PhD student, Department of Computer Science, Aarhus University
I work on how large language models reason, supervised by Akhil Arora and Lars Klein. My aim is to make that reasoning, as Daft Punk would put it, harder, better, faster, stronger. Before Aarhus I studied electrical engineering and then data science at EPFL, where my master's thesis at DLAB became Fleet of Agents.
ICML 2025 * equal contribution
A group of LLM agents explore a problem as a dynamic tree search, using genetic-type particle filtering to decide which branches deserve more compute.
Findings of EMNLP 2025 * equal contribution
An asynchronous, batched request layer that sits between your code and the LLM, cutting cost and latency while keeping experiments reproducible.
Tutorial, ACL 2026 and IJCAI-ECAI 2026
A tutorial on where LLM reasoning stands today: how models are trained and prompted to reason, how that reasoning is evaluated, and what remains open.
ICLR 2026 workshops
Treats an agent's tool calls like database transactions, so that a multi-step workflow can be kept consistent and recovered when a step fails.
REALM Workshop, EMNLP 2026 (spotlight) * equal contribution
A benchmark for how much an LLM's reasoning changes from run to run, so that reliability can be measured instead of assumed.
REALM Workshop, EMNLP 2026 (spotlight)
Simply retrying a failed reasoning attempt is a surprisingly strong baseline, often matching methods that ask the model to critique itself.
