Do language models possess tacit knowledge in Davies' sense?

The excerpt argues that transformer-based LLMs can acquire tacit knowledge in the sense Martin Davies (1990) gave the term: rules "not represented explicitly, but that nevertheless describe causally-relevant structures that guide behavior." Davies himself denied that connectionist networks meet the constraints, so the paper's contribution is to test those constraints against contemporary LLMs. It finds that certain architectural features satisfy semantic description, syntactic structure and causal systematicity. The causal constraint carries the most weight, and its support is a single case: the factual-association work of Meng et al. (2022), which the author calls "compelling—albeit preliminary—evidence that at least some LLMs meet the constraints for tacit knowledge."

The argument turns on an intermediate notion of knowledge. Explicit knowledge is a stored rule the system retrieves, as in the Cyc knowledge base. Weak knowledge is mere conformity: a network can be described as following "pointy ears → cat" without representing that rule, and one network may conform to several incompatible rule sets. Tacit knowledge sits between the two, describing rules that are not stored but causally shape input-output transitions. For LLMs the syntactic constraint is weakened to admit the embedding layer. The causal test has two steps. Causal tracing localizes where a fact, such as the location of the Eiffel Tower, is represented, in certain middle-layer MLP modules that act as key-value pairs. ROME then rewrites one such pair so the model outputs "Rome." If the edit propagates to the output, the stored association counts as a causal common factor.

Three library notes sit close to this. Do language models actually use their encoded knowledge? names the gap this paper has to cross: probing can show a fact is encoded without showing it does anything, and an edit that changes output is the kind of causal evidence probing lacks. Does refusing explicit knowledge harm AI system performance? faults tacit learning for producing uninterpretable, non-robust representations. This excerpt uses "tacit" as a descriptive handle that makes such representations nameable and open to intervention, without claiming they are interpretable. Its mid-layer placement of facts also cuts against Why does reasoning training help math but hurt medical tasks?, which puts retrieval in lower layers. The excerpt does not reconcile the two.

The excerpt does not establish much. It rests on one fact-editing study and a conceptual argument, with no measured test across models. The author concedes that ROME's specificity and generalization are "not perfect," that superposition may mean edits touch unrelated input-output pairs, and that Hase et al. (2023) found interventions at other locations about as effective as the original ones. That last finding weakens the claim that the located site is the causal common factor. The implication is modest: tacit knowledge is a usable hypothesis and vocabulary for intervention research, while attributing it to LLMs stays provisional until the causal constraint holds beyond one fact.