Understanding the origins of linguistic compositionality is a fundamental challenge in evolutionary linguistics. Prior workhas explored this topic through dynamical computational modeling and experiments in iterated learning. We explorethese questions using RL agents tasked with developing cooperative communication strategies in a signaling game. Weanalyze how various mechanisms (such as Bayesian pragmatic reasoning) and constraints (such as limited memory) mayaffect compositionality and generalizability in the invented communication protocols. In particular, our preliminary resultssuggest that incremental pragmatic reasoning induces a bias towards lexical compositionality. To evaluate the extensibilityof our model, we compare the behavior of the RL agents to the behavior of humans on the same task. That is, we askhumans to coordinate in a reference game task by repeatedly composing non-linguistic symbols. We discuss ways in whichthe resulting protocol mirrors and differs from that produced by the RL agents.

Large Language Models (LLMs) have driven extraordinary improvements in NLP. However, it is unclear how such models represent lexical concepts-i.e., the meanings of the words they use. We evaluate the lexical representations of GPT-4, GPT-3, and Falcon-40B through the lens of HIPE theory, a concept representation theory focused on words describing artifacts (such as “mop”, “pencil”, and “whistle”). The theory posits a causal graph relating the meanings of such words to the form, use, and history of the referred objects. We test LLMs with the stimuli used by Chaigneau et al. (2004) on human subjects, and consider a variety of prompt designs. Our experiments concern judgements about causal outcomes, object function, and object naming. We do not find clear evidence that GPT-3 or Falcon-40B encode HIPE's causal structure, but find evidence that GPT-4 does. The results contribute to a growing body of research characterizing the representational capacity of LLMs.

Human learning is sensitive to rule-like structure and the curriculum of examples used for training. In tasks governed by succinct rules, learning is more robust when related examples are blocked across trials, but in the absence of such rules, interleaving is more effective. To date, no neural model has simultaneously captured these seemingly contradictory effects. Here we show that this same tradeoff spontaneously emerges with “in-context learning” (ICL) both in neural networks trained with metalearning and in large language models (LLMs). ICL is the ability to learn new tasks “in context” — without weight changes — via an inner-loop algorithm implemented in activation dynamics. Experiments with pretrained LLMs and metalearning transformers show that ICL exhibits the blocking advantage demonstrated in humans on a task involving rule-like structure, and conversely, that concurrent in-weight learning reproduces the interleaving advantage observed in humans on tasks lacking such structure.

Models based on the Transformer neural network architecture have seen success on a wide variety of tasks that appear to require complex ``cognitive branching''-- or the ability to maintain pursuit of one goal while accomplishing others. In cognitive neuroscience, success on such tasks is thought to rely on sophisticated frontostriatal mechanisms for selective gating, which enable role-addressable updating-- and later readout-- of information to and from distinct ``addresses'' of memory, in the form of clusters of neurons. However, Transformer models have no such mechanisms intentionally built-in. It is thus an open question how Transformers solve such tasks, and whether the mechanisms that emerge to help them to do so bear any resemblance to the gating mechanisms in the human brain. In this work, we analyze the mechanisms that emerge within a vanilla attention-only Transformer trained on a simple sequence modeling task inspired by a task explicitly designed to study working memory gating in computational cognitive neuroscience. We find that, as a result of training, the self-attention mechanism within the Transformer specializes in a way that mirrors the input and output gating mechanisms which were explicitly incorporated into earlier, more biologically-inspired architectures. These results suggest opportunities for future research on computational similarities between modern AI architectures and models of the human brain.

Narratives (sequences of purposively related concrete situations) and arguments (reasoning and conclusions in an attempt to persuade) are distinct cornerstones of human discourse. While theories of their linguistic structures exist, it is unclear which theorized features influence perception of narrative and argument quality. Furthermore, differences in their usage over time and across formal versus informal mediums remain unexplored. Thus, we use an original dataset of news and Reddit discourse (consisted of >10,000 clauses), annotated for clause-level discourse elements (e.g., generic statements vs. events; Smith, 2003), and their coherence relations (e.g., cause/effect; Wolf & Gibson, 2005). We identify the features that correspond to differing perceptions of narrative and argument quality across multiple dimensions. Since the documents cover marijuana legalization discourse during a period of massive attitude shift in the U.S. (2008-2019), we also examine changes over time in discourse structure within this rapidly evolving sociopolitical context.

Text-generation algorithms like GPT-2 (Radford et al., 2019) and GPT-3 (Brown et al., 2020) produce documents which resemble coherent human writing. But no study has compared the discourse linguistics features of the artificial text with that of comparable human content. We used a sample of Reddit and news discourse as prompts to generate artificial text using fine-tuned GPT-2 (Grover; Zellers et al., 2019). Blind annotators identified clause-level discourse features (e.g., states and events; Smith, 2003), and coherence relations (e.g., contrast; Wolf and Gibson, 2005) in prompts and generated text. Comparing the >20000 clauses, Grover recreates human word co-occurrence patterns and clause types across discourse modes. However, its coherence relations have shorter length and lower quality, with many nonsensical instances. Therefore, annotators could perfectly guess the human/algorithmic source of documents. Using a corresponding GPT-3 sample, we discuss aspects of generation that have and have not improved since Grover.