EPFL Pre-ICML 2025 Research Highlights

The ELLIS Lausanne Unit and the EPFL AI Center are pleased to host a Pre-ICML 2025 Research Highlights Session, highlighting recent contributions from the EPFL community to ICML 2025 and other leading machine learning conferences.

This session will feature a series of short 10-min presentations, offering an opportunity for researchers to share their work and engage with colleagues across the EPFL machine learning ecosystem. The presentations will be followed by a networking coffee, providing space for further discussion and new connections.

Programme

The full schedule, including confirmed speakers and talk titles, will be shared closer to the event date.

13:45–14:00 - Arrival & Welcome

• 14:00–15:00 — Talks (5 × 10 min)
  • Suryanarayana Sankagiri, « Recommendations with Sparse Comparison Data: Provably Fast Convergence for Nonconvex Matrix Factorization »
  • Yiming Qin, « DeFoG: Discrete Flow Matching for Graph Generation»    
  • Ilia Badanin, «Benchmarking Concept-Spilling Across Languages in LLMs »
  • Agatha Duzan, « OS-Harm: A Benchmark for Measuring Safety of Computer Use Agent s»
  • Filippo Bigi, « The dark side of the forces: assessing non-conservative models for atomistic machine learning »
     
• 15:00–15:10 — Break
   
• 15:10–16:00 — Talks (5 × 10 min)
  • Emanuele Troiani, « Fundamental limits of learning in sequence multi-index models and deep attentionnetworks: High-dimensional asymptotics and sharp thresholds »
  • Adytia Varre, « Learning In-context n-grams with Transformers: Sub-n-grams Are Near-stationaryPoints »
  • Roman Bachmann, « FlexTok: Resampling Images into 1D Token Sequences of Flexible Length »
  • Ben Lonnqvist, « Contour Integration Underlies Human-Like Vision »
  • Hatef Otroshi, « Synthetic Face Datasets Generation via Latent Space Exploration from BrownianIdentity Diffusion »