cognitive training
the paradigm: a controlled self-improvement loop where a model creates its own training tasks, keeps the ones worth creating, and learns from them
xent — safely
shaping intelligence
xent is an AI research lab: we shape the cognitive abilities of language models — transparently and deliberately, so that what a model becomes capable of is a choice, not an accident
the idea is cognitive training: a model discovers relevant new skills by creating tasks for itself — with a principled way to measure which tasks are worth creating
done right, self-improvement becomes something we can steer: models that grow useful, competitive skills while their abilities stay measured and bounded
what we build
one research program, three concrete parts
xent games
the task space: a surprisingly rich family of games a model can pose to itself and score using nothing but its own probabilities
frost
the algorithm: our reinforcement-learning method, exploiting the differentiable structure of xent games to learn faster than Monte-Carlo methods
how it works
instead of an ever-growing pile of handcrafted environments, cognitive training turns task discovery itself into a measurable, steerable loop:
- create — the model proposes a new , drawing on knowledge already implicit in its own distribution
- measure — the values the game by its transfer: which skills it strengthens, which it opens, and which it leaves alone
- train — learns the worthwhile games efficiently
- repeat — the model returns to task creation stronger, while the objective that steers it stays fixed
shaping, not just scaling
today, models gain abilities as a side effect of scale — uneven, hard to predict, and not all of them wanted
cognitive training takes the other route: skills grow through explicit tasks under a fixed objective, so capability goes where it is useful — and stays bounded where it is not
the aim: models whose abilities are known in advance, and can be trusted where it matters most
counterfactual thinking — try it
one family of xent games asks: which piece of information would most change one's view? here is the human-scale version
“a new language model improved by 40% on its main benchmark.”
which new fact should change your interpretation the most?
an illustration for visitors — the real games are played, and scored, by the model itself.
our research
xent games — “Cross-Entropy Games for Language Models”
introduces the task space: games grounded in a model's own measures, from in-filling to counterfactual thinking
cognitive training — “Cognitive Training for Language Models”
formulates the self-improvement loop and derives the meta-objective that decides which games are worth creating
frost — “Cross-Entropy Games and Frost Training”
the training algorithm: exploiting differentiable game structure to move beyond the Monte-Carlo paradigm
notes — Clément's running notes on xent games at xent.blog
the people building xent
Clément and Franck, with Arthur Jacot, received the 2026 ICBS Frontiers of Science Award for their work on the neural tangent kernel
contact — write to hello@xent.io
the territory
this page is one path through xent's ideas. the map holds all of them — thirteen concepts, connected the way the ideas actually are