Safety-Constrained Abstention: Predictive Safety Networks Meet Collaborative Self-Play

Real-world assistants face safety and resource limits. Building on Predictive Safety Networks (Guo & Bürger, 2019), we add safety constraints to the self-play environment: certain tool invocations or actions are unsafe unless confidence exceeds a threshold or counter-evidence is fetched. The team is rewarded for task success and adherence to safety envelopes, and penalized for unsafe shortcuts. This conflicts with naïve optimization where agents might “darkly” rely on tools. It’s novel to tie abstention thresholds to learned safety predictions within the collaborative reward structure. The result is agents that internalize safety-aware selective prediction—especially valuable for embodied or operations settings (e.g., heterogeneous embodied teams; Liu et al., 2023) and high-stakes analysis pipelines (Plaat et al., 2025).

References:

1. Agentic Large Language Models, a survey. A. Plaat, M. V. Duijn, N. V. Stein, Mike Preuss, P. V. D. Putten, K. Batenburg (2025). arXiv.org.
2. Predictive Safety Network for Resource-constrained Multi-agent Systems. Meng Guo, Mathias Bürger (2019). Conference on Robot Learning.
3. Heterogeneous Embodied Multi-Agent Collaboration. Xinzhu Liu, Di Guo, Xinyu Zhang, Huaping Liu (2023). IEEE Robotics and Automation Letters.