 55.9% decrease in overall token costs.
Transforming failures into reusable safety constraints, this approach allows agents to explore creatively while maintaining strict safety protocols.
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.