An AI can learn complex safety rules just by being told yes or no when it makes a mistake.
April 29, 2026
Original Paper
Discovering Agentic Safety Specifications from 1-Bit Danger Signals
arXiv · 2604.23210
The Takeaway
We usually assume that to teach an AI a rule, we need to provide a detailed textual explanation of what it did wrong. This study shows that LLM agents can discover intricate safety constraints using only a single binary danger bit as feedback. The model can figure out that entering from the north is dangerous without any words describing the north or the danger. This is similar to how biological organisms learn to avoid pain without needing a manual. It suggests that AI safety can be built into agents using much simpler and more universal signals than we currently use.
From the abstract
Can large language model agents discover hidden safety objectives through experience alone? We introduce EPO-Safe (Experiential Prompt Optimization for Safe Agents), a framework where an LLM iteratively generates action plans, receives sparse binary danger warnings, and evolves a natural language behavioral specification through reflection. Unlike standard LLM reflection methods that rely on rich textual feedback (e.g., compiler errors or detailed environment responses), EPO-Safe demonstrates th