xent is a research lab working on the alignment of language models. the name is short for cross-entropy: the quantity a language model computes for every piece of text it reads, and the only thing about a model we take at face value
most of what one wants to know about a model — what it knows, whether it says what it predicts, what would change its mind — is a question about its probability distribution. the model itself usually cannot answer such questions. but they can be computed, and the answers can be checked. that is where our work starts
we build xent games, tasks scored by cross-entropy that need no human grader and no reward model, and cognitive training, a procedure that uses them to measure a model before, and after, anything is trained
the usual way to find out whether a model is honest, or consistent, or knows what it knows, is to ask it — or to run it on a test. both rely on the model's answers. a model can be wrong about itself, a capable one can answer strategically, and a test only covers what it covers
the model's probabilities are different. they are what it actually computes, for any text, and they do not depend on what it chooses to say. the probability a model gives to a statement, to the same statement after one more fact, to the same question asked another way — these are numbers one reads off directly, and their relationships are facts about the model, not reports from it
a is a task whose score is one of these numbers. in-filling: fill the gap so the model's cross-entropy on the whole passage is lowest. reverse-prompting: find the prompt under which the model gives a text the highest probability. counterfactual thinking: find the fact that most changes the model's cross-entropy on a passage. the model plays, and the model's own cross-entropy keeps score. there is no rubric to argue with and no second model to fool
two things, and the measurements in between
a family of tasks for a language model, each scored by the model's own cross-entropy. in-filling, reverse-prompting and counterfactual thinking are the simplest members. the first paper argues that the family is wide enough to express a large part of what one asks a language model to do
what the games measure: how much a given fact moves a model's predictions; what surprises it; whether it stays consistent when a question is rephrased; whether it knows which prompt leads it to which answer. each is a cross-entropy or a difference of two, so each is exact, cheap and repeatable
a training procedure in which the games are the only data. before a game is used, a fixed objective scores what training on it would change; after training, the same games are replayed and the change is measured. the second paper derives that objective rather than tuning it
cognitive training, step by step:
a model can play dumb on an evaluation. its cross-entropy is not something it reports; it is something it computes, and we read it directly. so a property stated in cross-entropy — self-prediction, consistency, the value of a fact — is checked against the model's computation. its cooperation is not required
the training side is narrower than it sounds. cognitive training does not set out to make a model more capable in general. it trains against explicit games, each chosen by a fixed objective and logged, and measures the result with the same games. capability that appears anyway shows up in the verify step as a number, where it can be looked at — or stopped
two things stay fixed throughout: the objective that selects games, which is derived once and not retuned, and the set of games used to measure. nothing in the loop rewrites either
what this does not do: it does not tell you a model's values, and it does not make a model safe by itself. it makes a specific class of claims about a model — what it knows, whether it is honest about what it knows, what moved during training — checkable by computation rather than by trust. we think that class is where alignment work has to start
one family of xent games asks: which piece of information would most change the model's view of a passage? the real game is played by the model and scored in cross-entropy. 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.
defines the games and shows how far the family reaches, from in-filling to counterfactual reasoning
formulates the measure–select–train–verify loop and derives the meta-objective that scores a game before it is used
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
this page is one path through xent's ideas. the map holds all of them — thirteen concepts, connected the way the ideas actually are