"Agency Is Frame-Dependent"

Below podcast on this paper is generated with Google's Illuminate.

https://arxiv.org/abs/2502.04403

The paper addresses the challenge of defining agency in systems, questioning if agency is an inherent property or if it depends on how we observe the system.

This paper argues that agency is not absolute. Agency is always determined relative to a chosen "reference frame".

-----

📌 Agency determination in AI systems is not absolute. Evaluation metrics for Reinforcement Learning agents must explicitly define the "reference frame" to be meaningful and consistent.

📌 Frame-dependence implies AI benchmarks are implicitly biased by their chosen evaluation frame. Varying the reference frame during testing reveals an agent's true robustness across contexts.

📌 Designing AI agents requires acknowledging frame commitments. State space, reward function, and action definitions are frame choices impacting perceived agency and should be consciously designed.

----------

Methods Explored in this Paper 🔧:

→ This paper explores the concept of agency through the lens of Reinforcement Learning.

→ It examines the four key properties of agency: individuality, source of action, normativity, and adaptivity, as defined by Barandiaran et al.

→ The paper argues that each of these four properties is not absolute. Each property is dependent on a chosen "reference frame".

→ A reference frame includes commitments like defining system boundaries, choosing relevant causal variables, interpreting goals, and setting criteria for adaptation.

→ The paper uses philosophical arguments and examples from existing research in Reinforcement Learning, causal modeling, and adaptivity to support its frame-dependent view of agency.

-----

Key Insights 💡:

→ Agency is not an invariant property of a system.

→ Determining if a system has agency fundamentally depends on the chosen reference frame.

→ The four essential properties of agency – individuality, source of action, normativity, and adaptivity – are each individually frame-dependent.

→ The choice of reference frame is arbitrary, influencing whether a system is considered agentic.

-----

Results 📊:

→ Individuality is frame-dependent: Agent properties change based on chosen boundaries, as supported by Jiang (2019) and Harutyunyan (2020).

→ Source of action is frame-dependent: Agent detection in causal models varies with chosen causal variables, as shown by Kenton et al. (2023).

→ Adaptivity is frame-dependent: Whether a system is adaptive depends on the reference class for behavioral change, as argued by Zadeh (1963) and Abel et al. (2023).