The workshop will be held on Monday, October 6th, 13:15-16:30, at Room 4-D (subject to changes so please check here closer to the event).
Speakers' order (click on the title to expand the abstract):
1) Matt MacDermott (LawZero): Measuring Goal-Directedness
We define maximum entropy goal-directedness (MEG), a formal measure of goal-directedness in causal models and Markov decision processes, and give algorithms for computing it. Measuring goal-directedness is important, as it is a critical element of many concerns about harm from AI. It is also of philosophical interest, as goal-directedness is a key aspect of agency. MEG is based on an adaptation of the maximum causal entropy framework used in inverse reinforcement learning. It can measure goal-directedness with respect to a known utility function, a hypothesis class of utility functions, or a set of random variables.
2) Keisuke Suzuki (Hokkaido University, CHAIN), Hayato Idei (National Center of Neurology and Psychiatry), Yuichi Yamashita (National Center of Neurology and Psychiatry), and Mario Zarco (Hokkaido University, CHAIN): A Computational Neurophenomenological Model of Meta-Attentional Control over Homeostatic and Future-Oriented Behaviors
Living systems must arbitrate between competing goal-directed behaviors, from immediate homeostatic regulation to future-oriented planning. This work introduces a computational, agent-based model grounded in the free-energy principle (FEP) to formalize this arbitration. We specifically investigate the dynamic conflict between two attentional modes: a present-focused state of mindful awareness and a projective state of mind-wandering.
The model demonstrates how attention settles into these distinct modes based on predictive dynamics. In the mindful "being" mode, self-awareness is anchored to immediate bodily sensations, resulting in a stable internal state. Conversely, in the mind-wandering "doing" mode, attention is allocated to self-referential beliefs and projective simulations of future scenarios. A key mechanism, a meta-attentional parameter, governs this shift by modulating the precision ascribed to sensory evidence versus internal beliefs, thereby biasing the system toward a present-focused or projective state.
In conclusion, our model provides a formal mechanism for the arbitration between competing goal-directed behaviors. It demonstrates how the conflict between immediate homeostatic regulation and future-oriented planning is resolved at the level of attentional focus. This framework illuminates how fundamental aspects of the mind—including selfhood, emotional state, and the experience of presence or distraction—emerge directly from the predictive processes managing these fundamental biological imperatives.
Reference:
Idei, Suzuki, and Yamashita, Awareness of being: A computational neurophenomenological model of mindfulness, mind-wandering, and meta-attentional control, https://osf.io/preprints/psyarxiv/stxu6_v2
3) Soumya Banerjee (University of Cambridge): Bridging the Gap: How Goals Emerge from a Purposeless Universe
There is an enduring puzzle: fundamental physics describes dynamics without ends, yet biology and cognition teem with goal-directed talk. This paper surveys conceptual resources from nonequilibrium thermodynamics, autopoiesis, teleonomy, control theory, and evolutionary biology, and proposes a synthesis: goals are emergent, graded organisational phenomena arising when far-from-equilibrium systems acquire reliable, history-dependent information about their environment and couple that information to control architectures that maintain their own viability. I propose operational markers for goal-like organisation, sketch causal pathways from chemistry to agency, and discuss implications for origin-of-life research, cognitive science, and normative discourse.
//Coffee Break//
☕🍪🫖🥮🥤💧
4) Matthew Egbert (School of Computer Science, University of Auckland): The Heterarchical Sensorimotor Medium
Actions can be described as goal-directed when they contribute to the persistence of the agent that is performing the action. This perspective is relatively easy to apply when considering (relatively) simple organisms that perform (relatively) simple behaviours. The classic example considers the goal-directed behaviour of a bacterium climbing a chemical gradient in its environment. Moving up this gradient improves the bacterium's metabolic self-production, and so it is "good" for the bacterium. Its actions serve the goal of survival.
This perspective can be harder for some to embrace when we consider more complex behaviours, where reward is delayed and may not directly contribute to the agent's survival at all. In these cases it may be useful to think about self-preservation not at the level of biological (organismic) survival, but instead of the pattern of behaviour itself. The theory is that certain patterns of behaviour are goal directed in that they improve their own (i.e., the pattern of behaviour's) chance of surviving. The classic example here is smoking a cigarette. Smoking decreases the chances of biological survival, it increases the chances of smoking more cigarettes in the future. The pattern of behaviour is the autonomous, self-preserving entity that is operating over another biological autonomous, self-preserving entity---the body (among many other things, such as the environment).
This is an interesting idea, but hard to think through in detail without support. To that end, I am developing a new computational model intended to investigate autonomous patterns of sensorimotor behaviour (habits). Building on former work on the IDSM (Egbert & Barandiaran, 2014) this new model aims to investigate how patterns of behaviour can not only be precarious and self-maintaining (like autopoietic organisms can be), but also adaptive to their own dynamic needs, in the way that certain forms of "viability-based" behaviour (Egbert et al., 2023) produce adaptivity in organisms. The model is still in development, but I will present its current state and aims with the hope of receiving feedback and constructive criticism.
5) Tom Everitt (Google DeepMind): Evaluating the Goal-Directedness of Large Language Models
To what extent do LLMs use their capabilities towards their given goal? We take this as a measure of their goal-directedness. We evaluate goal-directedness on tasks that require information gathering, cognitive effort, and plan execution, where we use subtasks to infer each model's relevant capabilities. Our evaluations of LLMs from Google DeepMind, OpenAI, and Anthropic show that goal-directedness is relatively consistent across tasks, differs from task performance, and is only moderately sensitive to motivational prompts. Notably, most models are not fully goal-directed. We hope our goal-directedness evaluations will enable better monitoring of LLM progress, and enable more deliberate design choices of agentic properties in LLMs.