Fumeng Yang

Taehyun Yang, Eunhye Kim, Zhongzheng Xu et al.

Generative AI tools have lowered barriers to producing branded social media images and captions, yet small-business owners (SBOs) still struggle to create on-brand posts without access to professional designers or marketing consultants. Although these tools enable fast image generation from text prompts, aligning outputs with a brand's intended look and feel remains a demanding, iterative task. In this position paper, we explore how SBOs navigate iterative content creation and how AI-assisted systems can support SBOs' content creation workflow. We conducted a preliminary study with 12 SBOs who independently manage their businesses and social media presence, using a questionnaire to collect their branding practices, content workflows, and use of generative AI alongside conventional design tools. We identified three recurring challenges: (1) translating brand "feel" into effective prompts, (2) difficulty revisiting and comparing prior image generations, and (3) difficulty making sense of changes between iterations to steer refinement. Based on these findings, we present a prototype that scaffolds brand articulation, supports feedback-informed exploration, and maintains a traceboard of branching image iterations. Our work illustrates how traces of the iterative process can serve as workflow support that helps SBOs keep track of explorations, make sense of changes, and refine content.

Nishant Balepur, Vishakh Padmakumar, Fumeng Yang et al.

LLMs are aligned to follow input instructions by learning which of two responses users prefer for a prompt. However, such preference data do not convey why users prefer responses that are chosen or rejected, so LLMs trained on these datasets cannot tailor responses to varied user needs. To surface these parameters of personalization, we apply abductive reasoning to preference data, inferring needs and interests of users, i.e., personas, that may prefer either response. We test this idea in two steps: Persona Inference (PI), abductively inferring personas of users who prefer chosen or rejected outputs, and Persona Tailoring (PT), training models to tailor outputs to personas from PI. We show: 1) LLMs infer personas accurately explaining why different users may prefer both chosen or rejected outputs; 2) Training on preference data augmented with PI personas via PT boosts personalization and generalizes to supporting user-written personas; and 3) Rejected response personas form harder personalization evaluations, showing PT better aids users with uncommon preferences versus typical alignment methods. We argue for an abductive view of preferences for personalization, asking not only which response is better but when, why, and for whom.

Nishant Balepur, Matthew Shu, Yoo Yeon Sung et al. · allen-ai, oxford

To assist users in complex tasks, LLMs generate plans: step-by-step instructions towards a goal. While alignment methods aim to ensure LLM plans are helpful, they train (RLHF) or evaluate (ChatbotArena) on what users prefer, assuming this reflects what helps them. We test this with Planorama: an interface where 126 users answer 300 multi-step questions with LLM plans. We get 4388 plan executions and 5584 comparisons to measure plan helpfulness (QA success) and user preferences on plans, and recreate the setup in agents and reward models to see if they simulate or prefer what helps users. We expose: 1) user/model preferences and agent success do not accurately predict which plans help users, so common alignment feedback can misalign with helpfulness; 2) this gap is not due to user-specific preferences, as users are similarly successful when using plans they prefer/disprefer; 3) surface-level cues like brevity and question similarity strongly link to preferences, but such biases fail to predict helpfulness. In all, we argue aligning helpful LLMs needs feedback from real user interactions, not just preferences of what looks helpful, so we discuss the plan NLP researchers can execute to solve this problem.

Caitlyn M. McColeman, Fumeng Yang, Steven Franconeri et al.

Data can be visually represented using visual channels like position, length or luminance. An existing ranking of these visual channels is based on how accurately participants could report the ratio between two depicted values. There is an assumption that this ranking should hold for different tasks and for different numbers of marks. However, there is little existing work testing assumption, especially given that visually computing ratios is relatively unimportant in real-world visualizations, compared to seeing, remembering, and comparing trends and motifs, across displays that almost universally depict more than two values. We asked participants to immediately reproduce a set of values from memory. With a Bayesian multilevel modeling approach, we observed how the relevant rank positions of visual channels shift across different numbers of marks (2, 4 or 8) and for bias, precision, and error measures. The ranking did not hold, even for reproductions of only 2 marks, and the new ranking was highly inconsistent for reproductions of different numbers of marks. Other factors besides channel choice far more influence on performance, such as the number of values in the series (e.g. more marks led to larger errors), or the value of each mark (e.g. small values are systematically overestimated). Recall was worse for displays with 8 marks than 4, consistent with established limits on visual memory. These results show that we must move beyond two-value ratio judgments as a baseline for ranking the quality of a visual channel, including testing new tasks (detection of trends or motifs), timescales (immediate computation, or later comparison), and the number of values (from a handful, to thousands).