Large language models promise a broad set of functions, but when not given a specific objective, they default to generic results. We demonstrate that inferring the user’s in-the-moment objective, then rapidly optimizing for that singular objective, enables LLMs to produce specialized tools, interfaces, and responses. Our work introduces just-in-time objectives, which model a user’s goals to specialize LLM systems on the fly. We contribute an architecture for automatically inducing such objectives by passively observing user behavior, then steering downstream AI systems through generation and evaluation against this objective. Inducing just-in-time objectives (e.g., “Clarify the abstract’s research contribution”) enables automatic generation of tools, e.g., those that critique a draft based on relevant HCI methodologies, anticipate related researchers’ reactions, or surface ambiguous terminology. In a series of experiments on participants’ own tasks, JIT objectives enable LLM outputs that achieve 66–86% win rates over typical LLMs. In-person use sessions confirm that JIT objectives produce specialized tools that are unique to each participant and are rated as significantly higher quality than a standard LLM chat tool.

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.

Figure 1: Policy maps chart LLM policy coverage over an unbounded space of model behaviors.Here, an AI practitioner is designing a policy for how an LLM should summarize violent text.Policy map abstractions (right) allow the policy designer to interactively author and test policies that govern a model's behavior using if-then rules over concepts.The designer can create any desired concept by providing a simple text definition to capture cases of model behavior.Our Policy Projector tool (center) renders cases, concepts, and policies as visual map layers to aid iterative policy design.

LLM chains enable complex tasks by decomposing work into a sequence of subtasks. Similarly, the more established techniques of crowdsourcing workflows decompose complex tasks into smaller tasks for human crowdworkers. Chains address LLM errors analogously to the way crowdsourcing workflows address human error. To characterize opportunities for LLM chaining, we survey 107 papers across the crowdsourcing and chaining literature to construct a design space for chain development. The design space covers a designer’s objectives and the tactics used to build workflows. We then surface strategies that mediate how workflows use tactics to achieve objectives. To explore how techniques from crowdsourcing may apply to chaining, we adapt crowdsourcing workflows to implement LLM chains across three case studies: creating a taxonomy, shortening text, and writing a short story. From the design space and our case studies, we identify takeaways for effective chain design and raise implications for future research and development.

Data analysts have long sought to turn unstructured text data into meaningful concepts. Though common, topic modeling and clustering focus on lower-level keywords and require significant interpretative work. We introduce concept induction, a computational process that instead produces high-level concepts, defined by explicit inclusion criteria, from unstructured text. For a dataset of toxic online comments, where a state-of-the-art BERTopic model outputs “women, power, female,” concept induction produces high-level concepts such as “Criticism of traditional gender roles” and “Dismissal of women’s concerns.” We present LLooM, a concept induction algorithm that leverages large language models to iteratively synthesize sampled text and propose human-interpretable concepts of increasing generality. We then instantiate LLooM in a mixed-initiative text analysis tool, enabling analysts to shift their attention from interpreting topics to engaging in theory-driven analysis. Through technical evaluations and four analysis scenarios ranging from literature review to content moderation, we find that LLooM’s concepts improve upon the prior art of topic models in terms of quality and data coverage. In expert case studies, LLooM helped researchers to uncover new insights even from familiar datasets, for example by suggesting a previously unnoticed concept of attacks on out-party stances in a political social media dataset.

The standard way to teach models is by feeding them lots of data. However, this approach often teaches models incorrect ideas because they pick up on misleading signals in the data. To prevent such misconceptions, we must necessarily provide additional information beyond the training data. Prior methods incorporate additional instance-level supervision, such as labels for misleading features or additional labels for debiased data. However, such strategies require a large amount of labeler effort. We hypothesize that people are good at providing textual feedback at the concept level, a capability that existing teaching frameworks do not leverage. We propose Clarify, a novel interface and method for interactively correcting model misconceptions. Through Clarify, users need only provide a short text description of a model’s consistent failure patterns. Then, in an entirely automated way, we use such descriptions to improve the training process. Clarify is the first end-to-end system for user model correction. Our user studies show that non-expert users can successfully describe model misconceptions via Clarify, leading to increased worst-case performance in two datasets. We additionally conduct a case study on a large-scale image dataset, ImageNet, using Clarify to find and rectify 31 novel hard subpopulations.

Mounting evidence indicates that the artificial intelligence (AI) systems that rank our social media feeds bear nontrivial responsibility for amplifying partisan animosity: negative thoughts, feelings, and behaviors toward political out-groups. Can we design these AIs to consider democratic values such as mitigating partisan animosity as part of their objective functions? We introduce a method for translating established, vetted social scientific constructs into AI objective functions, which we term societal objective functions, and demonstrate the method with application to the political science construct of anti-democratic attitudes. Traditionally, we have lacked observable outcomes to use to train such models-however, the social sciences have developed survey instruments and qualitative codebooks for these constructs, and their precision facilitates translation into detailed prompts for large language models. We apply this method to create a democratic attitude model that estimates the extent to which a social media post promotes anti-democratic attitudes, and test this democratic attitude model across three studies. In Study 1, we first test the attitudinal and behavioral effectiveness of the intervention among US partisans (N=1,380) by manually annotating (alpha=.895) social media posts with anti-democratic attitude scores and testing several feed ranking conditions based on these scores. Removal (d=.20) and downranking feeds (d=.25) reduced participants' partisan animosity without compromising their experience and engagement. In Study 2, we scale up the manual labels by creating the democratic attitude model, finding strong agreement with manual labels (rho=.75). Finally, in Study 3, we replicate Study 1 using the democratic attitude model instead of manual labels to test its attitudinal and behavioral impact (N=558), and again find that the feed downranking using the societal objective function reduced partisan animosity (d=.25). This method presents a novel strategy to draw on social science theory and methods to mitigate societal harms in social media AIs.

Machine learning practitioners often end up tunneling on low-level technical details like model architectures and performance metrics. Could early model development instead focus on high-level questions of which factors a model ought to pay attention to? Inspired by the practice of sketching in design, which distills ideas to their minimal representation, we introduce model sketching: a technical framework for iteratively and rapidly authoring functional approximations of a machine learning model’s decision-making logic. Model sketching refocuses practitioner attention on composing high-level, human-understandable concepts that the model is expected to reason over (e.g., profanity, racism, or sarcasm in a content moderation task) using zero-shot concept instantiation. In an evaluation with 17 ML practitioners, model sketching reframed thinking from implementation to higher-level exploration, prompted iteration on a broader range of model designs, and helped identify gaps in the problem formulation—all in a fraction of the time ordinarily required to build a model.

Algorithm audits are powerful tools for studying black-box systems without direct knowledge of their inner workings. While very effective in examining technical components, the method stops short of a sociotechnical frame, which would also consider users themselves as an integral and dynamic part of the system. Addressing this limitation, we propose the concept of sociotechnical auditing: auditing methods that evaluate algorithmic systems at the sociotechnical level, focusing on the interplay between algorithms and users as each impacts the other. Just as algorithm audits probe an algorithm with varied inputs and observe outputs, a sociotechnical audit (STA) additionally probes users, exposing them to different algorithmic behavior and measuring their resulting attitudes and behaviors. As an example of this method, we develop Intervenr, a platform for conducting browser-based, longitudinal sociotechnical audits with consenting, compensated participants. Intervenr investigates the algorithmic content users encounter online, and also coordinates systematic client-side interventions to understand how users change in response. As a case study, we deploy Intervenr in a two-week sociotechnical audit of online advertising (N = 244) to investigate the central premise that personalized ad targeting is more effective on users. In the first week, we observe and collect all browser ads delivered to users, and in the second, we deploy an ablation-style intervention that disrupts normal targeting by randomly pairing participants and swapping all their ads. We collect user-oriented metrics (self-reported ad interest and feeling of representation) and advertiser-oriented metrics (ad views, clicks, and recognition) throughout, along with a total of over 500,000 ads. Our STA finds that targeted ads indeed perform better with users, but also that users begin to acclimate to different ads in only a week, casting doubt on the primacy of personalized ad targeting given the impact of repeated exposure. In comparison with other evaluation methods that only study technical components, or only experiment on users, sociotechnical audits evaluate sociotechnical systems through the interplay of their technical and human components.

Whose labels should a machine learning (ML) algorithm learn to emulate? For ML tasks ranging from online comment toxicity to misinformation detection to medical diagnosis, different groups in society may have irreconcilable disagreements about ground truth labels. Supervised ML today resolves these label disagreements implicitly using majority vote, which overrides minority groups’ labels. We introduce jury learning, a supervised ML approach that resolves these disagreements explicitly through the metaphor of a jury: defining which people or groups, in what proportion, determine the classifier’s prediction. For example, a jury learning model for online toxicity might centrally feature women and Black jurors, who are commonly targets of online harassment. To enable jury learning, we contribute a deep learning architecture that models every annotator in a dataset, samples from annotators’ models to populate the jury, then runs inference to classify. Our architecture enables juries that dynamically adapt their composition, explore counterfactuals, and visualize dissent. A field evaluation finds that practitioners construct diverse juries that alter 14% of classification outcomes.

Because algorithm audits are conducted by technical experts, audits are necessarily limited to the hypotheses that experts think to test. End users hold the promise to expand this purview, as they inhabit spaces and witness algorithmic impacts that auditors do not. In pursuit of this goal, we propose end-user audits-system-scale audits led by non-technical users-and present an approach that scaffolds end users in hypothesis generation, evidence identification, and results communication. Today, performing a system-scale audit requires substantial user effort to label thousands of system outputs, so we introduce a collaborative filtering technique that leverages the algorithmic system's own disaggregated training data to project from a small number of end user labels onto the full test set. Our end-user auditing tool, IndieLabel, employs these predicted labels so that users can rapidly explore where their opinions diverge from the algorithmic system's outputs. By highlighting topic areas where the system is under-performing for the user and surfacing sets of likely error cases, the tool guides the user in authoring an audit report. In an evaluation of end-user audits on a popular comment toxicity model with 17 non-technical participants, participants both replicated issues that formal audits had previously identified and also raised previously underreported issues such as under-flagging on veiled forms of hate that perpetuate stigma and over-flagging of slurs that have been reclaimed by marginalized communities.

Expert crowdsourcing marketplaces have untapped potential to empower workers' career and skill development. Currently, many workers cannot afford to invest the time and sacrifice the earnings required to learn a new skill, and a lack of experience makes it difficult to get job offers even if they do. In this paper, we seek to lower the threshold to skill development by repurposing existing tasks on the marketplace as mentored, paid, real-world work experiences, which we refer to as micro-internships. We instantiate this idea in Atelier, a micro-internship platform that connects crowd interns with crowd mentors. Atelier guides mentor-intern pairs to break down expert crowdsourcing tasks into milestones, review intermediate output, and problem-solve together. We conducted a field experiment comparing Atelier's mentorship model to a non-mentored alternative on a real-world programming crowdsourcing task, finding that Atelier helped interns maintain forward progress and absorb best practices.