Why does articulatory probing predict SSL model performance better than phonetic probing?

This explores why articulatory probing beats phonetic probing as a predictor of self-supervised speech model performance — and the corpus has a direct answer that turns out to be a special case of a much broader pattern in how these models represent the world. The core finding is that self-supervised speech models don't learn language-specific sound categories at all; they infer the *causal, language-agnostic physics* of how a vocal tract produces acoustics Do speech models learn language-specific sounds or universal physics?. Phonetic categories (this is a /b/, that is a /p/) are surface labels — culturally specific taxonomies layered on top of the acoustic signal. Articulation (lips closing, tongue position, voicing onset) is the generative mechanism that *produces* those acoustics in the first place. A probe succeeds to the degree it reads out what the model genuinely encodes. So articulatory probing predicts performance better because it's asking the model about the thing it actually represents — the generative cause — while phonetic probing is asking about a downstream label the model never committed to.

What makes this interesting is that it's the same shape as a recurring theme across the collection: these models tend to encode the *underlying generative process or statistical substrate* rather than the human-facing semantic surface. The clearest cousin is the work on hidden trait transmission, where behavioral traits propagate between models through data bearing no semantic relationship to the trait — the mechanism embeds *statistical signatures*, not meaning, and the effect is architecture-specific in exactly the way you'd expect if it rides on the model's internal representation rather than on content humans can read Can language models transmit hidden behavioral traits through unrelated data?. In both cases, the explanatory variable is mechanistic and invisible to a semantically-framed probe.

There's also a useful tension with the metalinguistic-analysis result, where a reasoning model can explicitly construct phonological generalizations and syntactic trees through step-by-step reasoning Can language models actually analyze language structure?. That's the *phonetic-category* level made explicit — and notably it requires deliberate chain-of-thought to surface. The contrast sharpens the point: the categorical, taxonomic knowledge is something a model can be coaxed to articulate as analysis, but it's not what's doing the predictive work in a raw self-supervised speech representation. The physics is baked into the representation; the categories are a reasoning artifact.

A second lateral framing comes from the prompting literature: prompts (and by extension, probes) can only *activate* structure that already exists in the model's learned distribution — they can't inject what isn't there Can prompt optimization teach models knowledge they lack?. A probe is a readout, not a teacher. If the articulatory geometry is present in the representation and the phonetic partition isn't cleanly present, no clever phonetic probe will conjure it — you'll just measure the model failing to have organized itself the way your probe assumes. The probe's predictive power is therefore a *mirror* of the representation's actual organizing principle.

The thing you may not have known you wanted to know: this is quietly an argument about *why these models transfer across languages*. Because they learn the universal physics of speech production rather than the phonetic inventory of any one language, the articulatory representation is portable — and the probe that reads it out inherits that portability as predictive power. Phonetic probing is asking 'did the model learn English vowels?' when the model actually learned 'how a mouth makes sound,' which is the more fundamental — and more useful — thing.