Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Parametrically Retargetable Decision-Makers Tend To Seek Power
Alexander Matt Turner, Prasad Tadepalli
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 |
|---|---|---|---|
| Use a decision procedure that can be retargeted from staying in the initial room to leaving it | Greedy final-observation reward maximization in Montezuma's Revenge · Montezuma's Revenge, step-T observation | preference for leaving over staying | Improved |
| Use ϵ-greedy exploration to learn an exploitation policy | Bandit algorithm for Montezuma's Revenge initial actions · Montezuma's Revenge initial action | retargetability multiplicity from ↓ to other actions: 4 | Improved |
| Increase training trials so the learned exploitation policy is more likely to try every action | ϵ-greedy bandit algorithm · Montezuma's Revenge initial action | approximation to greedy reward maximization | Improved |
| Use DQN rather than a training process capable of exploring beyond the first room | Montezuma's Revenge policy training · Leaving the first room | retargetability away from the first room | Hurt |
| Enable exploration of the first three rooms to the right of the initial room | RL training with featurized item rewards · Montezuma's Revenge, leaving the first room | retargetability away from the first room | Improved |
| Increase exploration and ability to remember routes to item configurations | GO-EXPLORE-style training · Montezuma's Revenge, entire level | retargetability across final inventory outcomes | Improved |
Closest papers by meaning
- Emergent Risk Awareness in Rational Agents under Resource Constraints (NeurIPS 2025 · similarity 0.87)
- Learning Utilities from Demonstrations in Markov Decision Processes (ICML 2025 · similarity 0.85)
- AI Alignment with Changing and Influenceable Reward Functions (ICML 2024 · similarity 0.84)
- Sequential Decision Making in Stochastic Games with Incomplete Preferences over Temporal Objectives (AAAI 2025 · similarity 0.84)
- Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.84)
- Observation Interference in Partially Observable Assistance Games (ICML 2025 · similarity 0.84)
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.
