Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts
A method for evolving AI agent harnesses via counterfactual guidance may improve robustness but also raises concerns about more capable autonomous agents.
A method for evolving AI agent harnesses via counterfactual guidance may improve robustness but also raises concerns about more capable autonomous agents.
Auto-scored: A capability advance in agent harness evolution, with minor potential for misuse but no direct escalation.