Traditional computational accounts of gender representation and learning (e.g., Carroll, 1989, 1995) differ radically from cue-based and connectionist accounts. The latter but not the former predicts that access to noun gender will vary depending on the reliability of noun endings (and other sublexical elements and morphological constituents) in marking gender, and that agreement markers can be used strategically to constrain the genders of ambiguously marked nouns. Adult native (L1) speakers of Russian (Experiment 1) and advanced nonnative (L2) speakers (Experiment 2) read Russian sentences on a computer and were asked to choose one of two inflected past tense verbs in a forced choice task. The verbs either matched or mismatched the gender of the subject NP. Half of the target trials used opaque (end-palatalized) subject nouns, which were ambiguously marked for gender, and the other half used transparent (regularly marked) subject nouns. Noun type was crossed with the presence or absence of a gender-marked adjective in the subject NP. When an adjective was present in the subject NP, both L1 and L2 speakers were significantly faster at reading and selecting the correct verb form. L2 but not L1 speakers showed longer reading and choice latencies and made more errors when the subject noun was opaque. The data showed that L2 speakers used adjective inflections to disambiguate the gender of opaque subject nouns and to select gender appropriate verb inflections. The accuracy data for L1 and L2 speakers were tested against several connectionist models. The models' success in accounting for the data suggested that L1 and L2 speakers may depend on a common learning mechanism and thus resemble one another at the computational level, contrary to traditional computational accounts (Carroll, 1989, 1995).