Exemplar models of language have proven to be a promising line of research in recent decades, with a number of studies suggesting that such an approach can account for otherwise puzzling facts about language processing and language change (Bybee and McClelland, 2005; Pierrehumbert, 2002). One strand of research in this area explores whether, and how, exemplar models can handle the types of generalizations that are the focus of traditional generative phonology; phenomena such as category formation (Pierrehumbert, 2001) and categorical patterns and contrast across the lexicon (Wedel, 2004) have been successfully modeled in exemplar frameworks. This paper adds to this body of work by exploring positional neutralization, a pattern whose traditional analysis involves a unique underlying form for each word in the lexicon and a categorical rule that eliminates some distinctions among them. I show that it is possible to model positional neutralization – specifically, final devoicing – in an exemplar framework that relies exclusively on surface forms, but that this is possible only under certain conditions.The simulations presented here are similar to those in Wedel (2004) and adopt a fairly standard set of assumptions. The first set of simulations, exemplified by Simulation A in Table 1, implement a basic model with paradigm uniformity and no neutralization. Each simulation consists of a lexicon of ten lemmas, each associated with two cases (‘nominative’ and ‘accusative’), for a total of twenty distinct wordforms. Each wordform has a cloud of up to five exemplars, and is initially seeded with a single randomly generated wordform. All words have the form CVC (nominative) or CVCi (accusative; the ‘suffix’ [-i] is not allowed to vary). On each cycle of the simulation, a randomly selected exemplar is chosen as the base for a ‘production’ and subjected to analogical pressure from other exemplars in the lexicon – with special weight given to exemplars in the same paradigm – plus a small amount of noise. Each production is then probabilistically categorized as a member of the exemplar cloud it most closely resembles; the production is stored as a new exemplar in the cloud, replacing a randomly selected old exemplar if the cloud is already full. Under these conditions, the lexicon evolves to near-perfect paradigm uniformity, even when the initial nominative and accusative seeds are unrelated.Simulation B shows the problem that arises when bias is added to the production process in this simple model, such that word-final consonants have a small chance of being devoiced. Word-final consonants evolve to become consistently voiceless, as expected. However, the corresponding stem-final consonants in the accusative forms are also consistently voiceless, due to the pressure toward paradigm uniformity, despite the fact that they are not themselves word- final. In simulations of this type, it appears to be impossible to model positional neutralization: the consonants in the non-neutralizing position are ‘pulled’ towards the voiceless forms by paradigm uniformity, and there is no counteracting pressure to encourage retention of voiced consonants.The problem cannot be solved by adding another bias to production, one that encourages voicing in non-final consonants; this would lead to simple allophony, in which consonants are voiceless word-finally and voiced elsewhere. Nor can it be solved by looking for regularities in the distribution of voicing in non-neutralized forms; even when those regularities are real (e.g., Ernestus and Baayen 2003), we are still left with a residue of unpredictability to account for. What we need is a way to allow contrast word-medially, such that voiced and voiceless consonants are both allowed and voicing is specified unpredictably on a word-by-word basis.The solution adopted here is to allow paradigm uniformity to operate non-symmetrically: neutralized forms are under pressure to resemble non-neutralized forms, but not vice versa. This approach ensures the similarity of morphologically related forms without propagating neutralization throughout the paradigm. The approach has a principled basis; Albright (2010) and Albright and Kang (2009) present evidence that the base of an inflectional paradigm is the paradigm’s most informative member – effectively, the member that is least neutralized.Simulation C operationalizes ‘informativity’ as entropy (Shannon, 1948): the probability that some feature of a production of a given wordform will be altered to match another member of the same paradigm is proportional to the entropy of that feature across all words with the same case as the other member. As the bias toward final devoicing leads to consistently voiceless final consonants, the entropy of the [voice] feature of final consonants in the nominative quickly approaches zero, and accusative forms are therefore unlikely to be influenced by them. As shown in Table 1, the result is a robust pattern of final devoicing but non-final contrast.If exemplar-based approaches are to be viable models of language, they must be able to handle patterns such as positional neutralization that have been described so successfully in more traditional frameworks. The present study represents an important advance on previous implementations by demonstrating that it is indeed possible to model positional neutralization in an episodic, surface-form-based model without abstract underlying stems. Moreover, these results provide further evidence for a particular type of paradigm uniformity: one that, in effect, makes reference to a privileged base. Table 1: Results of simulations under various settings. Each simulation was run 100 times;representative examples are given here.Simulation ASimulation BSimulation CNOMACCNOMACCNOMACCpigbigibitbititaptabipukpukigitgitibapbapidakdakigutgutitattatipidpidipukpukibukbugibupbupidatdatidupdupitiktikipatpatibapbapipukpukigitgitibapbapipukpukigitgitibapbapidukdukigutgutiditdiditidtiditiktikituktagi