Though language models (LMs) have enabled great success across a wide variety of tasks, they still appear to fall short of one of the loftier goals of artificial intelligence research: creating an artificial system that can adapt its behavior to radically new contexts upon deployment (Shi et al., 2024).One important step towards this goal is to create systems that can induce rich representations of data that are seen in-context, and then flexibly deploy these representations to accomplish goals (Lampinen et al., 2024).Recently, Park et al. (2025a) demonstrated that current LMs are indeed capable of inducing such representation from context (i.e., in-context representation learning).The present study investigates whether LMs can use these representations to complete simple downstream tasks.We first assess whether open-weights LMs can use in-context representations for next-token prediction, and then probe models using a novel task, adaptive world modeling.In both tasks, we find evidence that open-weights LMs struggle to deploy representations of novel semantics that are defined in-context, even if they encode these semantics in their latent representations.Furthermore, we assess closed-source, state-of-the-art reasoning models on the adaptive world modeling task, and demonstrate that even the most performant LMs cannot reliably leverage novel patterns presented in-context.Overall, this work seeks to inspire novel methods for encouraging models to not only encode information presented in-context, but to do so in a manner that supports flexible deployment of this information.

Studies show that safety-tuned models may nevertheless divulge harmful information. In this work, we show that whether they do so depends significantly on who they are talking to, which we refer to as *user persona*. In fact, we find manipulating user persona to be more effective for eliciting harmful content than certain more direct attempts to control model refusal. We study both natural language prompting and activation steering as intervention methods and show that activation steering is significantly more effective at bypassing safety filters. We shed light on the mechanics of this phenomenon by showing that even when model generations are safe, harmful content can persist in hidden representations and can be extracted by decoding from earlier layers. We also show we can predict a persona’s effect on refusal given only the geometry of its steering vector. Finally, we show that certain user personas induce the model to form more charitable interpretations of otherwise dangerous queries.

Prior work has explored the writing challenges experienced by people with dyslexia, and the potential for new spelling, grammar, and word retrieval technologies to address these challenges. However, the capabilities for natural language generation demonstrated by the latest class of large language models (LLMs) highlight an opportunity to explore new forms of human-AI writing support tools. In this paper, we introduce LaMPost, a prototype email-writing interface that explores the potential for LLMs to power writing support tools that address the varied needs of people with dyslexia. LaMPost draws from our understanding of these needs and introduces novel AI-powered features for email-writing, including: outlining main ideas, generating a subject line, suggesting changes, rewriting a selection. We evaluated LaMPost with 19 adults with dyslexia, identifying many promising routes for further exploration (including the popularity of the “rewrite” and “subject line” features), but also finding that the current generation of LLMs may not surpass the accuracy and quality thresholds required to meet the needs of writers with dyslexia. Surprisingly, we found that participants’ awareness of the AI had no effect on their perception of the system, nor on their feelings of autonomy, expression, and self-efficacy when writing emails. Our findings yield further insight into the benefits and drawbacks of using LLMs as writing support for adults with dyslexia and provide a foundation to build upon in future research.

The latest generation of large neural language models such as GPT-3 have achieved new levels of performance on benchmarks for language understanding and generation. These models have even demonstrated an ability to perform arbitrary tasks without explicit training. In this work, we sought to learn how people might use such models in the process of creative writing. We built Wordcraft, a text editor in which users collaborate with a generative language model to write a story. We evaluated Wordcraft with a user study in which participants wrote short stories with and without the tool. Our results show that large language models enable novel co-writing experiences. For example, the language model is able to engage in open-ended conversation about the story, respond to writers’ custom requests expressed in natural language (such as ”rewrite this text to be more Dickensian”), and generate suggestions that serve to unblock writers in the creative process. Based on these results, we discuss design implications for future human-AI co-writing systems.

Transformer architectures show significant promise for natural language processing. Given that a single pretrained model can be fine-tuned to perform well on many different tasks, these networks appear to extract generally useful linguistic features. A natural question is how such networks represent this information internally. This paper describes qualitative and quantitative investigations of one particularly effective model, BERT. At a high level, linguistic features seem to be represented in separate semantic and syntactic subspaces. We find evidence of a fine-grained geometric representation of word senses. We also present empirical descriptions of syntactic representations in both attention matrices and individual word embeddings, as well as a mathematical argument to explain the geometry of these representations.

Homophily - our tendency to surround ourselves with others who share our perspectives and opinions about the world - is both a part of human nature and an organizing principle underpinning many of our digital social networks. However, when it comes to politics or culture, homophily can amplify tribal mindsets and produce "echo chambers" that degrade the quality, safety, and diversity of discourse online. While several studies have empirically proven this point, few have explored how making users aware of the extent and nature of their political echo chambers influences their subsequent beliefs and actions. In this paper, we introduce Social Mirror, a social network visualization tool that enables a sample of Twitter users to explore the politically-active parts of their social network. We use Social Mirror to recruit Twitter users with a prior history of political discourse to a randomized experiment where we evaluate the effects of different treatments on participants' i) beliefs about their network connections, ii) the political diversity of who they choose to follow, and iii) the political alignment of the URLs they choose to share. While we see no effects on average political alignment of shared URLs, we find that recommending accounts of the opposite political ideology to follow reduces participants» beliefs in the political homogeneity of their network connections but still enhances their connection diversity one week after treatment. Conversely, participants who enhance their belief in the political homogeneity of their Twitter connections have less diverse network connections 2-3 weeks after treatment. We explore the implications of these disconnects between beliefs and actions on future efforts to promote healthier exchanges in our digital public spheres.