Projects

Arthur Renard, Franck Gabriel, Valentin Hartmann, Clément Hongler

Xent Labs · Université Lyon 1

Frost Training improves Monte Carlo policy optimization for Cross-Entropy Games by exploiting the gradient of the reward in embedding space. The signal behind the GCG jailbreaking technique is repurposed to boost training: applied to GRPO for maximum-likelihood infilling, it reaches higher best-of-k scores faster.

Clément Hongler, Franck Gabriel, Valentin Hartmann, Arthur Renard, Andrew Emil

Xent Labs

A framework for automatically growing a curriculum of tasks that builds general capabilities in language models through relevant skill discovery. Using Cross-Entropy Games, the paper shows that greedy curriculum learning admits essentially a single meta-objective (up to a few hyper-parameters), yielding a principled process called cognitive training.

Stéphane d'Ascoli*, Arthur Renard*, Emmanuel Abbé, Clément Hongler, Vassilis Papadopoulos, Josh Susskind, Samy Bengio (* equal contribution)

EPFL · Apple

A Transformer trained end-to-end for symbolic regression of Boolean functions. Predicts compact formulas from truth tables, remains robust to noisy or partial observations, and is competitive with state-of-the-art genetic algorithms on gene regulatory network inference at orders-of-magnitude lower cost.

Vassilis Papadopoulos, Guilhem Doat, Arthur Renard, Clément Hongler

EPFL

An algorithm that efficiently locates parameters at the border between phases in high-dimensional dynamical systems. Applied to Lenia, it doubles the frequency of interesting emergent behaviours while remaining tractable for large-scale searches.