Cognitive Thermometers: Neural Networks Proposed as a Way to Measure the Complexity of Meaning

The preprint arXiv:2610.10724 (cs.CL) raises a question at the intersection of linguistics, cognitive science, and machine learning: why do natural languages systematically prefer some meanings over others? The authors propose measuring the complexity of meaning through the learnability of models, and they call such models "cognitive thermometers." The paper contains no new benchmarks or record numbers. It is a survey of accumulated evidence and a methodological position, but one with practical implications for AI researchers.

Why logical complexity turned out to be an awkward tool

Semantic typology studies which meanings appear across the world's languages and which almost never do. Typical objects of study include quantifiers such as "all," "some," and "most," as well as words for color, kinship, and modality. The classic explanation runs like this: frequent meanings are simpler. Simplicity is usually defined through logic, by looking at the length of the formula that describes a meaning or the complexity of its definition.

This approach has a weak spot, and the authors name it directly. The complexity estimate depends heavily on the choice of logical language. Add one operator to the set of primitives, and one meaning becomes "shorter" while another becomes "longer." Researchers make some of these choices without independent justification. The result is a metric that can be tuned to produce a desired outcome, even if unintentionally.

What the authors propose instead

The authors argue that machine learning offers a more agnostic way to measure the same quantity. The idea is to train a model to recognize a meaning and see how much data or how many steps it needs. The harder a meaning is to learn, the "more complex" it is. This does not require choosing a logical formalism, although other assumptions remain: the architecture, the data, and the training procedure.

The survey draws three conclusions:

• Logic and machine learning often give similar estimates of the relative complexity of meanings.
• Both sets of estimates predict equally well which meanings occur more often across the world's languages.
• Where the methods diverge, learnability, in the authors' view, explains typological data better than logical complexity does.

Hence the thermometer metaphor. A thermometer does not explain why it is warm, but it gives a reproducible number that does not depend on the observer's tastes. In this role, the model is a measuring instrument, not a model of human thinking.

A bridge between symbols and connectionism

The paper's main theoretical argument is that this approach unites symbolic logic and connectionist AI. For many years, the two camps have argued over how to represent knowledge: with rules or with distributed weights. Here they get a shared task. Logic generates hypotheses about the structure of meanings, and learning tests how well those hypotheses match what is actually easy to acquire.

The caveats are obvious. A neural network's learnability depends on its architecture, inductive biases, and data distribution, so the "thermometer" may have its own systematic error. Similarity between its results and human typology does not prove that humans learn the same way. It is a correlation between two measurement procedures, not proof of a shared mechanism.

Where this could be useful

For practitioners and researchers, the takeaway is fairly down-to-earth. If you study what language models learn and how, semantic typology offers a ready-made set of test tasks with known human data. If you work on interpretability, you can compare "complexity for the model" with "complexity for humans" and look for gaps: they would show where a model uses alien, non-human strategies. For those building neurosymbolic systems, the paper serves as an argument for hybrid metrics, in which formal complexity is checked against empirical learnability.

For now, this is more a research program than a ready-made standard. Its value will be determined by reproducible experiments across different architectures: if different models produce a consistent "degree of complexity," the thermometer can be considered to work.