xent games
a family of tasks scored by a model's own cross-entropy, from in-filling and reverse-prompting to counterfactual reasoning. introduced in our 2025 paper
xent — safely shaping
language models
xent is an AI safety research lab. we work on a basic problem: language models are deployed before anyone can say precisely what they know, what they can do, or what their training taught them. we want to make those things checkable
our main tool is xent games: tasks a model is given and scored on using its own probabilities, with no human grader and no second model involved. they let us test what a model knows about itself, whether it is consistent, and how a round of training changed it
we also work on cognitive training, a training method in which everything the model learns from is an explicit, recorded task. the aim is training whose effects can be predicted and then verified, rather than discovered later
the safety problem we work on
almost everything known about a deployed model comes from evaluations: prompts, benchmarks, red-teaming. these depend on the model's answers. a model can be wrong about itself, and a sufficiently capable model can answer in whatever way serves it. evaluations also cover only the cases someone thought to test
training is the other gap. a modern model is trained on thousands of hand-built environments and on rewards produced by other models. at the end, nobody can say exactly what was learned, and unintended abilities and habits are found afterwards, if at all
we think safety work needs a layer underneath this: ways of checking facts about a model that do not go through its answers, and a form of training whose effects are measured rather than inferred. that is what xent games and cognitive training are for
xent games
a language model assigns a probability to any text. a xent game is a task whose score is computed from those probabilities. for example: given a passage, find a short prompt that makes the model predict it as well as possible. or: find the piece of information that changes the model's predictions the most
the model does the scoring. there is no rubric to write and no reward model to train, and the same model can be asked thousands of such questions
for safety, this matters because the score does not depend on what the model chooses to say. it depends on probabilities the model computes anyway. that makes xent games a way to check a model's self-knowledge, its consistency, and what it finds surprising, without relying on its cooperation
what we build
three parts
safety checks
what we use the games for: testing whether a model knows its own predictions, whether it answers rephrased questions consistently, which facts change its mind, and what a round of training altered
cognitive training
a training method in which xent games are the only training data. each game's expected effect is estimated before training and measured afterwards, so the result can be checked against the intent. described in our 2026 paper
how it works
cognitive training, in four steps:
- measure — run a fixed set of on the model and record the results. this is the baseline
- select — score candidate training games with the , which estimates what training on each of them would change. choose games on that basis, and record the choice
- train — train on the chosen games. the training signal is the model's cross-entropy on them
- verify — run the baseline games again. compare the change with what was expected; anything unexpected is visible here
the meta-objective and the baseline games stay fixed throughout. the model does not choose its own objective, and the record of what it trained on is complete
counterfactual thinking — try it
one family of xent games asks: which piece of information would most change the model's view of a passage? here is a human-scale version of the same question
“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 xent games and shows the range of tasks they cover
cognitive training — “Cognitive Training for Language Models”
describes the training method and the meta-objective used to choose games
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