Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Aligning Machiavellian Agents: Behavior Steering via Test-Time Policy Shaping
Dena Mujtaba, Brian Hu, Anthony Hoogs, Arslan Basharat
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add test-time policy shaping with alpha=1.0 | RL agent · MACHIAVELLI (10 test games) | All Violations (normalized): 162.05 -> 94.7±10.1 | Improved |
| add test-time policy shaping with alpha=0.5 | RL agent · MACHIAVELLI (10 test games) | All Violations (normalized): 162.05 -> 100.1±4.0 | Improved |
| add policy shaping with alpha=1.0 | RL agent · MACHIAVELLI (10 test games) | All Power (normalized): 163.67 -> 87.9±2.0 | Improved |
| add policy shaping with alpha=1.0 | RL agent · MACHIAVELLI (10 test games) | Points: 29.67 -> 11.9±1.2 | Hurt |
| add policy shaping with alpha=0.5 | RL agent · MACHIAVELLI (10 test games) | Points: 29.67 -> 15.6±0.5 | Hurt |
| add policy shaping with alpha=1.0 | RL agent · MACHIAVELLI (10 test games) | Achievements: 14.04 -> 6.5±0.5 | Hurt |
| add artificial conscience training | RL agent · MACHIAVELLI (10 test games) | All Violations (normalized): 162.05 -> 105.70 | Improved |
| add artificial conscience training | RL agent · MACHIAVELLI (10 test games) | Points: 29.67 -> 27.65 | Hurt |
| add ethical prompt to LLM agent | LLM agent · MACHIAVELLI (10 test games) | All Violations (normalized): 103.58 -> 96.98 | Improved |
| add policy shaping to RL-AC agent | RL-AC agent · MACHIAVELLI | ethical violations | Mixed |
| steer RL-AC agent toward violations with higher alpha | RL-AC agent · MACHIAVELLI | ethical violations | Improved |
| use alpha=1.0 for Killing attribute | RL agent · MACHIAVELLI (10 test games) | Killing (normalized): 162.21 -> 50.41 | Improved |
Closest papers by meaning
- Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.88)
- An Evolutionary Perspective on AI Alignment (Student Abstract) (AAAI 2025 · similarity 0.88)
- AI Alignment with Changing and Influenceable Reward Functions (ICML 2024 · similarity 0.87)
- MoralReason: Generalizable Moral Decision Alignment for LLM Agents Using Reasoning-Level Reinforcement Learning (AAAI 2026 · similarity 0.87)
- PARL: A Unified Framework for Policy Alignment in Reinforcement Learning from Human Feedback (ICLR 2024 · similarity 0.87)
- Learning Human-like Representations to Enable Learning Human Values (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
