Calum McNamara

Understanding an agent's goals helps explain and predict its behaviour, yet there is no established methodology for reliably attributing goals to agentic systems. We propose a framework for evaluating goal-directedness that integrates behavioural evaluation with interpretability-based analyses of models' internal representations. As a case study, we examine an LLM agent navigating a 2D grid world toward a goal state. Behaviourally, we evaluate the agent against an optimal policy across varying grid sizes, obstacle densities, and goal structures, finding that performance scales with task difficulty while remaining robust to difficulty-preserving transformations and complex goal structures. We then use probing methods to decode the agent's internal representations of the environment state and its multi-step action plans. We find that the LLM agent non-linearly encodes a coarse spatial map of the environment, preserving approximate task-relevant cues about its position and the goal location; that its actions are broadly consistent with these internal representations; and that reasoning reorganises them, shifting from broader environment structural cues toward information supporting immediate action selection. Our findings support the view that introspective examination is required beyond behavioural evaluations to characterise how agents represent and pursue their objectives.

This dissertation is about the role that conditionals play in uncertain reasoning and deliberation. Specifically, I attempt to show that, by appealing to a particular semantics for conditionals---a contexutalist, sequence semantics, which has recently become popular in philosophy of language---several open problems in decision theory and epistemology can be solved. Chapter 1 is introductory. I set out the semantic view of conditionals in question, and I describe some of its historical background. Chapter 2 turns to a striking problem faced by causal decision theorists. A popular formulation of causal decision theory (CDT) appeals to counterfactual conditionals. However, the standard theory of these conditionals has unintuitive consequences in deterministic worlds. In particular, it says that if anything---including the choice you make---were different in the present, then either the laws or nature would be violated, or the distant past would be changed. And as several authors have recently shown, it's easy to transform this consequence of the standard theory of counterfactuals into full-blown counterexamples to CDT. In response to these counterexamples, I develop a contextualist version of CDT, which makes use of the sequence semantics. I then show that the deterministic counterexamples don't arise for my version of CDT. In Chapter 3, I deal with a different puzzle, about whether or not the so-called Desire-as-Belief (DAB) thesis is consistent with decision theory---something that famous arguments of David Lewis seem to show isn't the case. Once again, I show that, if we understand the DAB thesis in a contextualist way---and spell it out using the sequence semantics---then Lewis's arguments against that thesis don't go through. In fact, we can prove a tenability result for the DAB thesis, which shows that it's compatible with decision theory after all. Finally, in Chapter 4, I transition from decision-theoretic issues to epistemological ones. More precisely, I tackle the question of how our credences should change when we learn indicative conditionals. Several famous cases in the literature---notably, Bas van Fraassen's Judy Benjamin problem---seem to show that the standard Bayesian update rules deliver implausible results when we learn conditionals of this kind. However, in the chapter, I show that, if we adopt the sequence semantics, then the Bayesian update rules turn out to deliver the correct results after all. Better still, alternatives to these rules which have been put forward in the literature turn out to be equivalent to the Bayesian rules in my framework---at least in many contexts. Thus, what we end up with is a nice, unified account of how rational agents should update on conditional information: one which fits in well with recent work on the semantics of conditionals. My proposal also relates, in interesting ways, to discussions that have been happening elsewhere in the literature, like discussions about the tenability of the notorious Stalnaker's thesis.

The classic formulation of causal decision theory (CDT) appeals to counterfactuals. It says that you should aim to choose an option that would have a good outcome, were you to choose it. However, this version of CDT faces trouble if the laws of nature are deterministic. After all, the standard theory of counterfactuals says that, if the laws are deterministic, then if anything—including the choice you make—were different in the present, either the laws would be violated or the distant past would be changed. And as several authors have shown, it's easy to transform this upshot of the standard theory of counterfactuals into full-blown counterexamples to CDT. In response to these counterexamples, I argue here that the problem lies, not so much with CDT's guiding idea—that it's the expected causal consequences of your actions that matter for rational decision-making—but with the fact that the classic formulation of CDT doesn't pay sufficient attention to the context-sensitivity of counterfactuals. I develop a contextualist version of CDT which better accounts for this context-sensitivity. And I show that my theory avoids the problems faced by the classic formulation of CDT in determinstic worlds.

Our scientific theories, like our cognitive structures in general, consist of propositions linked by evidential, explanatory, probabilistic, and logical connections. Those theoretical webs ‘impinge on the world at their edges,’ subject to a continuing barrage of incoming evidence. Our credences in the various elements of those structures change in response to that continuing barrage of evidence, as do the perceived connections between them. Here we model scientific theories as Bayesian nets, with credences at nodes and conditional links between them modelled as conditional probabilities. We update those networks, in terms of both credences at nodes and conditional probabilities at links, through a temporal barrage of random incoming evidence. Robust patterns of punctuated equilibrium, suggestive of ‘normal science’ alternating with ‘paradigm shifts,’ emerge prominently in that change dynamics. The suggestion is that at least some of the phenomena at the core of the Kuhnian tradition are predictable in the typical dynamics of scientific theory change captured as Bayesian nets under even a random evidence barrage.

We model scientific theories as Bayesian networks. Nodes carry credences and function as abstract representations of propositions within the structure. Directed links carry conditional probabilities and represent connections between those propositions. Updating is Bayesian across the network as a whole. The impact of evidence at one point within a scientific theory can have a very different impact on the network than does evidence of the same strength at a different point. A Bayesian model allows us to envisage and analyze the differential impact of evidence and credence change at different points within a single network and across different theoretical structures.