ADR-031: compositionality and mathematical-statistics join the topics

Context

The manuscript batch of 2026-10-09 was read by several agents working in parallel. Each reported, without seeing the others’ reports, that the closed vocabulary was missing a word.

• Compositionality. GenEval, T2I-CompBench and Composable Diffusion were filed with representation-learning first. Their readers found the tag hard to defend. None of the three papers studies how a model represents its data. GenEval and T2I-CompBench ask whether generated images bind attributes to objects and get relations right. Composable Diffusion composes concepts by adding one guidance term per concept. What the three have in common is compositionality. So do LIT-667 and THEORY-117, which are about compositional generalization, and the compositional distributional semantics of LIT-272 and LIT-273, which are filed under linguistics, logic and mathematics.
• Mathematical statistics. Blackwell 1953 (LIT-781), Torgersen 1991 (LIT-778) and the THEORY filed from Blackwell’s paper went under probabilistic-modeling, whose blurb is “Bayesian inference and statistical models outside deep learning — nonparametric priors, topic models, samplers”. The comparison of experiments is none of these things. The same mismatch applies to Fritz’s theorems on sufficient statistics (LIT-772, and the THEORY filed from it), and to Rao 1945 on the accuracy attainable in estimation (LIT-349).

The owner authorised extending the vocabulary on 2026-10-09.

Decision

Add two topics.

• compositionality (Compositionality): how the structure and meaning of a whole come from its parts and the way they are combined. It covers the principle of compositionality and compositional semantics, and the emergence of compositional structure in languages under transmission. It also covers composition and binding in learned models: compositional generalization to unseen combinations, attribute binding, and composing distributions or concepts. It is a subject word. A work on how to make a model generalize compositionally is still ML practice, and goes to the anthology.
• mathematical-statistics (Mathematical statistics): the theory of statistical inference read for itself. It covers comparison and sufficiency of experiments (Blackwell, Le Cam deficiency), statistical decision theory, the sufficiency theorems (Fisher–Neyman, Basu, Bahadur), and the bounds on estimation. Bayesian modelling stays under probabilistic-modeling.

Retagged now:

• GenEval and T2I-CompBench take compositionality first in place of representation-learning. Both keep anthology-candidate last.
• Composable Diffusion takes compositionality first and drops representation-learning, which its reading did not support.
• LIT-667, THEORY-117, LIT-272 and LIT-273 add compositionality after their present primary tag.
• Blackwell, Torgersen and the THEORY filed from Blackwell’s paper take mathematical-statistics first and drop probabilistic-modeling, the nearest wrong word they had been given.
• Fritz and the THEORY filed from Fritz’s paper add mathematical-statistics after mathematics.
• Rao 1945 (LIT-349) takes mathematical-statistics first.

Works still being read when this was decided are retagged once their reading is filed. Kirby, Cornish and Smith is the obvious case for compositionality.

Alternatives

• generative-modeling. This is the subject the text-to-image papers share. It is also the anthology’s topic word, and having it here would invite ML practice to be filed in this record. The anthology-candidate flag already says where those papers may belong.
• Widening probabilistic-modeling. That would turn a word about building models into a word about all of statistics. The blurb would then cover everything, and the tag would distinguish nothing.
• translation. Gutt and Hofstadter are the only two works, and neither has been read, because no lawful copy of either was reachable. Under the record’s habit of waiting for a second instance, a decision can wait until one of them has been read.