Abstracts (Day 1)

Seminar Room 4: Philosophy of the Physical Sciences: Space and Time

Quasi-Local Approaches to Black Holes: From Global Ideals to Local Realities

Juliusz Doboszewski, Yichen Luo, Helen Meskhidze

What is a black hole? For relativists and mathematical physicists, it is often characterized by its event horizon, or the boundary of the causal past of future null infinity. But the event horizon is notoriously problematic. In this paper, we survey the philosophical literature surrounding event horizons to compile a list of desiderata. We use this list to evaluate quasi-local approaches to black hole horizons. In particular, we consider dynamical and isolated horizons and whether they represent successful de-idealizations of the event horizon. Finally, we consider whether these quasi-local horizons exhibit non-local behavior (i.e.,``clairvoyance’’). We argue that the characterization of these closed, trapped surfaces as clairvoyant is too critical. Though they might exhibit some non-local behavior, it does not undermine their fruitfulness or the prospects of the quasi-local approaches generally. We conclude by offering an optimistic outlook on the program and discussing open questions and ways forward.

Schwarzschild Black Holes & the Spacetime-Matter Dichotomy

Sanne Vergouwen, Niels Martens

What, if anything, makes the M parameter in the Schwarzschild metric a physical mass? In this paper we present and evaluate five possible interpretations of the mass of a Schwarzschild black hole. We argue that the concept of black hole mass is best understood by taking each of these five interpretations into account. We particularly show how a global interpretation of mass succeeds in referring to black hole mass, in spite of its significant difference from the traditional Newtonian notion of mass that is local. Finally, we argue not only a) that this global interpretation shows that the spacetime--matter dichotomy breaks down at the intensional level—in the sense that these conceptual categories become mixed—but also b) that the various interpretations suggest that the spacetime–matter dichotomy is best given up altogether, at least in the Schwarzschild context.

Diffeomorphism Invariant Quantities in Phase Space: More than Correlations

Alvaro Mozota Frauca

The mainstream view in the foundations of diffeomorphism-invariant theories is that their physical content is encoded in correlations or `observables’: phase-space functions thathave vanishing Poisson brackets with the constraints related to diffeomorphisms. I study the phase-space structure of models with a temporal diffeomorphism invariance and prove a series of formal results that challenge this view in a few ways. First, I show how this view is not applicable to all the phase space trajectories of every diffeomorphism invariant theory. Second, I show how correlations can be proved to be invariant only in a way that generalizes the standard definition and in a way that does not provide smooth functions. Third, I prove that spatiotemporal relations are also invariant. Fourth, I prove that spatiotemporal structures are indispensable fordefining the invariant content of diffeomorphism-invariant models. These results represent a challenge for some views in the foundations of general relativity and quantum gravity.

Seminar Room 6: Philosophy of the Life Sciences: Evidence, Normativity, and Precision Medicine

Evidence Synthesis via Formal Argumentation: Insights from the tobacco playbook

Juergen Landes, Paolo Baldi, Esther Anna Corsi, Hykel Hosni

Due to the presence of manufactured uncertainty, conflicts of interest, questionable research practices, and fraud, scientific evidence often cannot be taken at face value. The goal of this paper is to provide a proof of concept of how formal argumentation can be used to synthesise a body of evidence consisting of claims and counter-claims that may be affected by data imperfection. We argue that the theory of formal argumentation is a promising framework to take these issues into account by means of a stylised analysis of the so-called tobacco playbook.

What is “normal” about the “normal range”? Reference intervals, clinical decision limits, and varieties of normativity in laboratory medicine

Benjamin Chin-Yee

“Normal ranges” play a central role in contemporary biomedicine, yet their significance is contested. In laboratory medicine, the term has been replaced by “reference interval” (RI) to emphasize that RIs are not by themselves criteria for health or disease. Yet in practice, RIs are routinely treated as normatively significant, guiding judgments of abnormality and action.

Such uses are attributed to clinicians mistaking RIs for decision thresholds. This explanation, however, is insufficient. The difficulty stems from the capacity of RIs to support multiple, distinct forms of normativity.

I distinguish four forms relevant to RIs: statistical, salience, biological, and prescriptive normativity. Each relies on different justificatory practices and licenses different inferences. Many RIs warrant statistical and salience normativity; far fewer warrant biological or prescriptive normativity without additional justification. RIs, however, frequently invite stronger normative interpretations than they can properly support, a predictable consequence of a single representational tool serving multiple normative roles.

The Narrative Path Towards Precision

Hamed Tabatabaei Ghomi

Current precision medicine faces a paradox: striving for accuracy through traditional Evidence-Based Medicine (EBM) often increases clinical imprecision. By relying on stratified randomized controlled trials, “black box” algorithms, and reductionistic approaches, the field suffers from reduced statistical power, uninterpretable biomarkers, and a reductionist neglect of human complexity. This paper argues that the failure stems from applying general models to inherently unpredictable complex systems. Drawing on complex systems theory, I propose narrative precision medicine as a superior alternative. Unlike the data-centric EBM paradigm, narrative reasoning treats the patient as a unique trajectory rather than a population data point. By framing clinical care as a “coarse-grained simulation,” narrative reasoning integrates physiological, social, and personal causal factors into a temporal description. This shift from reductionist markers to narrative competence offers a “practically precise” framework capable of managing the non-linear dynamics of human health.

Seminar Room 7: Philosophy of Technology and AI: Mechanistic Interpretability and Explainability

What’s mechanistic in mechanistic interpretability?

Kristian Barman, André Curtis-Trudel

Mechanistic interpretability (MI) is an emerging research program in machine learning and AI safety that seeks to understand neural networks by identifying internal “components” and assigning them mechanistic and causal roles. Yet these components are virtual patterns in parameter and activation space rather than spatially discrete components capable of exchanging energy. This paper analyses MI case studies to clarify what, if anything, is genuinely mechanistic in such explanatory and causal claims, and what epistemic value mechanistic talk adds. We argue that existing neomechanist, computational-mechanist, and purely heuristic readings each capture something important but fail to fit MI practice as a whole. We develop a fourth, hybrid view on which MI operates with “fictional mechanisms”: structured, deliberately idealised models that misdescribe underlying ontology but, via well-calibrated links between representation and target (that capture counterfactual dependence), nonetheless support reliable interventions, and explanatory and predictive practices.

Concept Formation as the Basis of Explainable AI: Structural Conditions for Human–AI Understanding

Lee-Sun Choi

In this talk, I argue that the problem of explainable artificial intelligence cannot be resolved at the level of transparency or post-hoc justification alone. Explanation has an internal structure composed of concepts, and explanatory success requires that these concepts be understood. When artificial systems form concepts that are not translatable into human concepts, explanations articulated in those terms fail to be intelligible. I reframe explainable AI as a problem of conceptual translatability rooted in concept formation. I develop an integrated account of world representation that combines distributed and probabilistic representations, according to which concepts are individuated by structural properties of probability distributions over feature spaces, such as modality, density, and spikeness. Structural similarity across concept formation processes enables translation under conceptual pluralism, allowing multiple representations of the same world without collapsing into solipsism. On this view, explainability is the operational manifestation of successful conceptual translation between human and artificial intelligence.

Constructivism and the Binding Problem in AI

Xiaotong Li

In recent years, Artificial Neural Networks  have exhibited a series of compositional defects in multimodal tasks, including “superposition catastrophe” and “mismatching” in multi-objects representations. These phenomena constitute the “compositionality problem” . Empirical evidence suggests that  compositional failures are rooted in the model’s inability to perform “binding. “ Specifically, the “binding problem” refers to the ability to achieve reconfigurable structural representations within a distributed representation system. 

This study argues that the binding problem is an inherent lack of capability rather than a “patchable error”, requiring a conceptual reconstruction necessitated by  empirical failures. In this regard, existing theories face a dual dilemma. Structural Reductionism suffers from representational rigidity, while Weak Emergence fails to account for the downward causation required for logical reconfiguration. I propose Constructivism, treating binding as a “structurally constructive activity.” This approach provides policy-useful guidelines for future architectures by demonstrating that building binding capability requires  structural design rather than mere parameter expansion.

Seminar Room 8: General Philosophy of Science: Causal Inference

Causes Don’t Push

Angela Potochnik

Complex systems approaches have proved useful in a wide range of scientific fields. Here I explore implications of these approaches for our very concept of causation. Philosophical accounts of causation—physical and counterfactual alike—assume causation can be conceptualized as isolated direct influence. I conjecture that this model of causation, what I term billiard-ball causation, is deeply misleading. Instead, consideration of complex systems methods and the uses and limitations of experimental practices and modeling techniques supports a different conception of causation, what we might think of as a causal mesh. Shifting our conception of causation from billiard balls to a causal mesh clarifies some features and limitations of scientific techniques and resolves some philosophical puzzles about causation, especially puzzles that have emerged for the popular manipulability conception of causation.

Causal Inference Beyond Evidence-Based Medicine: A Bayesian Appraisal of Mechanistic Plausibility

Ethan Vorster

EBM prioritises statistical evidence from RCTs when assessing causal claims about medical interventions. While effective in pharmaceutical contexts, it encounters difficulties when applied to interventions whose effects are realised through complex engineered systems. This paper introduces machine-based medicine as a category of intervention characterised by strong biophysical rationale and mechanistic evidence, but limited statistical evidence. Using particle-beam therapy as a case study, I examine how existing evidential hierarchies struggle to accommodate interventions that exhibit high causal plausibility yet resist conventional trial-based evaluation. Three approaches to causal appraisal are analysed: EBM, EBM+, and Bayesianism. I argue that while pluralist frameworks successfully underscore the epistemic relevance of mechanistic evidence, they lack the resources for systematic evidence integration. A Bayesian framework, by contrast, offers a transparent and principled method for combining mechanistic and statistical evidence. The paper contributes to ongoing debates on confirmation, causal inference, and evidential pluralism in healthcare.

Measuring Implicit Bias: On the Applicability of the Vera Causa Ideal in Social Science

Karim Bschir

Implicit bias is a frequently invoked explanatory concept in social-psychological studies of discriminatory and prejudiced social behavior. A common way to detect implicit attitudes is via indirect measurement, e.g. in the so-called Implicit Association Test (IAT) developed in the late 1990s. In this paper, I contribute to the ongoing debate about the validity of indirect measures of implicit attitudes by referring to the so-called vera causa ideal that played a historically important role in the assessment of theoretical entities primarily in biology or the earth sciences. I discuss the applicability of the ideal to studies of implicit attitudes, insofar as they aim at causal explanations of empirical phenomena (e.g. prejudiced behavior) by reference to the causal agency of a theoretical entity (e.g. implicit bias).

Seminar Room 9: General Philosophy of Science: Confirmation and Induction

Confirmation is Value-laden

Adrià Segarra

In this article, I delineate a category of legitimate non-epistemic value influences in science that the current literature has largely overlooked or dismissed. In particular, I argue that confirmation, understood here as the logical relation of inductive support, is value-laden. In order to show this, I examine the main accounts of confirmation in the literature and expose some of the points at which non-epistemic values must play a role. I conclude that the project of managing non-epistemic values in science must extend to the very foundations of inductive logic.

Induction without Rules

Kevin Davey

It is commonly assumed that science depends on a rule of induction that lets us infer unobserved facts from observed ones. While thinkers like Norton propose a patchwork of local inductive rules, I argue that even this view is implausible. Instead, I suggest we abandon the notion of inductive rules altogether. Claims about the unobserved are justified directly by background knowledge and observation, without the need for overarching rules. Even in resolving doubts or disputes about induction, careful review or additional evidence—rather than rules—is typically sufficient, both in practice and in theory.

Seminar Room 4: Philosophy of the Physical Sciences: Thermal Physics and Information Theory

Thermodynamics at Small Scales: A Case for Stochastic Thermodynamics at Strong Coupling 

Aditya Jha

I argue that extending thermodynamics to small systems exposes background assumptions that are largely invisible in its traditional macroscopic domain. Using mesoscopic stochastic thermodynamics, I examine the underappreciated weak-coupling idealization—negligible system–environment interaction—which normally secures an unambiguous split between heat and work. Drawing on a current debate (Talkner & Hänggi 2016; Seifert 2016; Jarzynski 2017), I show that in the strong-coupling regime this demarcation becomes partly conventional and renders heat especially ambiguous. The key reason is that work is then defined via the Hamiltonian of Mean Force for an effective system, rather than by changes in the system’s bare Hamiltonian, which generally differs from the HMF. I then assess contrasting philosophical responses amongst physicists concerning this (dis)ambiguation—essentialism versus functionalism—found respectively in Pucci et al. (2013), Cohen & Mauzerall (2004), Talkner & Hänggi (2016, 2020), and in Jarzynski (2007, 2017) and Seifert (2016).

New Balances: Finding Equilibrium in Integrable Quantum Statistical Mechanics

Eugene Chua and Yichen Luo

Philosophers have worried that some systems -- black holes (Dougherty & Callender 2016), self-gravitating systems (Robertson, 2019), relativistic systems (Chua & Callender forthcoming) -- are not really thermodynamic. We argue that a similar breakdown occurs in quantum statistical mechanics. Classically, and (typically) in the quantum domain of non-integrable systems: macroscopic observables relax to microcanonical values, small subsystems look Gibbsian, finite temperatures track mutual equilibrium, and equilibration washes out initial conditions. However, in the fuller quantum domain, much of this package fails. Quantum `scarring’ preserves nonthermal behavior for atypical initial states even for non-integrable systems; integrable and many-body localized systems relax to generalized Gibbs ensembles constrained by additional (quasi-)conserved quantities, and temperature alone cannot play its Zeroth-Law role; driven Floquet systems approach infinite-temperature states. However, we argue that equilibrium reasoning -- generally construed -- can continue by finding the right scales, observables, and suitable regime-specific conditions of applicability, even when classical thermodynamics itself no longer applies.

Relational Information: Towards a New Kind of Information in Quantum Mechanics

Niccolò Covoni, Marco Sanchioni

This talk argues that von Neumann entropy plays two conceptually distinct roles in quantum theory. When applied to global mixed states, it expresses the quantum analouge of the informational entropy. But when applied to reduced states of entangled systems, it measures objective physical correlations. We propose that this second usage realizes a new informational kind, which we call relational information. Unlike information in communication theory, it reflects structural interdependence between systems, not probability about a certain outcome. We further suggest that this informational kind has ontological significance: relational information is a physically instantiated feature of entangled systems, and not merely an agent-relative concept.

Seminar Room 6: History and Philosophy of Science: Integrated HPS and Methodology

Inventing Risk: A Subjectivist Perspective

Caterina Sisti, Luca Zanetti

This paper examines the notion of objective probability, or chance, in the works of early subjectivists, particularly Frank Ramsey and Bruno de Finetti. To this end, we propose a clarification and characterisation of Ramsey’s notion of chance and of its objectivity as agreement among agents and convergence of their degrees of belief, drawing on an independent suggestion by de Finetti regarding Knight’s distinction between risk and uncertainty. The interpretation we propose allows us to preserve the epistemic and pragmatic spirit of subjectivism while explaining how a meaningful sense of objectivity can emerge within it, a feature that is still relevant in contemporary debates on the foundations and applications of probability.

An integrated HPS approach to methodological ideals in linguistic relativity research 

Ilir Isufi

The linguistic relativity hypothesis states that language influences thought, so that speakers of different languages think differently about the world. While this thesis was initially put forward within the field of anthropology, the cognitive revolution of the 1950s and the subsequent formation of the field of cognitive science dramatically changed the scene. Nowadays, most cognitive science is conducted by experimental psychologists. This has also impacted linguistic relativity research, which has moved from anthropology to experimental psychology in the lab, following specific research methods. According to these, language needs to be isolated from culture, which is treated as a confounding variable, so that language-specific causal contributions to cognition can be identified. I argue that replication failures in linguistic relativity research are partly the result of historically contingent methodological ideals inherited from experimental psychology. An integrated HPS perspective reveals why anthropologically grounded approaches have been systematically marginalized despite their epistemic advantages.

The Practical Function of Modern HPS

Fons Dewulf

Ever since the inception of HPS, its scholars have disagreed about its aims and methods. In these discussions, one mostly finds a similar origin story of HPS, focusing purely on intellectual motivations: in the early 1960s, Hanson, Feyerabend and Kuhn used the history of science to make philosophical theories of science concrete, countering the logical empiricists’ empty formal models. Supposedly, these efforts cemented widespread interest in HPS.

I will argue that this classical origin story is misguided for two reasons. First, it neglects important efforts by atom bomb scientist just after the Second World War. Second, it obscures the broader socio-cultural motivations for universities to invest in HPS. I point out that HPS emerged primarily due to its potential as a mediator between science and its socio-cultural environment. HPS, I argue, should take this cultural context of its own emergence to heart in discussions on its aims and methods.

Seminar Room 7: Philosophy of the Cognitive Sciences: Perception and Imagination

Do Androids Imagine Electric Sheep? Radical Simulationism and AI Imagination

Jake Hawthorne

According to orthodox simulationism, episodic memory is a past-directed form of episodic imagination. For a subject S to remember an episode e, S must possess a representation of e, and this representation must be produced by a construction system that is properly functioning, episodic, and aimed at producing a representation from S’s personal past. Recently, Michaelian (2024) has developed radical simulationism, which abandons the personal past condition. This is to remove the ambiguity Michaelian sees in the notion of the personal past. In this paper, I object that this entails that many of the videos produced by video-generating artificial intelligences (VGAIs) are imaginings. This misconstrues episodic imagination, as these videos do not possess an essential feature of episodic imagination—namely, autonoesis. I conclude by suggesting that the simulationist can avoid the ambiguity associated with the personal past by incorporating a representation of the subject’s body into it.

Illusionism in Intensity Discrimination: a psychophysical analysis of the user-illusion

Hyungrae Noh

This paper develops a psychophysical analysis of Dennett’s user-illusion account by examining intensity discrimination under the Weber–Fechner law. Qualia realists assert that distinct phenomenal states within a single sensory dimension (e.g., ‘100g-quale’ vs. ‘200g-quale’) are necessary for perceiving stimulus differences. Drawing on cross-species data—from human psychophysics to bacterial chemotaxis, both captured by the Weber–Fechner law—I show that logarithmic encoding of intensity arises from non-conscious algorithmic processes without invoking qualia. I then present a debunking argument for illusionism: an independent explanation of our phenomenological intuitions demonstrates that, if qualia spectra truly underlay discrimination, their systematic correlation with behavioral thresholds would constitute an implausible coincidence. Given this cross-species evidence, the phenomenological intuitions are debunked: neural mechanisms of intensity discrimination operate independently of qualia. Thus, sensory discrimination exemplifies Dennett’s view that consciousness is a representational interface lacking causal power, motivating a targeted elimination of qualia realism in perception.

A Geometrical Account of Imaginative Resistance

Jianlang Yin

Imaginative resistance refers to the difficulty in imagining scenarios such as morally deviant worlds or logical contradictions. Existing accounts either have a limited scope or appeal to unexplained inferential mechanisms shared between belief and imagination. Drawing on Gardenfors’ theory of conceptual spaces, I propose a unified geometrical explanation: imaginative resistance arises from the difficulty of voluntarily constructing representations that violate the geometrical constraints of our conceptual spaces. 

I explain three types of resistance: Constitutive resistance arises from separating integral dimensions or adding unrelated ones (e.g., imagining pain without intensity or a green melody). Topological resistance occurs when a representation must occupy disjoint regions simultaneously (e.g., imagining something red and green all over). Mapping resistance arises from distorting established mappings between different spaces (e.g., imagining killing a baby is morally right). By explaining resistance geometrically, this account unifies moral and non-moral cases and explains why resistance is graded rather than absolute.

Seminar Room 8: General Philosophy of Science: Models, Data, and Simulation

Beyond “More or Less” Detail: Abstraction Levels in Scientific Explanation

Jietong Xu

Evaluation of Bogen-Woodward Framework: Contextualising Phenomenon-Data Relationship in Scientific Enquiries

Varun Bhatta

Bogen and Woodward (BW) argued that scientific theories do not explain data, but only phenomena (1988). Woodward (1989) further clarified that phenomena are not inferred “downward” from theory, but rather “upward” from data. The BW framework has received a spectrum of responses. Some scholars like Schindler (2007), Tal (2011) and Lusk (2021) have identified case studies that challenge Woodward’s proposal that phenomenon-identification is only upward, from data to phenomena. None of these critics, however, further clarifies why the BW-framework fail in these cases. I argue that the current confusion is due to not contextualising inferences in specific enquiries. Woodward’s analysis of his claim is situated in hypothesis verification. In contrast, Schindler’s and Tal’s case studies are situated in other contexts and in these, upward/downward claims do not make sense. Through this analysis, I identify a few drawbacks in the current approach to analyse scientific practice through the case-study methodology.

The Delusion of Simulations as Higher-Order Evidence: A Comparative Study on Rarity-Based Cosmology

Tianzhe Cozette Shen

Computer simulations play a significant role as higher-order evidence in scientific practice. In contemporary cosmology, large-scale cosmological simulations are oftentimes used to test a cosmological model, commonly by locating analogs of rare celestial events/phenomena/systems, like a colliding galaxy cluster. Multiple studies based on different simulations were conducted to locate analogs of the Bullet Cluster, while the outcomes are conflicting. By reconstructing their logic(s), I argue that this method’s logic is inherently flawed, and accordingly, simulations as higher-order evidence can be delusional mainly for two reasons: (1) there is no proper measure of similarity or resemblance to our universe, and (2) the idea of rarity is not quantitatively well-defined. We can meaningfully compare neither these studies with each other, nor one such study with real observations, with generalizability to other sciences like (paleo)ecology and (paleo)climatology. I also consider alternative approaches to rarity-based simulation studies.

Seminar Room 9: Philosophy of Technology and AI: Machine Speech, Content Alignment, and LLMs

Machine speech: the very idea

Matthew McKeever & Xindi Ye-Martinsky

This paper is about the question of machine speech—whether sophisticated machines, such as LLM-based chatbots, speak. I argue that responses to the question of machine speech fall into two camps: the deepists, who say that internal architecture and processes determine whether a machine speaks (Block 1981, Bender and Koller 2020, Chomsky 2023, Shanahan 2023), and the shallowists, who say that external behaviour determines whether a machine speaks (Turing 1950, Dennett 1987, Chalmers 2023). Which side is right? I argue that both deepists and shallowists latch on to important aspects of how we think about speech. But given that the meaning of words like ‘speaks’ is too underdetermined, it’s unlikely that we will come to consensus about the question of machine speech via descriptive analysis alone. What we should focus on, instead, is a normative analysis of speech—of what speech should be, or what ‘speaks’ should mean.

The problem of human-AI content alignment

Nikolaj Jang Lee Linding Pedersen

The issue of how to align the values implemented by AI systems with human values has received a great deal of attention in AI research, as it is regarded as a key component of AI safety. Aligning AI values with human values will help mitigate AI risk. This talk introduces the problem of human-AI content alignment and explores its connection to value alignment. Let ‘human content’ denote content featured in human minds or expressed by human language, and let ‘AI content’ denote content configured by AI states or expressed by AI outputs. How can alignment between human content and AI content be achieved, and why is it important? This talk does three things. First, examples of content misalignment are provided. Second, it is argued that failures of content alignment sometimes bring with them failures of value alignment. Third, considerations on how to address the problem of content alignment are offered.

Large Language Models are Fundamentally Next Token Predictors

Yue Ma

The prevailing consensus posits that Large Language Models (LLMs) are next token predictors (NTPs). Recently, three philosophers have challenged this view based on two central ideas. First, they argue that characterizing models merely as NTPs is a significant underestimation. Second, based on a deconstruction of the training process, they advance the Function and Explanation Objections. I aim to defend the Grounding View—the thesis that LLMs are fundamentally NTPs. This position acknowledges that models possess emergent properties, but maintains that the generation of these properties remains grounded in the process of token prediction. I argue that only the latter idea constitutes a potential rebuttal to my position. Furthermore, I contend that objections based on the training process are mistargeted. I will illustrate my point using the “Swamp Model” thought experiment.