Abstract: AI agents execute multi-step complex tasks directly on artifacts that users own and are accountable for. Interfaces for these agents are progressing from returning only a final output, to surfacing the execution as a sequence of steps, to letting users edit a step in place. Editing in place, however, overwrites the path an edit replaces: the steps that followed are regenerated, and the user gives up a sequence they have not yet evaluated for one they cannot preview. Therefore, users hesitate to intervene where it would be most helpful. We present Pista, a spreadsheet agent that enables users to branch: an edit at any step forks the execution; both paths persist, and users can move between them to compare outcomes. It also pairs each step with an in-situ explanation and surfaces the formula, source range, and output range behind every change. In a formative study (𝑁 = 8) with a stepwise agent probe and a within-subjects evaluation (𝑁 = 16) against a baseline agent, we find gains on both. At matched task success, participants explained the agent’s changes in significantly greater depth (4.38 vs. 2.94, 𝑝 = .015) and identified more issues (2.12 vs. 1.75). They also intervened differently: branching was the most-used feature (15 of 16 participants, median 3 uses per session), and 13 of those 15 used it to check a regenerated path against their expectations. Step-specific edits displaced whole-task re-prompts, cutting re-prompts (2.19 vs. 4.00, 𝑝 < .001) and halving the words each correction took (12.90 vs. 27.63, 𝑝 < .001), because an edit inherits its context from the step it attaches to. Understanding an agent’s actions is therefore not sufficient for steerability. We should provide a safe mechanism for users to explore outputs, weigh different outcomes, and steer the artifact toward what they intended.

Abstract: Creativity is fundamentally human. As AI takes on more of the generative work that once required human imagination, despite documented limitations in creative ability, a critical question emerges: How does GenAI affect users' creativity? Through a within-subject study followed by retrospective interviews with (N=20) programmers, we investigated the impact of LLMs on participants' process of creative thinking in programming and the creativity of generated solutions. Across two conditions (LLM-assisted vs. unassisted), participants using LLMs had significantly shorter idea-generation periods (p=0.0004), leading to fewer creative moments (p=0.002). Qualitative analysis of participants' interactions and interviews revealed four different human-LLM collaboration modes supporting various problem-solving strategies. However, a comparative analysis of the generated solutions shows that while LLMs can help generate more correct and functional code, their solutions contain roughly the same number of ideas as participant-generated ones. Based on our findings, we discuss design implications and considerations for effectively using LLMs to support user creativity.

Abstract: Is AI taking over software engineering research? By 2025, AI appeared in more than half of the research-track papers at ICSE, FSE, and ASE. Yet ubiquity alone cannot show whether AI expanded the publication portfolio, displaced other work, or redirected the intellectual agenda. We investigate this shift through a census of 4,790 research papers (three SE conferences; 2018–2026) and free-form responses from 244 SE researchers discussing the future of the SE research community (ICSE 2026 FoSE survey). We found that while AI is taking over, it is not simply replacing non-AI research; it is unevenly reorganizing SE research around its capabilities. AI accounted for 93.1% of publication growth, but expansion in some venues and problem areas concealed contraction in others. The LLM wave also changed the field's tools more than its problem boundaries: 92.6% of LLM-wave AI papers revisit established SE problems, while most new problems arise from the introduction of LLMs in the SE process themselves. Engagement with AI remains predominantly downstream: 95.7% of AI papers apply AI or evaluate and support its use, while only 4.3% interrogate the paradigm or develop alternatives. The FoSE survey reveals a similar imbalance between recognition and response. AI-driven change was the second-most frequent concern about the field's future 20.2%, yet only 7.0% of responses mentioned AI when researchers were asked to identify one change the community should make. Together, our findings reveal a field reorganizing an inherited agenda around AI faster than it is questioning that reorganization. We discuss how SE can respond by building cumulatively across technological waves, strengthening evaluation and assurance, developing SE-native foundations, and studying software engineering in an AI-transformed world.

Abstract: Large language models are known to be sycophantic, over-complying with users even when doing so is unproductive. When these models act as coding agents, this can be costly: an agent that keeps editing after no edit improves the code writes churn, technical debt, and review burden into the codebase. Yet the consequences of sycophancy in software engineering remain largely understudied. To understand them, we run a controlled study of iterative refactoring, an open-ended task with no stopping criterion, where any edit an agent makes once the code is already clean reflects compliance rather than improvement. Across algorithmic (n=50) and object-oriented (n=10) suites, we repeatedly ask two widely used agents, Codex and Claude Code, to refactor further across five models. Over 3,000 turns, we define six behavioral signals of over-compliance. We hypothesized that a capable agent would stop once it recognized an optimized codebase; none did. Every model kept editing, rewriting up to 31x the original codebase before first stopping and resuming edits even after declaring the task complete. Over-compliance is a costly, multi-turn behavior that compounds with every turn, and our six signals give developers and tool builders a way to detect and curb it.

Abstract: Many agent tasks recur with new data or small changes, yet agents regenerate their reasoning and actions on each run, producing different results. Distilling a prior run into a skill helps the agent avoid earlier mistakes, but it still re-runs the full process each time, so runs can vary and incur that cost. We compile a successful agent run into a reusable neuro-symbolic program (NSP) with a set of guidelines we call FlashScript. Actions and data transformations become code, instance-specific values become parameters, and decisions that need model judgment become LLM calls. We identify three layers where these calls belong and three techniques that replace most of them with code. The compilation does not remove nondeterminism; it localizes it, so the deterministic program and its neural calls can be reviewed and tested separately. Across 120 WorkBench analytics tasks with Claude Haiku 4.5, the NSPs were correct and made identical tool calls on every run, using about 70 times fewer input tokens and a third of the runtime of a skill-guided agent whose runs were inconsistent. In a composite version of the tasks, the skill-guided agent's accuracy fell to 23%, and its output inconsistency increased, while the NSPs remained correct and identical across all runs. The NSPs successfully generalized to all 108 unseen instances of the same templates. Our findings suggest that NSPs from agent trajectories can be considered cumulative software assets for end-users to test, maintain, and reuse in their recurring workflows.

Abstract: Cognitive fatigue transitions from focused attention to inexact responses can cause catastrophic failures in high-stakes environments, yet current black-box assessment techniques ignore the brain's non-Markovian and time-varying interdependent properties, limiting realtime phase transition detection. We develop a fractional dynamical networks-based machine learning (FDNML) framework using coupled fractional-order differential equations to capture brain signal interdependencies and detect cognitive fatigue transitions. Multifractal properties of brain activity exhibit distinct generalized fractal dimension signatures across fatigue levels, with Wasserstein distances of 0.10, 0.13, and 0.08 between states 0-1, 1-2, and 0-2, respectively. The framework achieves 93.33% classification accuracy and 95% AUROC enabling prevention of performance degradation through early detection of neural state transitions.

Abstract: Background: As artificial intelligence (AI) chatbots become an increasingly common source of quick medical guidance, it is important to understand whether their responses meet users' needs and support well-informed health decisions. Yet, existing evaluation frameworks rely primarily on expert-defined evaluation dimensions that have not been empirically validated with end users. It remains unclear whether these frameworks capture the criteria people actually use when judging a response to be useful, trustworthy, and satisfying, or which evaluation dimensions matter most to users in practice. Objective: This study empirically examined how commonly used dimensions such as Accuracy and Comprehensiveness shape caregivers' satisfaction with AI-generated answers to pediatric health questions. We further investigated what expectations and communication needs may be overlooked by current evaluation frameworks. Methods: We conducted a mixed methods vignette survey with 191 caregivers recruited through Prolific. Participants evaluated GPT-4o responses to a set of clinician-approved pediatric health questions across 8 dimensions and rated overall satisfaction. We quantified the influence of each dimension on overall satisfaction using a Cumulative Link Mixed Model and performed an inductive thematic analysis of open-ended comments to identify gaps in established frameworks. Results: A total of 191 caregivers evaluated 1146 chatbot responses. Initially, caregivers rated Accuracy and Credibility as the most important dimensions. However, Cumulative Link Mixed Model analysis identified Usefulness as the strongest driver of overall satisfaction (odds ratio [OR] 2.53, 95% CI 1.95-3.27; P<.001), followed by Thoroughness (OR 2.15, 95% CI 1.69-2.73; P<.001). Comprehensiveness did not significantly influence satisfaction (OR 1.05, 95% CI 0.82-1.33; P=.71). Qualitative feedback helped explain this: participants frequently criticized responses as “too long” and preferred concise, actionable guidance. Empathy/Warmth was significantly associated with overall satisfaction (OR 1.48, 95% CI 1.27-1.72; P<.001) but elicited polarized reactions (Van der Eijk agreement μ=0.38): some caregivers valued emotional support, while others found AI-generated empathy insincere and undermined credibility. Medical disclaimers increased trust in higher-risk situations but reduced confidence in lower-risk scenarios. Conclusions: Existing evaluation frameworks only partially capture how caregivers assess medical chatbots. Caregivers valued actionable guidance, credible references, and clear reasoning over lengthy, exhaustive detail. Rather than passively receiving dense information, they preferred an interactive style in which the chatbot proactively proposed follow-up suggestions, helping them steer the conversation toward their specific needs. Reactions to empathetic language and medical disclaimers were context-dependent: features that built trust in some situations could seem insincere, excessive, or unnecessary in others. These findings suggest that future medical chatbots should move beyond one-size-fits-all communication and adapt to individual users' situations, preferences, and information needs. Evaluation protocols should likewise assess not only whether chatbot responses are accurate and comprehensive but also whether they are actionable, appropriately toned, and responsive to users' evolving needs over the course of a conversation.

Abstract: The widespread adoption of Large Language Models (LLMs) in software development is transforming programming from a generative activity into one dominated by prompt engineering and AI-generated solution evaluation. This shift opens the pathway for new cognitive challenges that amplify existing decision-making biases or create entirely novel ones. One such type of challenge is cognitive biases, which are thinking patterns that lead people away from logical reasoning, often resulting in errors, poor decisions, or sub-optimal actions. Despite LLMs becoming integral to modern development workflows, we lack a systematic understanding of how cognitive biases manifest and impact developer decision-making in these AI-collaborative contexts. This paper presents the first comprehensive study of cognitive biases in LLM-assisted programming using a mixed-methods approach of observational studies with n=14 student and professional developers, followed by surveys with n=22 additional developers. First, we qualitatively analyze our data based on bias categorization in the traditional non-LLM workflow reported in prior work. Our findings suggest that the traditional software development biases are inadequate to explain/understand why LLM-related actions are more likely to be biased. Through systematic analysis of 239 cognitive bias types, we develop a novel taxonomy of 15 bias categories of 90 biases specific to developer-LLM interactions, validated with cognitive psychologists. We found that 48.8% of total programmer actions are biased, and programmer-LLM interactions included 56.4% of biased actions. Based on our survey analysis, we present practical tools and practices for programmers, along with recommendations for LLM-based code generation tool builders to help mitigate cognitive biases in human-AI programming.

Abstract: Large Language Models (LLMs) such as ChatGPT have quickly become part of student programmers' toolkits, whether allowed by instructors or not. This paper examines how introductory programming (CS1) students integrate LLMs into their problem-solving processes. We conducted a mixed-methods study with 14 undergraduates completing three programming tasks while thinking aloud and permitted to access any resources they choose. The tasks varied in open-endedness and familiarity to the participants and were followed by surveys and interviews. We find that students frequently adopt a pattern we call pseudo-apprenticeship, where students engage attentively with expert-level solutions provided by LLMs but fail to participate in the stages of cognitive apprenticeship that promote independent problem-solving. This pattern was augmented by disconnects between students' intentions, actions, and self-perceived behavior when using LLMs. We offer design and instructional interventions for promoting learning and addressing the patterns of dependent AI use observed.

Abstract: