Pierre Beckmann

I am a researcher working in Philosophy of AI and Philosophically-motivated AI. I am currently a PhD student in the Neuro-Symbolic AI group of Idiap and EPFL, working for the M-RATIONAL project. I am also a MATS Scholar. I keep close ties to the Institute for Philosophy of Bern.

Email / Scholar / Twitter / CV / ORCID

Research

My research bridges philosophy and AI, with a focus on LLMs. In philosophy, I draw on mechanistic interpretability to study understanding and individuation in AI systems. At EPFL, I work on neuro-symbolic approaches to reasoning in LLMs, with an upcoming focus on moral reasoning.

Philosophy of AI
	Where is the Mind? Persona Vectors and LLM Individuation Pierre Beckmann, Patrick Butlin arXiv preprint arXiv / thread We argue that the strongest candidates for minds in LLMs are the virtual instance view and two novel persona-based views, drawing on mechanistic interpretability and persona vectors.
	Mechanistic Indicators of Understanding in Large Language Models Pierre Beckmann, Matthieu Queloz Philosophical Studies, 2026 paper Building on mechanistic interpretability, we argue that LLMs exhibit signs of understanding — across three tiers: conceptual, state-of-the-world, and principled understanding.
	Why We Care About Understanding: Competence through Predictive Compression Matthieu Queloz, Pierre Beckmann PhilPapers preprint paper By reverse-engineering the function of both the state and the concept of understanding, we develop a compression account: understanding arises from pressures to build predictive models that are accurate yet compressed enough to store, demonstrate, and transmit.
	New Horizons in Machine Understanding: Explanatory and Objectual Understanding in Deep Learning Video Generation Models Pierre Beckmann Synthese, 2025 paper Using explanatory and objectual understanding, I evaluate in what sense video generation models like SORA can be said to understand the physical world.
	An alternative to cognitivism: computational phenomenology and deep learning Pierre Beckmann, Guillaume Köstner, Ines Hipolito Minds and Machines, 2023 paper
Reasoning in LLMs
	SciR: A Controllable Benchmark for Scientific Reasoning in LLMs Pierre Beckmann, Marco Valentino, Andre Freitas arXiv preprint paper / code / data A multi-paradigm benchmark for scientific reasoning — deduction, induction, and causal abduction — where tasks are generated as formal objects with verifiable answers, then rendered into scientific documents. Two difficulty axes, inference complexity and premise obfuscation, can be varied independently to decompose model failure into extraction and inference.
	Adaptive LLM-Symbolic Reasoning via Dynamic Logical Solver Composition Lei Xu, Pierre Beckmann, Marco Valentino, Andre Freitas EACL, 2026 paper A neuro-symbolic framework that automatically identifies formal reasoning strategies from natural language problems and dynamically selects specialized logical solvers via autoformalization interfaces.
Mechanistic Interpretability
	Probing Persona-Dependent Preferences in Language Models Oscar Gilg, Pierre Beckmann, Daniel Paleka, Patrick Butlin arXiv preprint arXiv Using linear probes on residual-stream activations, we find a preference vector that tracks a model's preferences across prompts and situations — and show this representation is largely shared across personas, predicting and steering choices even for an adversarial persona.
Deep learning for speech (2019–2023) I am no longer active in this field.
Deep speech inpainting of time-frequency masks Pierre Beckmann, Mikolaj Kegler, Milos Cernak Interspeech, 2020 arXiv
Word-Level Embeddings for Cross-Task Transfer Learning in Speech Processing Pierre Beckmann, Mikolaj Kegler, Milos Cernak EUSIPCO, 2021 arXiv
SERAB: A multi-lingual benchmark for speech emotion recognition Neil Scheidwasser-Clow, Mikolaj Kegler, Pierre Beckmann, Milos Cernak ICASSP, 2021 arXiv
Hybrid Handcrafted and Learnable Audio Representation for Analysis of Speech Under Cognitive and Physical Load Gasser Elbanna, Alice Biryukov, Neil Scheidwasser-Clow, Lara Orlandic, Pablo Mainar, Mikolaj Kegler, Pierre Beckmann, Milos Cernak Interspeech, 2022 arXiv
Byol-s: Learning self-supervised speech representations by bootstrapping Gasser Elbanna, Neil Scheidwasser-Clow, Mikolaj Kegler, Pierre Beckmann, Karl El Hajal, Milos Cernak PMLR, 2023 arXiv 3rd place in HEAR benchmark.

Research

Philosophy of AI

Reasoning in LLMs

Mechanistic Interpretability

Deep learning for speech (2019–2023)