Language models routinely produce grammatical sentences that go beyond their training corpora, a feat that suggests they pick up on indirect evidence rather than memorizing examples. Prior work has inferred that such evidence exists, but its precise location has remained unclear. A new arXiv paper, "Where's the Indirect Evidence? Investigating Grammatical Generalization with Influence Functions," takes up this question directly.
The authors propose using influence functions, a technique for tracing a model's predictions back to specific training data points, to find which examples actually drive grammatical generalization. This approach could reveal whether models rely on particular constructions, lexical patterns, or some other form of indirect signal when learning grammar.
Because the source is only an abstract, the paper's methods and results are not described in detail here. The abstract itself is cut off mid-sentence, so no empirical findings or conclusions can be reported from this source alone.