There is widespread concern about the negative impacts of social media feed ranking algorithms on political polarization. Leveraging advancements in large language models (LLMs), we develop an approach to re-rank feeds in real-time to test the effects of content that is likely to polarize: expressions of antidemocratic attitudes and partisan animosity (AAPA). In a preregistered 10-day field experiment on X/Twitter with 1,256 consented participants, we increase or decrease participants' exposure to AAPA in their algorithmically curated feeds. We observe more positive outparty feelings when AAPA exposure is decreased and more negative outparty feelings when AAPA exposure is increased. Exposure to AAPA content also results in an immediate increase in negative emotions, such as sadness and anger. The interventions do not significantly impact traditional engagement metrics such as re-post and favorite rates. These findings highlight a potential pathway for developing feed algorithms that mitigate affective polarization by addressing content that undermines the shared values required for a healthy democracy.

Language models are aligned to emulate the collective voice of many, resulting in outputs that align with no one in particular. Steering LLMs away from generic output is possible through supervised finetuning or RLHF, but requires prohibitively large datasets for new ad-hoc tasks. We argue that it is instead possible to align an LLM to a specific setting by leveraging a very small number ($<10$) of demonstrations as feedback. Our method, Demonstration ITerated Task Optimization (DITTO), directly aligns language model outputs to a user's demonstrated behaviors. Derived using ideas from online imitation learning, DITTO cheaply generates online comparison data by treating users' demonstrations as preferred over output from the LLM and its intermediate checkpoints. We evaluate DITTO's ability to learn fine-grained style and task alignment across domains such as news articles, emails, and blog posts. Additionally, we conduct a user study soliciting a range of demonstrations from participants ($N=16$). Across our benchmarks and user study, we find that win-rates for DITTO outperform few-shot prompting, supervised fine-tuning, and other self-play methods by an average of 19% points. By using demonstrations as feedback directly, DITTO offers a novel method for effective customization of LLMs.

Many real-world applications of language models (LMs), such as writing assistance and code autocomplete, involve human-LM interaction. However, most benchmarks are non-interactive in that a model produces output without human involvement. To evaluate human-LM interaction, we develop a new framework, Human-AI Language-based Interaction Evaluation (HALIE), that defines the components of interactive systems and dimensions to consider when designing evaluation metrics. Compared to standard, non-interactive evaluation, HALIE captures (i) the interactive process, not only the final output; (ii) the first-person subjective experience, not just a third-party assessment; and (iii) notions of preference beyond quality (e.g., enjoyment and ownership). We then design five tasks to cover different forms of interaction: social dialogue, question answering, crossword puzzles, summarization, and metaphor generation. With four state-of-the-art LMs (three variants of OpenAI's GPT-3 and AI21 Labs' Jurassic-1), we find that better non-interactive performance does not always translate to better human-LM interaction. In particular, we highlight three cases where the results from non-interactive and interactive metrics diverge and underscore the importance of human-LM interaction for LM evaluation.

Generating a Chain of Thought (CoT) has been shown to consistently improve large language model (LLM) performance on a wide range of NLP tasks. However, prior work has mainly focused on logical reasoning tasks (e.g. arithmetic, commonsense QA); it remains unclear whether improvements hold for more diverse types of reasoning, especially in socially situated contexts. Concretely, we perform a controlled evaluation of zero-shot CoT across two socially sensitive domains: harmful questions and stereotype benchmarks. We find that zero-shot CoT reasoning in sensitive domains significantly increases a model's likelihood to produce harmful or undesirable output, with trends holding across different prompt formats and model variants. Furthermore, we show that harmful CoTs increase with model size, but decrease with improved instruction following. Our work suggests that zero-shot CoT should be used with caution on socially important tasks, especially when marginalized groups or sensitive topics are involved.

Prior work has identified a resilient phenomenon that threatens the performance of human-AI decision-making teams: overreliance, when people agree with an AI, even when it is incorrect. Surprisingly, overreliance does not reduce when the AI produces explanations for its predictions, compared to only providing predictions. Some have argued that overreliance results from cognitive biases or uncalibrated trust, attributing overreliance to an inevitability of human cognition. By contrast, our paper argues that people strategically choose whether or not to engage with an AI explanation, demonstrating empirically that there are scenarios where AI explanations reduce overreliance. To achieve this, we formalize this strategic choice in a cost-benefit framework, where the costs and benefits of engaging with the task are weighed against the costs and benefits of relying on the AI. We manipulate the costs and benefits in a maze task, where participants collaborate with a simulated AI to find the exit of a maze. Through 5 studies (N = 731), we find that costs such as task difficulty (Study 1), explanation difficulty (Study 2, 3), and benefits such as monetary compensation (Study 4) affect overreliance. Finally, Study 5 adapts the Cognitive Effort Discounting paradigm to quantify the utility of different explanations, providing further support for our framework. Our results suggest that some of the null effects found in literature could be due in part to the explanation not sufficiently reducing the costs of verifying the AI's prediction.

Modern multi-agent reinforcement learning frameworks rely on centralized training and reward shaping to perform well. However, centralized training and dense rewards are not readily available in the real world. Current multi-agent algorithms struggle to learn in the alternative setup of decentralized training or sparse rewards. To address these issues, we propose a self-supervised intrinsic reward ELIGN - expectation alignment - inspired by the self-organization principle in Zoology. Similar to how animals collaborate in a decentralized manner with those in their vicinity, agents trained with expectation alignment learn behaviors that match their neighbors' expectations. This allows the agents to learn collaborative behaviors without any external reward or centralized training. We demonstrate the efficacy of our approach across 6 tasks in the multi-agent particle and the complex Google Research football environments, comparing ELIGN to sparse and curiosity-based intrinsic rewards. When the number of agents increases, ELIGN scales well in all multi-agent tasks except for one where agents have different capabilities. We show that agent coordination improves through expectation alignment because agents learn to divide tasks amongst themselves, break coordination symmetries, and confuse adversaries. These results identify tasks where expectation alignment is a more useful strategy than curiosity-driven exploration for multi-agent coordination, enabling agents to do zero-shot coordination.

With the emergence of conversational artificial intelligence (AI) agents, it is important to understand the mechanisms that influence users' experiences of these agents. We study a common tool in the designer's toolkit: conceptual metaphors. Metaphors can present an agent as akin to a wry teenager, a toddler, or an experienced butler. How might a choice of metaphor influence our experience of the AI agent? Sampling metaphors along the dimensions of warmth and competence---defined by psychological theories as the primary axes of variation for human social perception---we perform a study (N=260) where we manipulate the metaphor, but not the behavior, of a Wizard-of-Oz conversational agent. Following the experience, participants are surveyed about their intention to use the agent, their desire to cooperate with the agent, and the agent's usability. Contrary to the current tendency of designers to use high competence metaphors to describe AI products, we find that metaphors that signal low competence lead to better evaluations of the agent than metaphors that signal high competence. This effect persists despite both high and low competence agents featuring human-level performance and the wizards being blind to condition. A second study confirms that intention to adopt decreases rapidly as competence projected by the metaphor increases. In a third study, we assess effects of metaphor choices on potential users' desire to try out the system and find that users are drawn to systems that project higher competence and warmth. These results suggest that projecting competence may help attract new users, but those users may discard the agent unless it can quickly correct with a lower competence metaphor. We close with a retrospective analysis that finds similar patterns between metaphors and user attitudes towards past conversational agents such as Xiaoice, Replika, Woebot, Mitsuku, and Tay.

Generative models often use human evaluations to measure the perceived quality of their outputs. Automated metrics are noisy indirect proxies, because they rely on heuristics or pretrained embeddings. However, up until now, direct human evaluation strategies have been ad-hoc, neither standardized nor validated. Our work establishes a gold standard human benchmark for generative realism. We construct Human eYe Perceptual Evaluation (HYPE) a human benchmark that is (1) grounded in psychophysics research in perception, (2) reliable across different sets of randomly sampled outputs from a model, (3) able to produce separable model performances, and (4) efficient in cost and time.

Visual knowledge bases such as Visual Genome power numerous applications in computer vision, including visual question answering and captioning, but suffer from sparse, incomplete relationships. All scene graph models to date are limited to training on a small set of visual relationships that have thousands of training labels each. Hiring human annotators is expensive, and using textual knowledge base completion methods are incompatible with visual data. In this paper, we introduce a semi-supervised method that assigns probabilistic relationship labels to a large number of unlabeled images using few labeled examples. We analyze visual relationships to suggest two types of image-agnostic features that are used to generate noisy heuristics, whose outputs are aggregated using a factor graph-based generative model. With as few as 10 labeled examples per relationship, the generative model creates enough training data to train any existing state-of-the-art scene graph model. We demonstrate that our method outperforms all baseline approaches on scene graph prediction by5.16 recall@100 for PREDCLS. In our limited label setting, we define a complexity metric for relationships that serves as an indicator (R^2 = 0.778) for conditions under which our method succeeds over transfer learning, the de-facto approach for training with limited labels.