Humans act via a nuanced process that depends both on rational deliberation and also on identity and contextual factors. In this work, we study how large language models (LLMs) can simulate human action in the context of social dilemma games. While prior work has focused on "steering" (weak binding) of chat models to simulate personas, we analyze here how deep binding of base models with extended backstories leads to more faithful replication of identity-based behaviors. Our study has these findings: simulation fidelity vs human studies is improved by conditioning base LMs with rich context of narrative identities and checking consistency using instruction-tuned models. We show that LLMs can also model contextual factors such as time (year that a study was performed), question framing, and participant pool effects. LLMs, therefore, allow us to explore the details that affect human studies but which are often omitted from experiment descriptions, and which hamper accurate replication.

Large language models (LLMs) show potential as simulators of human behavior, offering a scalable way to study responses to interventions. However, because LLMs are trained largely on observational data, interventions in experiments with LLM-simulated synthetic users can induce unintended shifts in latent user attributes, causing user drift where the implicit simulated population differs across treatment conditions, potentially distorting effect estimates. We formalize the confounding or selection bias that can arise due to user drift and show how intervention-dependent shifts can inflate or attenuate observed differences in user responses under intervention. To diagnose confounding, we propose using negative control outcomes–attributes that should remain invariant under intervention–to identify distribution shifts across intervention conditions, providing evidence of user drift. To mitigate drift, we study adjusting the persona specification by eliciting additional confounders, finding that targeted, setting-relevant confounders can substantially reduce bias across survey-style and multi-turn agent evaluations.

Large language models (LLMs) are increasingly capable of simulating human behavior, offering cost-effective ways to estimate user responses to various surveys and polls. However, the questions in these surveys usually reflect socially understood attitudes: the patterns of attitudes of old/young, liberal/conservative, as understood by both members and non-members of those groups. It is not clear whether the LLM binding is deep, meaning the LLM answers as a member of a particular in-group would, or shallow, meaning the LLM responds as an out-group member believes an in-group member would. To explore this difference, we use questions that expose known in-group/out-group biases. This level of fidelity is critical for applying LLMs to various political science studies, including timely topics on polarization dynamics, inter-group conflict, and democratic backsliding. To this end, we propose a novel methodology for constructing virtual personas with synthetic user "backstories" generated as extended, multi-turn interview transcripts. This approach is justified by the theory of narrative identity which argues that personality at the highest level is constructed from self-narratives. Our generated backstories are longer, rich in detail, and consistent in authentically describing a singular individual, compared to previous methods. We show that virtual personas conditioned on our backstories closely replicate human response distributions (up to an 87

The language generation and reasoning capabilities of large language models (LLMs) have enabled conversational systems with impressive performance in a variety of tasks, from code generation, to composing essays, to passing STEM and legal exams, to a new paradigm for knowledge search. Besides those short-term use applications, LLMs are increasingly used to help with real-life goals or tasks that take a long time to complete, involving multiple sessions across days, weeks, months, or even years. Thus to enable conversational systems for long term interactions and tasks, we need language-based agents that can plan for long horizons. Traditionally, such capabilities were addressed by reinforcement learning agents with hierarchical planning capabilities. In this work, we explore a novel architecture where the LLM acts as the meta-controller deciding the agent's next macro-action, and tool use augmented LLM-based option policies execute the selected macro-action. We instantiate this framework for a specific set of macro-actions enabling adaptive planning for users' personal plans through conversation and follow-up questions collecting user feedback. We show how this paradigm can be applicable in scenarios ranging from tutoring for academic and non-academic tasks to conversational coaching for personal health plans.

Many online platforms incorporate engagement signals, such as likes, into their interface design to boost engagement. However, these signals can unintentionally elevate content that may not support normatively desirable behavior, especially when toxic content correlates strongly with popularity indicators. In this study, we propose structured prosocial feedback as a complementary signal, which highlights content quality based on normative criteria. We design and implement an LLM-based feedback system, which evaluates user comments based on principles from positive psychology, such as individual well-being. A pre-registered user study then examines how existing peer-based (popularity) and the new expert-based feedback interact to shape users' reposting behavior in a social media setting. Results show that peer feedback increases conformity to popularity cues, while expert feedback shifts choices toward normatively higher-quality content. This illustrates the added value of normative cues and underscores the potential benefits of incorporating such signals into platform feedback systems to foster healthier online environments.

Moral foundations theory (MFT) is a psychological assessment tool that decomposes human moral reasoning into five factors, including care/harm, liberty/oppression, and sanctity/degradation (Graham et al., 2009). People vary in the weight they place on these dimensions when making moral decisions, in part due to their cultural upbringing and political ideology. As large language models (LLMs) are trained on datasets collected from the internet, they may reflect the biases that are present in such corpora. This paper uses MFT as a lens to analyze whether popular LLMs have acquired a bias towards a particular set of moral values. We analyze known LLMs and find they exhibit particular moral foundations, and show how these relate to human moral foundations and political affiliations. We also measure the consistency of these biases, or whether they vary strongly depending on the context of how the model is prompted. Finally, we show that we can adversarially select prompts that encourage the moral to exhibit a particular set of moral foundations, and that this can affect the model's behavior on downstream tasks. These findings help illustrate the potential risks and unintended consequences of LLMs assuming a particular moral stance.

Despite recent advances in multimodal pre-training for visual description, state-of-the-art models still produce captions containing errors, such as hallucinating objects not present in a scene. The existing prominent metric for object hallucination, CHAIR, is limited to a fixed set of MS COCO objects and synonyms. In this work, we propose a modernized open-vocabulary metric, ALOHa, which leverages large language models (LLMs) to measure object hallucinations. Specifically, we use an LLM to extract groundable objects from a candidate caption, measure their semantic similarity to reference objects from captions and object detections, and use Hungarian matching to produce a final hallucination score. We show that ALOHa correctly identifies 13.6% more hallucinated objects than CHAIR on HAT, a new gold-standard subset of MS COCO Captions annotated for hallucinations, and 30.8% more on nocaps, where objects extend beyond MS COCO categories.

Online Knowledge Repositories (OKRs) like Wikipedia offer communities a way to share and preserve information about themselves and their ways of living. However, for communities with low-resourced languages—including most African communities—the quality and volume of content available are often inadequate. One reason for this lack of adequate content could be that many OKRs embody Western ways of knowledge preservation and sharing, requiring many low-resourced language communities to adapt to new interactions. To understand the challenges faced by low-resourced language contributors on the popular OKR Wikipedia, we conducted (1) a thematic analysis of Wikipedia forum discussions and (2) a contextual inquiry study with 14 novice contributors. We focused on three Ethiopian languages: Afan Oromo, Amharic, and Tigrinya. Our analysis revealed several recurring themes; for example, contributors struggle to find resources to corroborate their articles in low-resourced languages, and language technology support, like translation systems and spellcheck, result in several errors that waste contributors’ time. We hope our study will support designers in making online knowledge repositories accessible to low-resourced language speakers.

Current Large Language Models (LLMs) are not only limited to some maximum context length, but also are not able to robustly consume long inputs. To address these limitations, we propose ReadAgent, an LLM agent system that increases effective context length up to 20x in our experiments. Inspired by how humans interactively read long documents, we implement ReadAgent as a simple prompting system that uses the advanced language capabilities of LLMs to (1) decide what content to store together in a memory episode, (2) compress those memory episodes into short episodic memories calledgist memories, and (3) take actions to look up passages in the original text if ReadAgent needs to remind itself of relevant details to complete a task. We evaluate ReadAgent against baselines using retrieval methods, using the original long contexts, and using the gist memories. These evaluations are performed on three long-document reading comprehension tasks: QuALITY, NarrativeQA, and QMSum. ReadAgent outperforms the baselines on all three tasks while extending the effective context window by 3.5-20x.

The evaluation of machine-generated image captions poses an interesting yet persistent challenge. Effective evaluation measures must consider numerous dimensions of similarity, including semantic relevance, visual structure, object interactions, caption diversity, and specificity. Existing highly-engineered measures attempt to capture specific aspects, but fall short in providing a holistic score that aligns closely with human judgments. Here, we propose CLAIR, a novel method that leverages the zero-shot language modeling capabilities of large language models (LLMs) to evaluate candidate captions. In our evaluations, CLAIR demonstrates a stronger correlation with human judgments of caption quality compared to existing measures. Notably, on Flickr8K-Expert, CLAIR achieves relative correlation improvements over SPICE of 39.6% and over image-augmented methods such as RefCLIP-S of 18.3%. Moreover, CLAIR provides noisily interpretable results by allowing the language model to identify the underlying reasoning behind its assigned score.

Investigative journalists and public defenders conduct the essential work of examining, reporting, and arguing critical cases around police use-of-force and misconduct. In an ideal world, they would have access to well-organized records they can easily navigate and search. In reality, records can come as large, disorganized data dumps, increasing the burden on the already resource-constrained teams. In a cross-disciplinary research team of stakeholders and computer scientists, we worked closely with public defenders and investigative journalists in the United States to co-design an AI-augmented tool that addresses challenges in working with such data dumps. Our Document Organization Tool (DOT) is a Python library that has data cleaning, extraction, and organization features. Our collaborative design process gave us insights into the needs of under-resourced teams who work with large data dumps, such as how some domain experts became self-taught programmers to automate their tasks. To understand what type of programming paradigm could support our target users, we conducted a user study (n=18) comparing visual, programming-by-example, and traditional text-based programming tools. From our user study, we found that once users passed the initial learning stage, they could comfortably use all three paradigms. Our work offers insights for designers working with under-resourced teams who want to consolidate cutting-edge algorithms and AI techniques into unified, expressive tools. We argue user-centered tool design can contribute to the broader fight for accountability and transparency by supporting existing practitioners in their work in domains like criminal justice.

Traditional automated metrics for evaluating conditional natural language generation use pairwise comparisons between a single generated text and the best-matching gold-standard ground truth text. When multiple ground truths are available, scores are aggregated using an average or max operation across references. While this approach works well when diversity in the ground truth data (i.e. dispersion of the distribution of conditional texts) can be ascribed to noise, such as in automated speech recognition, it does not allow for robust evaluation in the case where diversity in the ground truths represents signal for the model. In this work we argue that existing metrics are not appropriate for domains such as visual description or summarization where ground truths are semantically diverse, and where the diversity in those captions captures useful additional information about the context. We propose a novel paradigm for multi-candidate evaluation of conditional language generation models, and a new family of metrics that compare the distributions of reference and model-generated caption sets using small sample sets of each. We demonstrate the utility of our approach with a case study in visual description: where we show that existing models optimize for single-description quality over diversity, and gain some insights into how sampling methods and temperature impact description quality and diversity.

Recent work in machine learning and cognitive science has suggested that understanding causal information is essential to the development of intelligence. The extensive literature in cognitive science using the ``blicket detector'' environment shows that children are adept at many kinds of causal inference and learning. We propose to adapt that environment for machine learning agents. One of the key challenges for current machine learning algorithms is modeling and understanding causal overhypotheses: transferable abstract hypotheses about sets of causal relationships. In contrast, even young children spontaneously learn and use causal overhypotheses. In this work, we present a new benchmark -- a flexible environment which allows for the evaluation of existing techniques under variable causal overhypotheses -- and demonstrate that many existing state-of-the-art methods have trouble generalizing in this environment. The code and resources for this benchmark are available at https://github.com/CannyLab/casual_overhypotheses.

If you ask a human to describe an image, they might do so in a thousand different ways. Traditionally, image captioning models are trained to generate a single “best’ (most like a reference) image caption. Unfortunately, doing so encourages captions that are “informationally impoverished,’ and focus on only a subset of the possible details, while ignoring other potentially useful information in the scene. In this work, we introduce a simple, yet novel, method: “Image Captioning by Committee Consensus’ (IC3), designed to generate a single caption that captures high-level details from several annotator viewpoints. Humans rate captions produced by IC3 at least as helpful as baseline SOTA models more than two thirds of the time, and IC3 can improve the performance of SOTA automated recall systems by up to 84%, outperforming single human-generated reference captions, and indicating significant improvements over SOTA approaches for visual description. Code is available at [https://davidmchan.github.io/caption-by-committee/](https://davidmchan.github.io/caption-by-committee/)

Online knowledge repositories like Wikipedia offer a way for communities to share and preserve information about themselves and their ways of living. However, there is a huge gap in the volume and quality of content available for communities that speak high-resourced languages versus communities that speak low-resourced languages—including a majority of African communities.Usually, such online repositories are situated in Western ways of knowledge preservation and sharing, requiring low-resourced language communities to adapt to new forms of interaction and preservation.To understand the challenges faced by low-resourced language content creators on the popular knowledge repository Wikipedia, we collected Wikipedia forum discussions in low-resourced languages and conducted a thematic analysis. For the purpose of this work, we focused on three Ethiopian languages: Amharic, Tigrinya, and Afan Oromo. In this short paper, we report findings from ongoing research aimed at understanding the challenges faced by such content creators.Our analysis reveals several recurrent themes, including (1) how typing in non-Latin scripts on Latin keyboards is challenging and slow, and (2) how content creators' time is consumed by cleaning duplicate and low-quality articles.We hope our study will help inform designers' choices in making such platforms accessible to low-resourced language speakers.

A number of recent self-supervised learning methods have shown impressive performance on image classification and other tasks. A somewhat bewildering variety of techniques have been used, not always with a clear understanding of the reasons for their benefits, especially when used in combination. Here we treat the embeddings of images as point particles and consider model optimization as a dynamic process on this system of particles. Our dynamic model combines an attractive force for similar images, a locally dispersive force to avoid local collapse, and a global dispersive force to achieve a globally-homogeneous distribution of particles. The dynamic perspective highlights the advantage of using a delayed-parameter image embedding (a la BYOL) together with multiple views of the same image. It also uses a purely-dynamic local dispersive force (Brownian motion) that shows improved performance over other methods and does not require knowledge of other particle coordinates. The method is called MSBReg which stands for (i) a Multiview centroid loss, which applies an attractive force to pull different image view embeddings toward their centroid, (ii) a Singular value loss, which pushes the particle system toward spatially homogeneous density, (iii) a Brownian diffusive loss. We evaluate downstream classification performance of MSBReg on ImageNet as well as transfer learning tasks including fine-grained classification, multi-class object classification, object detection, and instance segmentation. In addition, we also show that applying our regularization term to other methods further improves their performance and stabilize the training by preventing a mode collapse.

With the growing adoption of short-form video by social media platforms, reducing the spread of misinformation through video posts has become a critical challenge for social media providers. In this paper, we develop methods to detect misinformation in social media posts, exploiting modalities such as video and text. Due to the lack of large-scale public data for misinformation detection in multi-modal datasets, we collect 160,000 video posts from Twitter, and leverage self-supervised learning to learn expressive representations of joint visual and textual data. In this work, we propose two new methods for detecting semantic inconsistencies within short-form social media video posts, based on contrastive learning and masked language modeling. We demonstrate that our new approaches outperform current state-of-the-art methods on both artificial data generated by random-swapping of positive samples and in the wild on a new manually-labeled test set for semantic misinformation.

Generating representations of video data is of key importance in advancing the field of machine perception. Most current techniques rely on hand-annotated data, which can be difficult to work with, expensive to generate, and hard to scale. In this work, we propose a novel learning approach based on contrastive learning, LAVA, which is capable of learning joint language, audio, and video representations in a self-supervised manner. We pre-train LAVA on the Kinetics 700 dataset using transformer encoders to learn representations for each modality. We then demonstrate that LAVA performs competitively with the current state-of-the-art self-supervised and weakly-supervised pretraining techniques on UCF-101 and HMDB-51 video action recognition while using a fraction of the unlabeled data.

Despite recent progress in reinforcement learning (RL), RL algorithms for exploration still remain an active area of research. Existing methods often focus on state-based metrics, which do not consider the underlying causal structures of the environment, and while recent research has begun to explore RL environments for causal learning, these environments primarily leverage causal information through causal inference or induction rather than exploration. In contrast, human children - some of the most proficient explorers - have been shown to use causal information to great benefit. In this work, we introduce a novel RL environment designed with a controllable causal structure, which allows us to evaluate exploration strategies used by both agents and children in a unified environment. In addition, through experimentation on both computation models and children, we demonstrate that there are significant differences between information-gain optimal RL exploration in causal environments and the exploration of children in the same environments. We conclude with a discussion of how these findings may inspire new directions of research into efficient exploration and disambiguation of causal structures for RL algorithms.

A fundamental challenge in manipulating fabric for clothes folding and textiles manufacturing is computing "pick points" to effectively modify the state of an uncertain manifold. We present a supervised deep transfer learning approach to locate pick points using depth images for invariance to color and texture. We consider the task of bed-making, where a robot sequentially grasps and pulls at pick points to increase blanket coverage. We perform physical experiments with two mobile manipulator robots, the Toyota HSR and the Fetch, and three blankets of different colors and textures. We compare coverage results from (1) human supervision, (2) a baseline of picking at the uppermost blanket point, and (3) learned pick points. On a quarter-scale twin bed, a model trained with combined data from the two robots achieves 92% blanket coverage compared with 83% for the baseline and 95% for human supervisors. The model transfers to two novel blankets and achieves 93% coverage. Average coverage results of 92% for 193 beds suggest that transfer-invariant robot pick points on fabric can be effectively learned.