Understanding And Controlling A Maze-solving Policy Network

Abstract

To understand the goals and goal representations of AI systems, we carefully study a pretrained reinforcement learning policy that solves mazes by navigating to a range of target squares. We find this network pursues multiple context-dependent goals, and we further identify circuits within the network that correspond to one of these goals. In particular, we identified eleven channels that track the location of the goal. By modifying these channels, either with hand-designed interventions or by combining forward passes, we can partially control the policy. We show that this network contains redundant, distributed, and retargetable goal representations, shedding light on the nature of goal-direction in trained policy networks.

Stats

• citations0
• S2 citations—
• github stars0
• HF likes0
• heat score0.00
• arxiv keymini2023understanding

Related papers

• Backward Learning For Goal-conditioned Policies (2023)0.00
• Towards Measuring Goal-directedness In AI Systems (2024)0.00
• Learning When To Switch: Adaptive Policy Selection Via Reinforcement Learning (2025)0.00
• Open-world Multi-task Control Through Goal-aware Representation Learning And Adaptive Horizon Prediction (2023)8.09
• Three Pathways To Neurosymbolic Reinforcement Learning With Interpretable Model And Policy Networks (2024)0.00
• GRAIL: Goal Recognition Alignment Through Imitation Learning (2026)0.00
• Real-world Human-robot Collaborative Reinforcement Learning (2020)9.41
• A Deep Policy Inference Q-network For Multi-agent Systems (2017)0.00