Hi, I'm

Frédéric Berdoz

PhD Student in Computer Science

Distributed Computing Group, ETH Zürich

I’m a third-year PhD student at ETH Zurich under the supervision of Prof. Dr. Roger Wattenhofer. My research focuses on AI alignment, particularly inference-time and post-training alignment of LLMs and other modalities (audio, 3D). In addition to fundamental research, I apply alignment methods to enhance inference efficiency (speculative decoding), content moderation (distributional alignment in media streams), and to align the opinions of LLMs.

News

Mar 2026 Two papers accepted at ICLR 2026 workshops: “Subliminal Signals in Preference Labels” at the Agents in the Wild Workshop, and “Reasoning Structure of Large Language Models” at the Logical Reasoning of LLMs Workshop.
Jan 2026 “Reasoning Boosts Opinion Alignment in LLMs” has been accepted at ICLR 2026!
Jan 2026 Heading to AAAI 2026 in Singapore with two oral presentations and a workshop paper!
Jan 2026 “High-Fidelity Speech Enhancement via Discrete Audio Tokens” has been accepted at ICASSP 2026 in Barcelona.
Dec 2025 Our paper “Data Attribution in Large Language Models via Bidirectional Gradient Optimization” has been accepted for an oral presentation at the AI Governance Workshop @ AAAI 2026.
Dec 2025 Traveling to NeurIPS 2025 in San Diego!
Aug 2025 Two papers accepted for oral presentation at AAAI 2026: “Text-to-Scene with Large Reasoning Models” and “Steering Pretrained Drafters during Speculative Decoding”. See you in Singapore!
Aug 2025 Presenting “Recommender Systems for Democracy” as an oral at IJCAI 2025 in Montreal!
Jul 2025 Presenting “Recommender Systems for Democracy” at IC2S2 2025 in Norrköping, Sweden!
Dec 2024 Excited to present “Can an AI Agent Safely Run a Government?” at NeurIPS 2024 in Vancouver!