Paper deep dive
Towards Cognitive Process-Aware Proactive Writing Support
Masahiro Yoshida, Atsuya Kobayashi, Kei Tateno, Xiang 'Anthony' Chen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 9/15/2026, 8:38:55 AM
Summary
This paper introduces AToM CoWriter, a proactive writing support system that leverages Flower and Hayes' cognitive process theory to infer user needs from writing interactions. By mapping observable behaviors (keystrokes, cursor movement) to six cognitive processes, the system selects from 14 specific support types via Functional Agents. Two user studies (N=21) demonstrate that this approach improves expressiveness and increases engagement with suggestions compared to prompt-based methods.
Entities (8)
Relation Signals (8)
Masahiro Yoshida â affiliatedwith â Sony Group Corporation
confidence 95% ¡ Affiliation: Sony Group Corporation... Masahiro Yoshida
Xiang 'Anthony' Chen â affiliatedwith â University of California, Los Angeles
confidence 95% ¡ Affiliation: University of California, Los Angeles... Xiang 'Anthony' Chen
AToM CoWriter â implements â Flower and Hayes' cognitive process theory
confidence 95% ¡ We instantiate this framework in AToM CoWriter... using an LLM-based prediction grounded in the FlowerâHayes model
AToM CoWriter â uses â Functional Agents
confidence 95% ¡ The system consists of... fourteen Functional Agents corresponding to the support types
AToM CoWriter â uses â Main Agent
confidence 95% ¡ The system consists of a Main Agent, which coordinates support selection
Main Agent â coordinates â Functional Agents
confidence 90% ¡ Main Agent, which coordinates support selection, and fourteen Functional Agents
Flower and Hayes' cognitive process theory â defines â Idea Generation
confidence 90% ¡ six cognitive processes derived from the Flower and Hayes modelâGoal Setting, Idea Generation (Generating)...
Flower and Hayes' cognitive process theory â defines â Translating
confidence 90% ¡ characterizes writing through six cognitive processes... Translating
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models can support writing, but existing tools require users to explicitly articulate prompts-particularly burdensome in creative writing, where intentions are often ambiguous. Proactive support that infers users' needs from writing interactions could alleviate this burden, but raises two challenges: determining what support to provide and when to intervene. This work focuses on the former. We hypothesize that Flower and Hayes' cognitive process theory of writing-which characterizes writing through six cognitive processes-offers an interpretable bridge between observable writing behavior and appropriate support types. Through a formative study and literature review, we identify 14 writing support types associated with these cognitive processes, along with characteristic interaction behaviors linked to each process. We then instantiate this framework in AToM CoWriter, which infers support needs from writing interactions and document context. Two within-subjects studies (N = 21) provide initial evidence that this approach improves expressiveness and idea exploration, and that cognitive process inference increases engagement with proactive suggestions. These findings suggest that cognitive processes can provide a promising basis for support selection in proactive writing systems.
Tags
Links
- Source: https://arxiv.org/abs/2608.30424v1
- Canonical: https://arxiv.org/abs/2608.30424v1
Trouble viewing inline? Open PDF directly â
Full Text
125,541 characters extracted from source content.
Expand or collapse full text
AToM CoWriter: Towards Cognitive ProcessâAware Proactive Writing Support Masahiro Yoshida Note: Both authors contributed equally to this research. Affiliation: Sony Group Corporation, Tokyo, Japan email: Masahiro.C.Yoshida@sony.com , Atsuya Kobayashi Affiliation: Sony Group Corporation, Tokyo, Japan email: Atsuya.Kobayashi@sony.com , Kei Tateno Affiliation: Sony Group Corporation, Tokyo, Japan email: Kei.Tateno@sony.com and Xiang âAnthonyâ Chen Affiliation: University of California, Los Angeles, Los Angeles, California, USA email: xac@ucla.edu Š none Figure 1. Proactive support workflow in AToM CoWriter. The system (1) continuously monitors writing interaction behaviors (keystrokes, cursor movement, editing patterns), (2) infers the userâs current cognitive process from a summary of these interaction signals together with the current document text, using an LLM-based prediction grounded in the FlowerâHayes model (Flower and Hayes, 1981), and (3) upon detecting a potential writing block, automatically activates appropriate Functional Agents to generate targeted support. The user can then access these suggestions to advance writing without explicit prompting. Abstract. Large language models can support writing, but existing tools require users to explicitly articulate promptsâparticularly burdensome in creative writing, where intentions are often ambiguous. Proactive support that infers usersâ needs from writing interactions could alleviate this burden, but raises two challenges: determining what support to provide and when to intervene. This work focuses on the former. We hypothesize that Flower and Hayesâ cognitive process theory of writingâwhich characterizes writing through six cognitive processesâoffers an interpretable bridge between observable writing behavior and appropriate support types. Through a formative study and literature review, we identify 14 writing support types associated with these cognitive processes, along with characteristic interaction behaviors linked to each process. We then instantiate this framework in AToM CoWriter, which infers support needs from writing interactions and document context. Two within-subjects studies (N = 21) provide initial evidence that this approach improves expressiveness and idea exploration, and that incorporating interaction-derived cognitive-process cues was associated with greater engagement with proactive suggestions. These findings suggest that cognitive processes can provide a promising basis for support selection in proactive writing systems. Keywords: Human-AI Collaboration, Proactive Agent, Writing Assistant, Creative Writing 1. Introduction Large language models (LLMs) are reshaping how people approach writing tasks. Chat-based LLM services such as ChatGPT11 1 ChatGPT by OpenAI, https://chat.openai.com and Claude22 2 Claude by Anthropic, https://claude.ai support all phases of writingâfrom ideation and outlining to drafting and reviewingâthrough appropriately crafted prompts. These advances highlight the potential of AI agents as collaborators that augment human writing capabilities (Lin et al., 2023). However, in these user-initiated systems, writers must pause their work to reflect on their current state and challenges, and then articulate an appropriate prompt that conveys their intent and context (Guo et al., 2025). This process involves two key burdens: the cognitive load of verbalizing problems and the physical effort of prompting (Subramonyam et al., 2024; Kuo et al., 2026). As a result, prompt-driven support, while flexible, can interfere with the natural writing flow. These burdens are particularly salient in creative writing, such as storytelling, where writersâ intentions are often ambiguous and their support needs may change dynamically throughout the writing process. In contrast, proactive writing support shifts part of this burden from the writer to the system: rather than requiring writers to specify each request explicitly, the system can infer likely support needs from naturally occurring writing interactions and offer assistance accordingly. Such inference may draw on signals including keystrokes, mouse operations, cursor movements, editing patterns, and the evolving document (Liu et al., 2026b). In the domain of proactive support, many systems have been proposed that automatically complete or continue text (Bhat et al., 2023; Jakesch et al., 2023). While highly useful, such support primarily targets the Translating process, in which writers transform ideas into linguistic form (Flower and Hayes, 1981). Writing difficulties, however, also arise while generating ideas, organizing thoughts, evaluating drafts, and revising text. Supporting these different activities requires not only detecting when assistance may be needed, but also determining what kind of assistance is appropriate at a given moment. Designing such proactive support therefore involves two distinct challenges: determining what to offer and when to offer it. This work focuses on the what problem. A central challenge is to connect observable writing behaviors with the types of support that writers may need at a given moment. We investigate cognitive processes as a theory-grounded bridge between these behaviors and appropriate support types. Specifically, we ask whether observable writing interactions can provide useful cues about the writerâs current cognitive process, and whether this process can in turn guide the selection of appropriate writing support. We distinguish this question from the when problem. Existing proactive systems commonly use heuristic triggers such as inactivity (Buschek et al., 2021; Bhat et al., 2023). Our system adopts the same heuristic: it does not adapt intervention timing to the writer, and we leave adaptive timing to future work (Sec. 6.2). Our scope is creative story writing, where writers repeatedly move among ideation, organization, translation, evaluation, and revision, providing a suitable setting for investigating process-aware support selection. To operationalize this approach, we conducted a formative study grounded in the cognitive process model of writing by Flower and Hayes (Flower and Hayes, 1981), which characterizes writing as an iterative interplay of six cognitive processes and has been widely applied in the design and analysis of writing support tools (Wan et al., 2024a; Gero et al., 2022a). Through this study and a complementary literature review, we identified fourteen types of writing support associated with six cognitive processes, together with characteristic writing interaction behaviors related to each process. These findings provide a theory-grounded mapping from observable writing interactions, through cognitive processes, to candidate support types (Sec. 3). We instantiate this framework in AToM CoWriter, a proactive writing support system for creative writing. AToM CoWriter continuously observes writing interactions such as typing pauses and speed, deletion and revision patterns, cursor location and movement, and scrolling, estimates the writerâs current cognitive process together with document context, and uses this information to select relevant forms of assistance. The system consists of a Main Agent, which coordinates support selection, and fourteen Functional Agents corresponding to the support types identified in the formative study. When support is triggeredâin the current implementation, after a brief period of typing inactivityâthe Main Agent selects up to three relevant Functional Agents, which generate suggestions in the background for users to inspect at their own pace. The name AToM (AI with Theory of Mind) is inspired by the concept of Theory of Mindâthe ability to infer othersâ intentions, goals, and emotions (Wang and Goel, 2024; Wang et al., 2021). In AToM CoWriter, we draw on this idea to infer writersâ likely cognitive processes from their writing interactions. Suggestions are surfaced through an unobtrusive notification rather than a pop-up, and writers can additionally request support at any moment through a user-initiated, prompt-less mechanism. We evaluated this approach through two complementary within-subjects studies on short story writing. Study 1 (n=11n=11) examined the overall experience of system-inferred support relative to conventional user-articulated, prompt-based support, providing initial evidence of benefits for expressiveness and idea exploration while also revealing tensions around timing and user control. Study 2 (n=10n=10) held the proactive interaction paradigm constant and examined whether adding interaction-derived cognitive process cues to document-based support selection affected engagement with proactive suggestions. Interaction results showed substantially greater engagement with proactively generated suggestions when cognitive process inference was incorporated, providing evidence that behavioral process cues can contribute to support selection. In summary, this paper makes the following contributions: (1) A theory-grounded framework for incorporating cognitive process inference into proactive writing support, connecting observable writing interactions to the six cognitive processes defined in the FlowerâHayes model and to fourteen support types identified in our study (Sec. 3). (2) The design and implementation of AToM CoWriter as an instantiation of this framework, using writing interaction behaviors and document context to infer cognitive processes and provide context-appropriate proactive support (Sec. 4). (3) Empirical evidence from two complementary user studies suggesting that system-inferred support can facilitate creative exploration, and that incorporating interaction-derived cognitive process cues can increase writersâ engagement with proactively generated suggestions (Sec. 5). Together, these contributions suggest that cognitive processes can serve as a useful intermediate representation for determining what support to provide in proactive writing systems. 2. Related Work 2.1. Proactive AI: What and When to Offer Systems that respond only to explicit requests require users to first recognize their own difficulty and then articulate it, which itself imposes cognitive and interaction costs (Subramonyam et al., 2024; Kuo et al., 2026). Proactive and mixed-initiative interaction instead shifts part of the initiative from users to the system, and has been extensively studied to reduce these costs (Horvitz, 1999; Allen et al., 1999), with a longstanding challenge being how to balance system initiative with user control (Chu-Carroll, 2000). Poorly timed or misaligned interventions can increase interaction costs or undermine usersâ sense of agency (Horvitz, 1999; Clark et al., 2018), making proactivity particularly challenging in open-ended activities such as writing. We characterize proactive support in terms of two key design questions: when assistance should be offered and what assistance should be provided. Regarding the when question, existing systems have used a range of triggering mechanisms, including real-time intervention after each input (Chang and Chang, 2015; Arnold et al., 2020; Tsai et al., 2020), activation upon inactivity (Buschek et al., 2021; Jakesch et al., 2023; Bhat et al., 2023), and threshold-based triggers on editing activity (Lin et al., 2024). More recent work on proactive agents has explored adaptive approaches to intervention timing, including personalized proactivity, modeling the costs of unnecessary assistance, and timing-aware decision policies (Kaur et al., 2026; Steyvers and Mayer, 2025; Ding et al., 2026; Liu et al., 2025). The what question concerns selecting an appropriate form of assistance once support is considered useful. Many proactive systems are designed around a particular form of assistance (Buschek et al., 2021; Tsai et al., 2020), while some systems adaptively select among multiple support types. For example, Lin et al. used a Multi-Armed Bandit algorithm to select among three support types (Lin et al., 2024), and Chen et al. proposed a proactive programming assistant that uses an LLM to predict which of eight support categories is required (Chen et al., 2025). Our work focuses primarily on this what-to-offer problem. Support selection is particularly relevant to writing because writers iteratively move among qualitatively different cognitive processes, and the forms of assistance that are useful can differ across these processes. The following subsection reviews this process-oriented view of writing and prior work relating writing processes to support needs. For the when question, we adopt a pragmatic combination of inactivity-based triggering and a user-initiated mechanism, so that the selection of support can be investigated without simultaneously introducing a new timing mechanism. 2.2. Cognitive Processes and Writing Support Writing is not a single, linear activity but an iterative process comprising qualitatively different forms of cognitive work. The Cognitive Process Theory of Writing by Flower and Hayes (Flower and Hayes, 1981) decomposes writing into six cognitive processes: Generating, Organizing, and Goal Setting, which together constitute Planning; Translating; and Evaluating and Revising, which together constitute Reviewing. Later models have extended this account to accommodate additional factors such as motivation, working memory, and the social context of writing (Hayes, 2012; Becker, 2006). The process-oriented perspective has likewise been employed to analyze and design writing support systems (Gero et al., 2022a; Beauvais et al., 2011; Wan et al., 2024a). Prior work suggests that different stages or processes of writing call for different forms of assistance. Chakrabarty et al. (Chakrabarty et al., 2024) analyzed LLM-assisted writing through the FlowerâHayes model and found that the value of LLM support differed across cognitive processes. GĂśldi et al. (GĂśldi et al., 2024) similarly demonstrated the importance of aligning intelligent writing support with relevant cognitive processes. More broadly, design spaces of intelligent writing assistants have organized prior systems according to the stages of writing that they support (Lee et al., 2024; Gero et al., 2022a), and recent work has investigated how specific writing tools correspond to writersâ needs and writing activities (Zhao et al., 2025). Together, these studies suggest that writersâ support needs vary with the processes in which they are engaged. However, as noted above, existing systems generally require writers to recognize and articulate those needs themselves. Our work therefore uses cognitive processes as a theory-grounded intermediate layer between observable writing behavior and support selection: rather than treating process identification as an end in itself, we investigate whether it can help determine what support a proactive writing system should provide. 2.3. Inferring User States from Interaction Traces Observable interaction traces have long been used as signals for modeling usersâ activities, abilities, and task states (Gajos and Weld, 2004; Cockburn et al., 2007; Liu et al., 2026b). In writing, such traces provide particularly rich signals for understanding and modeling writersâ ongoing cognitive processes. A substantial literature has analyzed keystroke dynamics and temporal patterns in writing, associating pauses, revision behavior, cursor activity, and other interaction measures with different writing processes and cognitive demands (Baaijen et al., 2012; Guo et al., 2018; Zhang et al., 2025a; Delgado-Osorio et al., 2025). Some studies have explicitly attempted to infer phases or cognitive processes of writing from keystroke logs, textual context, or combinations of behavioral features (Conijn et al., 2024; Wang et al., 2025). Mouse and cursor behavior have likewise been studied as signals of usersâ cognitive activity and workload (Arshad et al., 2015; Zhang et al., 2024). Related approaches have recently appeared in AI-assisted programming, where systems observe interactions with code editors to infer whether and how assistance may be useful. For example, proactive programming assistants have used editing activity and task context to determine potential assistance (Pu et al., 2025; Mozannar et al., 2024). These studies demonstrate how ongoing interaction can provide signals for inferring task-relevant user states and needs without relying solely on explicit commands. Our work builds on this tradition of interaction-based user modeling in the domain of writing. Rather than directly mapping low-level behavioral signals to a particular intervention, we use interaction traces to estimate the writerâs current cognitive process. The inferred process then structures the selection among multiple support types, connecting interaction-based user modeling with a theory of writing and providing an interpretable basis for proactive support selection. 2.4. Interaction Paradigms for AI Writing Support AI-assisted writing systems have explored a broad range of interaction paradigms for supporting writers. Existing systems provide functions such as text continuation, rewriting, critique, summarization, and idea generation through inline suggestions, dedicated interfaces, or conversational interaction (Bhat et al., 2023; Jakesch et al., 2023; Lee et al., 2024; Chung et al., 2022; Mirowski et al., 2023; Bao et al., 2022; Talaei et al., 2025; Zhang et al., 2025c; Yao et al., 2026; Dang et al., 2022; Laban et al., 2024). While chat-based LLMs provide flexible access to many such capabilities, obtaining support through them depends on the writerâs own articulation of the problem, with the costs described above. Recent research has therefore explored interaction paradigms that go beyond conventional prompt-based chat. For example, Texterial treats text as directly manipulable material for LLM-mediated writing (J Shen et al., 2026), while Visual Story-Writing supports story development through manipulation of visual representations (Masson et al., 2025). Polymind uses parallel visual diagramming and microtasks to support prewriting (Wan et al., 2025), and Narrix supports story writing by allowing writers to remix narrative strategies from examples (Zhang et al., 2026). Other work has examined how writers integrate LLMs into prewriting and creative practice (Wan et al., 2024b; Guo et al., 2025), illustrating that useful AI assistance can take substantially different forms across the writing process. Beyond the utility of the support itself, interaction design also shapes writersâ sense of agency and ownership, and prior studies have documented tensions between AI assistance, authorial control, and perceived ownership (Carrera et al., 2026; Draxler et al., 2024; Biermann et al., 2022; Yoshida et al., 2026). Together, these studies illustrate substantial diversity in both the forms of writing support and the ways in which writers interact with AI. Our work is complementary to efforts that develop new interaction techniques for co-writing. Rather than proposing a single interaction paradigm intended to accommodate all writing activities, we focus on an upstream problem: determining what support is likely to be relevant to the writerâs current process without requiring the writer to articulate that need explicitly. 3. Formative Study This formative study aimed to deepen our understanding of the relationships among writersâ interaction behaviors, momentary cognitive processes, and the support they require at each moment, as a foundation for designing proactive writing support. In this study, we use the term cognitive process to refer to the momentary cognitive process that is most prominent at a given moment during writing, as characterized in the Flower and Hayes model (Flower and Hayes, 1981). Unless otherwise noted, all mentions of cognitive processes in the remainder of the paper follow this definition. To structure this investigation, we adopted the Cognitive Process Model of Writing proposed by Flower and Hayes (Flower and Hayes, 1981), which provides a well-established framework linking observable writing interaction behaviors to the underlying cognitive processes. To collect empirical data, we developed a custom data-logging editor grounded in this framework. Participants used the tool to write short stories, during which we recorded both detailed writing interaction logs and self-reported cognitive processes whenever they felt âblocked.â Building on these foundations, this section addresses the following three research questions (RQs): RQ-F1.: What relationships exist between usersâ cognitive processes and the types of support they require? (Sec. 3.4) RQ-F2.: How are usersâ observable writing interaction behaviors related to their cognitive processes during writing blocks? (Sec. 3.5) RQ-F3.: Can usersâ cognitive processes be predicted from their behavioral data? (Sec. 3.6) 3.1. Data Logging Editor Figure 2. The data-logging editor used in the formative study. It collects cognitive process annotations via voice input to minimize disruption during writing. The developed data-logging editor was a lightweight rich-text editor augmented with logging and annotation capabilities for later analysis (Fig. 2). It recorded two types of data: (1) the Writing Interaction Log, which captures usersâ interactions with the editor during writing; and (2) the Self-Report Log, which records usersâ reflections when they encounter a writing block. The Writing Interaction Log automatically tracks various writing-related activities, including keystrokes, mouse operations, cursor positions, and text content. The Self-Report Log allows users to annotate their current process and thoughts whenever they feel blocked or in need of support. In other words, users primarily continue writing as usual but are encouraged to provide annotations at moments when they experience a block in their writing process. 3.1.1. Self-Report Procedure The Self-Report Log was collected through an interactive floating panel called the âInterviewerâ, which guided participants in recording their thoughts during writing (Fig. 3). Participants were instructed to write as they normally would and to engage with the Interviewer only when they felt blocked or needed AI support. Figure 3. Self-reporting flow: The Interviewer prompts after 10 seconds of inactivity. Users select a cognitive process and verbally describe their difficulty and desired support. The Interviewer displayed the message âIs there anything youâre stuck or having trouble with?â after 10 seconds of inactivity. When participants hovered the mouse over the panel, six cognitive processes derived from the Flower and Hayes modelâGoal Setting, Idea Generation (Generating), Organizing Ideas, Translating, Evaluating, and Revisingâappeared as selectable options. Participants could also manually initiate an annotation session at any time when they felt blocked, as the six cognitive process options were always accessible regardless of whether the question prompt was currently displayed. After selecting a cognitive process, the Interviewer entered a listening mode. Participants then responded verbally to the displayed question, âIs there anything youâre stuck on or having trouble with?â Once the participant had stopped speaking for more than one second or pressed the âEnd Speechâ button, the system proceeded to the second question, âIs there anything youâd like the agent to help you with?â, to which the participant also responded verbally. Each self-report session thus recorded three elements: (1) The current cognitive process (selected from six options) (2) The content of the difficulty (spoken response) (3) The desired support (spoken response) Participants repeated this process throughout the writing session whenever they encountered a block. To encourage consistent self-reporting, the system played a notification sound after 10 seconds of inactivity. Overly frequent prompts distracted participants, while infrequent ones led them to forget to self-report. Based on qualitative feedback from four pilot participants, we determined that a 10-second interval provided the best balance. The notification served as a gentle reminder for self-reflection; users could ignore it if they were not actually blocked. The Interviewer was carefully designed to minimize cognitive and physical burden. Excessive annotation effort could distort natural writing behavior and reduce data validity. The current process selection was thus kept simple, and verbal responses were used instead of typing to reduce interruption. While a Think-Aloud approach was also a possible option, it would have been too demanding and intrusive. Our design instead balances data richness with minimal cognitive and physical load. 3.2. Experiment Setup Sixteen engineers from a manufacturing company participated in the study (13 men and 3 women, aged 20â39; nine in their 20s and seven in their 30s). Recruitment was conducted through several internal Slack channels. All participants were regular LLM users. Three had written a story once before, and four had experience writing several creative works as a hobby. We recruited participants with engineering backgrounds because the study required individuals familiar with LLM-based tools who could engage fluently with the data-logging interface, allowing us to focus on capturing behavioral patterns and support needs rather than tool-onboarding effects. The creative writing task itself was novel for most participants, ensuring that observed difficulties reflected the creative writing challenge rather than prior writing expertise. Each session began with a 10-minute tutorial explaining tool usage and the six cognitive processes based on Flower and Hayesâ definitions. Participants were instructed to use only the provided editorâno external tools were allowed. They were also instructed to note ideas and outlines within the same editor when needed. They were given 40 minutes to write a short story with their own theme in Japanese, followed by a 15-minute semi-structured interview. All examples shown in figures were translated into English for this paper. 3.3. Behavior During Writing Sessions All participants began by jotting down notesâsuch as themes, settings, or character profilesâbefore starting the main text. Although the boundary between notes and main text was not explicitly defined, they were easily distinguishable in the final documents. The average main text length (excluding notes) was 581.3 Japanese characters. Across sessions, 152 self-reports were recorded (mean = 9.5 per participant, SD = 5.39). The distribution of self-reported cognitive processes at blocked moments per user is shown in Fig. 4. The Generating process (idea generation) was reported most frequently. Participants often selected this process repeatedly in the early stage of writing and again later when developing new story ideas, indicating its frequent use throughout the session. Additionally, Table 1 presents the median pause duration preceding each self-reported block, across reported cognitive processes. Figure 4. Distribution of average annotation count per participant. Interviews revealed that articulating oneâs difficulties and support needs was sometimes challenging. For instance, P8 stated, âWhen Iâm struggling to organize my ideas and the notification rings and starts annotation, I sometimes lose track of what I was thinking.â This suggests that explicitly describing oneâs needsâsimilar to prompting an LLMâsometimes imposes a heavy cognitive load, underscoring the importance of proactive systems that can infer user needs implicitly. Table 1. Median Pause Duration before reporting Block by Cognitive Process Process Median Pause Duration (sec) Goal-Setting 76.04 Generating 26.87 Organizing 29.94 Translating 15.87 Evaluating 12.33 Revising 18.90 3.4. Relationship Between Cognitive Processes and Support Needs To clarify the relationship between cognitive processes and required support types, we conducted qualitative coding of participantsâ verbal self-reports (see Appx. A.1 for the codebook). The coding categories were developed from the self-report data rather than defined a priori. One researcher first coded the transcribed self-reports, and a second researcher subsequently reviewed the assigned codes to identify potential inconsistencies or questionable interpretations. This analysis allowed us to identify and organize usersâ support requirements corresponding to each cognitive process. In parallel, we performed a literature review to understand what kinds of writing support functions have been proposed in prior research. Specifically, we began with papers included in Lee et al.âs survey of writing support systems (Lee et al., 2024), focusing on those tagged with their defined Writing Stage categories and published since 2018. For each category, we collected up to the top 10 most-cited papers as of January 22, 2025 (fewer when fewer papers were available). In addition, we reviewed papers from UIST 2023, and IUI, UIST, and CHI 2024, retrieved by searching titles and author-specified keywords containing âwriteâ or âeditâ and their derivatives (e.g., writing, writer, written). From these, we filtered papers related to AI-based writing assistance, defined as those whose title, keywords, or abstract included AI (Artificial Intelligence), LLM (Large Language Model), or Machine Learning, resulting in a total of 28 reviewed papers (22 from the work of Lee et al. (Lee et al., 2024)). Through this process, we defined 13 distinct support types. We then integrated the findings from both analysesâself-report data and the literature reviewâto map cognitive processes, user requirements, and support types, as summarized in Table 2. Because no support types corresponding to the Goal Setting process were identified in our initial literature review, we defined one based on observed requirements in our dataset. This resulted in a total of 14 support types. Among them, three typesâSynopsis Generation, Scoring, and Summarizationâwere reported in prior studies but not explicitly mentioned in our self-reports. We therefore inferred their most plausible corresponding cognitive processes based on observed patterns and contextual relevance. After constructing this taxonomy, we conducted a supplementary review of the venues that had appeared since the initial reviewâCHI 2025 and 2026, UIST 2025, and IUI 2025 and 2026âusing the same search and filtering criteria. All identified papers fell within the existing 14 support types and are shown in parentheses in the Related Work column of Table 2. Table 2 also includes the Implemented Support Functions integrated into AToM CoWriter. While all 14 support types correspond to distinct requirements, their complexity varies. For example, some users expressed vague intentions, such as wanting general ideas related to a theme (Support Type: Idea Generation), whereas others had specific intentions, such as seeking ideas for a particular scene or element (Support Type: Specified Idea Generation). Depending on this complexity, some supports can be provided directly, while others require a simple interaction with the user to narrow down the intended direction (see Sec. 4 for details). Table 2. Summary of the identified support types and their corresponding cognitive processes, synthesized from a formative study and a literature review, with representative user requirements and implemented functions in AToM CoWriter. âRelated Workâ reflects prior literature, while âRequirementsâ and âExample Quoteâ are derived from the formative study. Cells marked with â-â indicate no corresponding evidence from that source, while evidence is present in the other. In the âRelated Workâ column, citations enclosed in an additional pair of parentheses, e.g., â([N])â, were identified in an additional review of papers published after the initial literature review (CHI 2025/2026, UIST 2025, and IUI 2025/2026) Cognitive Process Support Type Related Work Requirements Example Quote Support Function Description Implemented in AToM CoWriter Goal Setting Target Setting ((El Alaoui et al., 2025; Yao et al., 2026; Liu et al., 2026a)) Uncertain about what aspects to decide first when initiating a story-writing process. P4: âIâd like to see a list of the typical elements people set when creating a story.â âIâd like a kind of template showing how to organize ideas that come up midway, or how to build the basic structure of a story.â Assist in deciding writing goals, style, and target audience through lightweight interaction with the AI. Generate Idea Generation (Gero et al., 2022b; Clark et al., 2018) ((J Shen et al., 2026; Siddiqui et al., 2025)) Seeks broad, open-ended ideas when the user has not yet clarified what they are looking for. P13: âIâd like you to come up with a few ideasâlike âHow about a story like this?â or âWouldnât this genre be interesting?ââ P15: âPlease give me a few fun ideas.â P8: âIâd like you to think about the ending of this story.â Generate a list of concrete story ideas that align with the current document context. Generate Specified Idea Generation (Schmitt and Buschek, 2021; Mirowski et al., 2023) Requests ideas focusing on specific aspects based on their idea. P1: âPlease give me several ideas about how AI could help with housework.â Confirm the desired direction of ideas through lightweight interaction with the AI (Figure 6 IV-i), then generate a tailored list of idea candidates. Generate Ideation Expression Generation (Buschek et al., 2021; Liu et al., 2019) Wants to explore alternative phrasings or stylistic variations for existing text. P7: âI think the opening of the story is a bit weak, so Iâd like to change it a bit. Please think of a better way to express it.â After confirming the userâs intent through lightweight interaction (Figure 6 IV-i), generate alternative phrasing or stylistic variations of the selected text. Generate Synopsis Generation (Osone et al., 2021) ((TĂźtĂźncĂź et al., 2026)) - - Generate a concise story synopsis by considering notes and contextual information in the current document. Organize Plot Generation (Goldfarb-Tarrant et al., 2019; Singh et al., 2023; Mirowski et al., 2023) ((Qin et al., 2025; Talaei et al., 2025; TĂźtĂźncĂź et al., 2026; Hu et al., 2026)) Needs assistance in combining ideas into a coherent storyline or plot. P8: âIâd like suggestions on how to combine the ideas I have and what flow would make a good story.â P11: âIâve gathered fragments that could be used in a story, but Iâm thinking about how to blend them together.â Organize ideas and notes into a coherent storyline, generating a bullet-point outline that reflects the narrative flow. Translate Next Sentence Generation (Yuan et al., 2022; Clark et al., 2018; Jakesch et al., 2023; Roemmele and Gordon, 2018; Dang et al., 2023; Bhat et al., 2023; Lee et al., 2022) Requests suggestions for the next sentence that align with the existing context and tone. P4: âSince Iâve written the setup, please just write the opening for me.â Generate a list of next-sentence candidates that maintain tone and consistency with the current document context. Translate Specified Next Sentence Generation (Dang et al., 2023; Yuan et al., 2022) Requests context-aware next-sentence suggestions that follow a specific narrative direction. P8: âPlease add a passage describing the part where the man realizes he has been reincarnated.â Confirm the desired narrative direction through lightweight interaction with the AI (Figure 6 IV-i), then generate next-sentence candidates that reflect it. Evaluate Commenting (Peng et al., 2020; Afrin et al., 2021; Wambsganss et al., 2022; MacLean et al., 1991) ((Hu et al., 2026; Siddiqui et al., 2025; Liu et al., 2026a; Qin et al., 2025; Zhang et al., 2025b)) Seeks feedback and critique on the written text to improve readability and engagement. P8: âPlease review the overall text so it becomes something that makes readers want to keep reading.â Provide structured feedback on the document, including identified strengths and suggested areas for improvement. Evaluate Scoring (Peng et al., 2020; Wambsganss et al., 2022; Afrin and Litman, 2023) - - Return an overall quality score (0â100) accompanied by a brief rationale. Evaluate Summarization (Dang et al., 2022) - - Generate concise paragraph-level summaries to help the user review the document efficiently. Revise Error Auditing (Laban et al., 2024; Iqbal et al., 2018) ((Hu et al., 2026)) Requests identification and correction of grammatical or typographical errors. P7: âPlease tidy up the text, fixing typos and awkward expressions.â P5: âPlease point out any mistakes in the text and correct them if possible.â Highlight grammatical and logical issues in the document and suggest corrections. Revise Refining (Wu et al., 2019; Iqbal et al., 2018; Goldfarb-Tarrant et al., 2019; Padmakumar and He, 2022) ((J Shen et al., 2026; Zhang et al., 2025c; Zhou et al., 2025; Hu et al., 2026)) Seeks improvements to enhance fluency, coherence, and overall writing quality. P2: âIâd like the text Iâm writing to be reconstructed more smoothly, but I donât have any ideas.â Generate revised versions of the selected text that maintain meaning while improving clarity and readability. Revise Style Transfer (Yuan et al., 2022; Laban et al., 2024) ((Siddiqui et al., 2025; Yao et al., 2026)) Requests stylistic adjustments such as tense consistency or tone unification. P5: âPlease consider whether it should be unified in past tense, and revise it if necessary.â Confirm the desired stylistic direction through lightweight interaction with the AI (Figure 6 IV-i), then generate a rewritten version that follows the requested tone or style. 3.5. Relationship Between Writing Interaction Behavior and Cognitive Processes We then analyzed writing interaction logs to examine behavioral patterns characteristic of the moments when each cognitive process was self-reported. The results of this analysis, organized by cognitive process and evaluated from multiple perspectives, are summarized in Table 3. Table 3. Writing interaction metrics across cognitive processes at the moment users initiated self-reports. Elapsed time (min) indicates the average time from session start to self-report. Chars in Main and Chars in Memo represent the average number of characters written in the main text and memo areas, respectively. Complete Sentence Ratio (%) denotes the percentage of cases where the final sentence was complete. Cursor PositionâMain/Memo (%) shows the proportion of time the cursor was located in the main text area. Writing Speed refers to the average change in character count per minute. Cursor Move Distance is the total number of characters the cursor moved within the last two minutes. Scroll Distance indicates the total number of pixels scrolled within the last two minutes. Delete Count is the total number of characters deleted within the last two minutes. Metric Goal Setting Generating Organizing Translating Evaluating Revising Elapsed time (min) 7.90 17.20 15.30 23.50 28.90 36.30 Chars in Main 14.06 155.58 187.87 305.42 678.90 805.09 Chars in Memo 175.81 306.84 289.40 493.53 357.30 528.91 Complete Sentence Ratio (%) 50.00 69.14 73.33 52.63 90.00 100.00 Cursor Position-Main/Memo (%) 28.57 23.73 30.00 25.00 88.89 80.00 Writing Speed (characters/min) 23.81 21.22 4.74 15.17 15.21 13.41 Cursor Move Dist. (characters) 135.63 357.28 215.93 293.11 526.80 802.82 Scroll Dist. (pixels) 15.88 59.32 91.07 104.95 151.60 240.91 Delete Count 16.63 41.27 41.27 46.89 66.40 35.45 Temporal order. Regarding the timing of each cognitive process, Goal Setting was self-reported earliest, followed by Generating and Organizing. Subsequently, Translating, Evaluating, and Revising appeared in this order on average. Text balance. During Planning (Goal Setting, Generating, Organizing) and Translating, note length exceeded main text length, while during Reviewing (Evaluating, Revising), the reverse was true. Sentence completeness and cursor position. In Revising, 100% of documents ended with complete sentences; Evaluating showed a similar trend, while Translating had lower completeness. Interaction patterns. Writing speed was highest during Goal Setting and Generating but dropped sharply during Organizing. Cursor movement and scrolling distances increased progressively from Planning to Reviewing. Delete events were most frequent during Evaluating. These findings suggest that observable behaviors can serve as useful indicators of cognitive processes, motivating the prediction approach described next. 3.6. Cognitive Process Prediction Given the consistent behavioral characteristics observed above, we explored whether cognitive processes could be predicted from the log and document data collected at the moments of usersâ self-reports. As a feasibility demonstration, we evaluated this using the 152 self-reported data points collected from the 16 participants, each corresponding to a writing block event annotated by participants. While this dataset is relatively small for supervised learning, collecting large-scale labeled data for cognitive process inference is costly, motivating approaches that operate effectively under limited data. Accordingly, we adopt an LLM-based approach that operates in a near zero-shot manner, leveraging commonsense reasoning about writing processes rather than requiring large-scale supervised training. The LLM-based predictor (Claude 3.5 Sonnet) received three types of inputs: (a) the document content at the time of reporting, (b) descriptions of the behavioral tendencies for each cognitive process derived from Sec. 3.5, and (c) a summary of the recent writing interaction log. To construct this summary, continuous log features were discretized into qualitative categories (e.g., âwriting speed is relatively highâ) using percentile thresholds and converted into natural-language descriptions suitable for inclusion in the prompt (see Appx. A.2 and A.4 for the prompt and feature details). We examined the feasibility of this prediction approach through two exploratory comparisons. First, we compared LLM (log + document), our approach, with LightGBM (Ke et al., 2017) trained on log features alone (evaluated with cross-validation) to test whether a lightweight, lower-latency model could serve as a practical alternative. The LLM outperformed LightGBM across all metrics, suggesting that semantic understanding of document content contributes to cognitive process prediction beyond what behavioral features alone can capture. Second, an ablation comparing LLM (log + document) with LLM (document only) yielded higher prediction scores when behavioral log features were included, particularly when disambiguating cases where document content alone is ambiguous (e.g., distinguishing a writer preparing to continue from one reviewing prior text). We evaluated prediction accuracy, macro-F1, and weighted-F1 for both the six-process classification and a coarser three-process mapping (Planning, Translating, Reviewing); the LLM (log + document) achieved the highest performance across all metrics (full results in Appx. A.3, Tables 4 and 5). 3.7. Summary of Findings RQ-F1: Relationship between cognitive processes and required support. We identified relationships between writersâ cognitive processes during writing blocks and the types of support they required (Table 2). Each cognitive process was associated with multiple support types, resulting in a taxonomy of 14 types (Sec. 3.4). RQ-F2: Relationship between writing interaction behaviors and cognitive processes. Behavioral analyses revealed distinctive activity patterns for each cognitive process (Table 3), such as differences in writing speed, cursor movement, and editing operations. These findings suggest that observable behaviors can serve as useful indicators of cognitive processes (Sec. 3.5). RQ-F3: Predicting cognitive processes. Prediction experiments showed that an LLM supplied with document content and behavioral descriptions could recover self-reported cognitive-process labels to a useful extent. (Sec. 3.6). Including a summary of the recent interaction log produced higher prediction scores than using document content alone. Together, these insights establish a foundation for designing proactive writing support systems that can anticipate user needs based on both cognitive process and behavioral cues. 4. System Design and Implementation The formative study mapped writing interactions, through cognitive processes, to support types (Table 2); we instantiate this mapping in AToM CoWriter, a proactive writing support system that infers the writerâs momentary cognitive process from writing interaction traces and document context, and uses it as an intermediate representation for selecting which forms of assistance to offer. Consistent with the scope of this work, the system addresses the what problem of proactive support, while deliberately keeping the proactive when mechanism simple: we adopt inactivity, a well-established trigger in prior proactive writing systems (Buschek et al., 2021; Bhat et al., 2023). Keeping the when mechanism simple helps isolate the effect of what the system offers from that of novel timing behavior. 4.1. Design Goal As discussed in Sec. 2, existing prompt-based writing tools place the initiative burden entirely on users (Subramonyam et al., 2024), while current proactive systems address only a narrow slice of writersâ support needs (Buschek et al., 2021; Tsai et al., 2020). Furthermore, while seeking creative writing support, users value maintaining a sense of ownership in both the process and outcomes (Draxler et al., 2024; Guo et al., 2025). These observations motivated two design goals: DG1 Reduce the need for explicit instructions to the agent by inferring and addressing likely challenges throughout the writing process. DG2 Stimulate and expand usersâ creativity and expressiveness without compromising their sense of ownership in the process and outcomes. 4.2. Design and Implementation Figure 5. AToM CoWriter interface. (i) Chat input to the Main Agent. (i) Functional Agent question UI for intent clarification. (i) Panel Output displaying generated content. (iv) Suggest Output providing inline editing suggestions in the editor. Building on the design rationale and prediction approach above, we instantiated this framework in AToM CoWriter, a proactive writing support system designed to examine whether cognitive process inference can meaningfully guide support selection in practice (Fig. 1). The UI combines a rich text editor with a chat interface and a dedicated area for reviewing agent outputs (Fig. 5). The system employs a multi-agent architecture: a Main Agent serves as the orchestrator, dynamically assigning up to three Functional Agents in parallel from the fourteen support types in Table 2. All inference and generation run entirely in the background to minimize interruption to the writerâs typing flow, supporting the elimination of explicit prompting (DG1). Moreover, the combination of unobtrusive notifications and parallel suggestions is intended to let users retain control over engagement while exploring diverse ideas, aiming to support creativity without compromising ownership (DG2). While cognitive process prediction is inherently imperfect, the parallel suggestion design helps mitigate this limitation by increasing the likelihood that at least one of the generated suggestions aligns with the userâs needs. Agent outputs include Panel Output displayed in a document pane and Suggest Output providing inline suggestions within the editor (Fig. 6). The system provides three AI support scenarios: Figure 6. (I): Main Agent and tabs UI. (I): Main Agent activates appropriate Functional Agents (up to three simultaneously). (I): Tabs display a spinner during generation; upon completion, an unread badge appears. (IV): Some agents generate final outputs directly (IV-i), while others create Question UIs to confirm intent (IV-i). Proactive. In the Proactive scenario, the system detects a potential writing block when typing activity remains idle for a predefined period (Fig. 1). The Main Agent then automatically infers support needs based on the inferred cognitive processes and interaction context, activating appropriate Functional Agents in parallel to generate support content in the background. As a design choice intended to minimize disruption, we implemented an unobtrusive notification system that only displays dot indicators. Users can access the system at their own pace without pop-ups or forced interruptions, minimizing the impact of false positives on trigger events. Prompt-less. Because the inactivity-based trigger cannot capture all moments when writers desire support, we introduce a complementary Prompt-less scenario that gives users direct control over when to receive assistance: users manually trigger support by pressing a âHelp buttonâ (Fig. 6-I), while the system still automatically predicts what support to provide and selects Functional Agents. Together, the Proactive and Prompt-less scenarios form a dual mechanismâsystem-initiated and user-initiated timing, respectivelyâintended to compensate for the limitations of either approach alone. Both scenarios can be enabled simultaneously; when either triggers generation, both are temporarily disabled until completion. Prompt-based. Users directly provide instructions to the Main Agent through the chat UI, specifying both timing and type of assistance. The Main Agent may invoke Functional Agents as needed. For online user testing, we implemented AToM CoWriter as a web application integrating a rich-text editor using Lexical33 3 lexical v0.30 https://lexical.dev/, and developed multi-agent workflows powered by Anthropic Claude Sonnet 444 4 claude-sonnet-4-20250514-v1:0 using ai-sdk55 5 ai-sdk v5 https://ai-sdk.dev/. The development process involved multiple iterations of prototyping and pilot studies with three participants, including two authors. Prompting Strategy. The Main Agentâs prompt includes labels for each Functional Agentâs corresponding cognitive process, guiding agent selection (Appx. C.5, C.6). The specific support type for each Functional Agent was derived from the categories in Table 2. The prompts for the 14 Functional Agents were iteratively refined through pilot studies (Appx. C.7). Context Management. All agents share the complete message history. The system programmatically forces editor read at the beginning of every interaction turn, injecting the current editor content directly into the context window. When Functional Agents require additional context, they use ask-* tools (Appx. B.1) to generate interactive UI components (e.g., select buttons, sliders, text inputs) prompting users for clarification. Cognitive Process Prediction Integration. To integrate the LLM-based cognitive process prediction (Sec. 3.6) into the application, we developed a prediction API backend using Pythonâs FastAPI. The application performs cognitive process predictions every 30 seconds. When a prediction is made, the system automatically sends a templated message to the Main Agent in the background (not visible to the user), which includes the predicted cognitive process. The Main Agent then uses this information to select and activate appropriate Functional Agents. 5. User Evaluation 5.1. Experiment Setup We conducted two within-subjects user studies as an exploratory evaluation of AToM CoWriter, aiming to probe the validity and limitations of the underlying design concept. Study 1 (n=11n=11) examined the impact of system-inferred proactive support compared to user-articulated prompt-based support. Study 2 (n=10n=10) isolated the effect of incorporating cognitive process inference from interaction logs within the proactive paradigm. Twenty-one participants (engineers at a technology company, all frequent LLM users; IRB approved) were randomly assigned to Study 1 or Study 2. We recruited participants with engineering backgrounds because the study required individuals highly familiar with LLM-based tools and capable of quickly adapting to a novel AI-assisted writing interface. Although the writing task (short story writing) was unfamiliar to most participants, this ensured that observed difficulties reflected the creative writing challenge rather than unfamiliarity with AI tools. Fourteen participants had also participated in the formative study; however, the two studies involved fundamentally different tasks and interfaces. The formative study used a self-report annotation tool (Fig. 2) with no AI assistance, whereas AToM CoWriter is a fully proactive AI co-writing system, minimizing direct transfer of learned strategies. Nevertheless, prior exposure to the cognitive process framework may have heightened metacognitive awareness, which we discuss as a limitation in Sec. 5.4. Participants completed two 20â25 minute short story writing tasks on assigned themes, with conditions counterbalanced via Latin square. Each session lasted approximately 90 minutes, including a tutorial, two writing tasks with post-task questionnaires, and a 10â15 minute semi-structured interview according to the guide (Appx. C.2). Study 1 Conditions: System-Inferred vs. User-Articulated Support. The System-Inferred condition enabled both the automatic (inactivity-triggered) and prompt-less (button-triggered) support scenarios, where the system infers what support to provide without requiring the user to articulate their needs. The User-Articulated condition provided only the prompt-based chat interface, requiring users to formulate explicit requests. Both conditions also allowed free use of the chat interface. We note that this comparison intentionally bundles proactivity (system-initiated timing) with prompt-free interaction (system-inferred support type), because our primary design goal (DG1) is to eliminate the need for users to articulate support needs. The inactivity threshold for proactive triggering was set to 20 seconds based on pilot testing. Study 2 Conditions: Effect of Cognitive Process Inference. Both conditions employed the same interaction paradigm (proactive + prompt-less + chat), thereby holding the timing and interaction modality constant. The w/ Cog condition incorporated real-time cognitive process inference from writing interaction logs (e.g., writing speed, cursor movement, editing patterns) alongside document content to select appropriate Functional Agents. The w/o Cog condition used only the evolving document content for inference, without behavioral log features. This design complements Study 1 by isolating the contribution of cognitive process inference from the proactivity factor: whereas Study 1 evaluated the bundled experience, Study 2 controls for proactivity and tests whether behavioral logâbased inference improves the what-to-offer decision. Evaluation Metrics. Evaluation metrics included weighted Creativity Support Index (CSI) (Cherry and Latulipe, 2014), weighted NASA-TLX (Hart and Staveland, 1988; Haga and Mizukami, 1996), custom evaluation items (Ability, Control, Ownership, Deprivation, see questionnaire items in Appx. C.3), and for Study 2, interaction logâbased conversion metrics measuring how proactively generated outputs influenced user actions. 5.2. Results: Impact of Proactive and Prompt-free Support (Study 1) Figure 7. Results for all subscales of survey metrics in Study 1. Each raw score is multiplied by 10, and Performance in NASA TLX and Deprivation in Custom Evaluation Items are inversely scaled as 100âx100-x. Left of each pair: System-Inferred (green); right: User-Articulated (orange). 5.2.1. Quantitative Results The CSI results showed a significant improvement in Expressiveness in the System-Inferred condition (Holm-adjusted p=0.034p=0.034, RBC =0.924=0.924), with Total Score and Results Worth Effort also showing large effect sizes (RBC >0.58>0.58, CLES >0.64>0.64). The NASA-TLX showed no significant differences; total workload was essentially equivalent. Custom evaluation items showed no significant differences, with marginally greater efficiency in the System-Inferred condition. (see Fig. 7 and Table 7 in Appx. C.4 for full statistics) 5.2.2. Qualitative Findings We analyzed semi-structured interview transcripts using a thematic analysis approach. Two researchers independently reviewed the transcripts and developed initial codes. An LLM (Claude 3.5 Sonnet) was used as a supplementary tool to propose candidate codes from the transcripts; these proposals were then reviewed, revised, and finalized by the researchers through discussion. The resulting codes were organized into themes aligned with the research questions. We report the qualitative findings below organized by theme rather than by individual codes, as our goal was to surface design-relevant insights rather than to produce an exhaustive codebook. Participants described two main advantages of system-inferred support: it helped them continue writing during blocks and exposed them to more diverse ideas. Notably, participantsâ comments emphasize the diversity of generated ideas rather than mere convenience of not writing prompts, suggesting that the expressiveness gain stems from the system simultaneously activating multiple Functional Agents that span different cognitive processes, thereby expanding the idea space beyond what writers would explore on their own. âWhile I was thinking, the AI generated 5 or 6 different ideas, so I could quickly pick a good one. It saved time and reduced stress.â (P12) âMultiple suggestions let me choose what I neededâvery helpful.â (P05) At the same time, participants also reported challenges. The main challenges concerned timing mismatches and a perceived loss of control. Some suggestions were contextually irrelevant or increased verification costs. âI realized I donât really need proactive support. I want to focus on my work, and the thought process itself has value.â (P10) âWhen it acts on its own, it feels meddlesome. I want to be left alone, but itâs helpful that itâs thinking in the background. Itâs a very selfish feeling.â (P15) Regarding writing process and ownership, NASA-TLX showed no significant workload differences. Proactive support encouraged exploration but sometimes disrupted focus, highlighting a trade-off. Custom evaluation items likewise showed no significant decline in perceived ownership. Many participants retained a sense of authorship, as they filtered and edited AI suggestions. âIn terms of concentration, the former (Proactive) was lower, but the degree of exploration was higher.â (P15) âI havenât fixed most of it, but I can still tell itâs my work. Itâs me who accepts, chooses, and fixes the weird parts.â (P03) âI had the final say on the ideas and could steer the AI in the direction I wanted, so my sense of ownership was not compromised.â (P21) Overall, these results provide support for DG2: the system helped expand usersâ expressiveness and idea exploration without clearly undermining their sense of ownership. Participants often described the AI as broadening the space of possible ideas, while still emphasizing that they themselves made the final decisions about what to adopt, revise, or discard. At the same time, the reported tensions around control and timing suggest that this goal was achieved only partially and under certain interaction conditions. 5.3. Results: Effect of Incorporating Cognitive Process Inference (Study 2) As described in the Experiment Setup, Study 2 compared two proactive conditions that shared the same interaction paradigm but differed in how user needs were inferred: w/ Cog incorporated real-time cognitive process inference from writing interaction logs alongside document content, while w/o Cog relied solely on the evolving document text. This design examines whether incorporating log-based cognitive process prediction affects engagement with proactive support while holding the interaction paradigm constant. Predicted cognitive processes in the w/ Cog condition were unevenly distributed: Translating dominated (49%), followed by Generating (24%) and Evaluating (13%)âand transitioned dynamically over time (Fig. 8), reflecting the iterative, non-linear nature of creative writing. Figure 8. Temporal distribution of predicted cognitive processes across all Study 2 participants. Generating dominates early stages while Evaluating and Revising become more prominent later, illustrating the iterative, non-linear nature of writing. Despite differences in system behavior, most participants did not consciously perceive differences between conditions. However, interaction logs revealed systematic variations. We analyzed three conversion metrics: Tabs, Panel, and Suggestion conversion. Respectively, these capture whether participants opened a generated Functional Agent tab, interacted with generated content shown in the Panel Output, or opened each inline Suggest Output in the editor. A one-sided Wilcoxon signed-rank test revealed a significant raw p-value for Suggestion conversion (W=26,p=0.031W=26,p=0.031), though this did not retain significance after Holm correction (paâdâj=0.094p_adj=0.094). The effect size was very large (r=0.86r=0.86, well above Cohenâs threshold of r=0.5r=0.5 for a large effect (Cohen, 1988)), and the suggestion open rate more than doubled (20.5% in w/ Cog vs. 7.4% in w/o Cogâa 177% relative increase). For Tabs and Panel conversions, no significant differences were found, though Tabs showed a large effect size (r=0.51r=0.51). The combination of the very large effect size and the higher suggestion open rate points to substantially greater behavioral engagement with suggestions in the w/ Cog condition (Fig. 9). We interpret this pattern as suggestive evidence of improved contextual appropriateness of proactive support and of a contribution from interaction-derived process cues to support selection, though this pattern should be confirmed with a larger sample. Figure 9. Comparison of conversion metrics w/ and w/o cognitive process prediction. (Left) Average conversions per session (error bars: SE). (Right) Individual participant trajectories; green lines indicate increase in w/ Cog, dashed gray lines indicate decrease or no change. Line thickness represents participant count. 5.4. Summary and Limitations As an exploratory investigation, these studies aimed to demonstrate the feasibility of a new paradigmâcognitive processâaware proactive writing supportâand to derive design implications, rather than to establish population-level effects. The two studies provided complementary insights. Study 1 showed that system-inferred proactive support significantly improved expressiveness, and qualitative findings suggested benefits for idea exploration during writing blocks; however, most other quantitative measures (CSI, TLX, custom items) showed no significant differences, and participants reported challenges with intervention timing and perceived control. Study 2 found that incorporating cognitive process inference from interaction logs increased usersâ engagement with proactive suggestions (very large effect size, r=0.86r=0.86), though the difference did not retain statistical significance after multiple comparison correctionâconsistent with the limited statistical power of the sample size. We note several aspects of the study scope that contextualize these findings. As an exploratory investigation aimed at deriving design implications for cognitive processâaware proactive support, we prioritized depth of interaction analysis over breadth of population coverage. All 21 participants were engineers from a single technology company with high LLM familiarity. This participant profile was an intentional design choice: because our goal was to evaluate the systemâs interaction paradigmânot LLM capabilities per seâwe needed participants who were already fluent with prompt-based AI tools and could serve as a meaningful baseline for assessing the added value of proactive, prompt-free support. Recruiting experienced LLM users eliminated tool-onboarding noise and allowed us to isolate the effect of cognitive processâaware proactive assistance. At the same time, the creative writing task was novel for most participants, naturally eliciting the writing difficulties and support needs that the system is designed to address. The sample sizes (n=11n=11 and n=10n=10) and relatively short semi-structured interviews (10â15 minutes) are appropriate for surfacing design-relevant patterns, though the results should be interpreted as indicative rather than conclusive for broader populations. Fourteen participants had also participated in the formative study under a substantially different paradigm (self-report annotation tool vs. AI co-writing system); while this overlap may have heightened metacognitive awareness, future studies should recruit independent participant pools to rule out such effects. We focused on creative story writing with short sessions (20â25 minutes); generalizability to other writing domains and longer sessions remains to be examined. Our system adopts a fixed inactivity threshold (20 seconds) as the primary trigger for proactive intervention. This heuristic is widely used in prior work (Buschek et al., 2021; Bhat et al., 2023) and was validated through our pilot testing in the formative study (Sec. 3). However, Study 1 also revealed that timing challenges were more prominent than anticipated. When the system offers diverse support types spanning multiple cognitive processes, rather than only text continuation, the appropriateness of intervention timing becomes more context-dependent. For instance, a suggestion to restructure an outline requires a different moment of receptivity than an inline text continuation. This suggests that as proactive systems expand what they can offer, they must correspondingly adapt when they intervene, incorporating the nature of the support, the userâs cognitive process, and individual preferences. We discuss this direction further in Sec. 6.2. 6. Discussion 6.1. Balancing Proactivity and User Autonomy Our findings revealed substantial individual differences in how participants received proactive support. Some welcomed idea-level proposals, while others found them intrusiveâP14 reported that even viewing unsolicited thematic ideas felt as if âmy territory was invaded,â despite appreciating phrasing suggestions. This variation indicates that even contextually appropriate support may fail when it does not account for usersâ preferences about the scope of AI involvement. Future systems should incorporate personalization mechanisms based on naturally occurring interaction data rather than explicit feedback, reducing cognitive load while preserving agency. Despite our expectation that system-initiated suggestions might reduce ownership, no such decline was observed. Two factors may explain this. First, proactive and on-demand suggestions appeared in the same interface, making their origin often indistinguishable. Second, the âlow cognitive cost of rejectionâ may have reinforced autonomy: since users did not invest effort in prompting, they felt freer to reject outputs, paradoxically strengthening their sense of control. This observation â that the absence of prompting effort lowers the psychological barrier to discarding suggestions â has broader implications beyond ownership preservation. When users craft explicit prompts, they invest cognitive effort that creates a sunk-cost bias: they may feel compelled to use or adapt the resulting output even when it does not fully align with their intent. In contrast, proactively generated suggestions carry no such investment, allowing users to evaluate them with less anchoring bias. Moreover, because the system generates diverse suggestions spanning multiple cognitive processes simultaneously, users encounter ideas they would not have thought to request, potentially expanding their creative exploration beyond the boundaries of self-directed prompting. We propose this as a hypothesis for future investigation: that unsolicited, low-investment suggestions may not only preserve autonomy but actively broaden the writerâs idea space by reducing selection bias inherent in prompt-driven workflows. The manner of presenting proactive suggestions also significantly affected cognitive load. Participants reported that even minimal notifications (red dot badges) sometimes increased time pressure, despite the absence of pop-ups. The user experience differs substantially between noticing that something has been generated and discovering it only when intentionally checking. Future work should establish a formal notion of âunobtrusiveness in proactive suggestion designâ and empirically examine its cognitive and emotional impacts. 6.2. Future Directions for Intent Inference From âwhatâ to âwhenâ: the next frontier. Our results suggest that grounding proactive support in cognitive process inference can help address the what-to-offer problemâin our studies, the system provided support types that users engaged with meaningfully. The natural next question is when to offer it. The current system uses a fixed inactivity threshold (20 seconds) as a pragmatic trigger, combined with unobtrusive presentation that lets writers access suggestions at their own pace. However, as Study 1 revealed, this approach does not distinguish between productive reflection and genuine writing blocks: writers may be thinking through a problem during pauses, or conversely may struggle while actively typing. This mismatch was a primary source of the timing challenges and perceived intrusiveness reported by participants. Promisingly, the same cognitive process inference that addresses the what problem may also inform the when problem: our formative study showed that the average pause duration preceding a block varied across cognitive processes, suggesting that optimal intervention timing is process-dependent. Future systems should move beyond simple inactivity detection toward adaptive triggering mechanisms that incorporate cognitive process context, behavioral patterns (e.g., repeated deletions suggesting struggle), and individual workflow preferences. Beyond timing, the limits of inference from interaction data alone became apparent. Some functional agents in AToM CoWriter were designed to clarify user intent through follow-up questions, but these features were rarely used in practice. The closed-question interface may have lacked sufficient precision and perceived relevance. Future research should explore low-burden, context-aware ways to elicit additional information, improving both accuracy and user trust. Observer effect in behavioral monitoring. An important consideration for future work is whether continuous behavioral monitoring introduces an observer effect, where users modify their writing behavior because they know (or suspect) that the system is inferring their cognitive processes. Such behavioral changes could affect perceived agency, ownership, and trust, and may be particularly relevant for broader user populations beyond experienced LLM users. Our current study did not explicitly measure this effect, but it warrants investigation as cognitive processâaware writing systems become more prevalent. Finally, as AI-assisted writing becomes increasingly common, the classical Flower and Hayes model may require extension. Users in our study engaged in meta-cognitive activitiesâsuch as formulating prompts or evaluating AI suggestionsâthat are not captured in the original six-process model. Future work should aim to extend the cognitive process theory for the AI era, modeling new processes and transitions that emerge from humanâAI collaboration, enabling more precise and context-sensitive proactive assistance. 7. Conclusion Providing proactive writing support requires addressing two problems: what support to offer and when to offer it. This work explored whether cognitive processes can provide a theory-grounded basis for determining what support to offer. Through a formative study and literature review, we developed a framework linking observable writing interactions to FlowerâHayes cognitive processes and fourteen support types identified in our study. We instantiated this framework in AToM CoWriter, which uses document context and writing interaction logs to infer cognitive processes and select appropriate forms of writing assistance. Two user studies provide initial empirical evidence for this approach. System-inferred support was associated with improved expressiveness, while qualitative findings suggested benefits for idea exploration. Incorporating interaction-derived cognitive-process cues alongside document context was associated with substantially greater engagement with proactively generated suggestions. These findings suggest that behavioral interaction signals can help inform support selection in proactive writing systems. At the same time, our studies surfaced intervention timing as an important complementary challenge. The current inactivity-based trigger, combined with unobtrusive presentation, provides a pragmatic but coarse solution; future systems should explore adaptive timing mechanisms that account for cognitive processes, interaction patterns, and individual preferences. Overall, our findings suggest that cognitive processes can serve as a useful theory-grounded bridge between observable writing behavior and the selection of proactive writing support. References Afrin et al. (2021) Tazin Afrin, Omid Kashefi, Christopher Olshefski, Diane Litman, Rebecca Hwa, and Amanda Godley. 2021. Effective interfaces for student-driven revision sessions for argumentative writing. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA. Afrin and Litman (2023) Tazin Afrin and Diane Litman. 2023. Predicting desirable revisions of evidence and reasoning in argumentative writing. In Findings of the Association for Computational Linguistics: EACL 2023. Association for Computational Linguistics, Stroudsburg, PA, USA, 2550â2561. Allen et al. (1999) J E Allen, C I Guinn, and E Horvtz. 1999. Mixed-initiative interaction. IEEE Intell. Syst. 14, 5 (Sept. 1999), 14â23. Arnold et al. (2020) Kenneth C Arnold, Krysta Chauncey, and Krzysztof Z Gajos. 2020. Predictive text encourages predictable writing. In Proceedings of the 25th International Conference on Intelligent User Interfaces. ACM, New York, NY, USA. Arshad et al. (2015) Syed Arshad, Yang Wang, and Fang Chen. 2015. Interactive mouse stream as real-time indicator of userâs cognitive load. In Proceedings of the 33rd Annual ACM Conference Extended Abstracts on Human Factors in Computing Systems. ACM, New York, NY, USA. Baaijen et al. (2012) Veerle M Baaijen, David Galbraith, and Kees de Glopper. 2012. Keystroke analysis: Reflections on procedures and measures. Writ. Commun. 29, 3 (July 2012), 246â277. Bao et al. (2022) Jianzhu Bao, Yasheng Wang, Yitong Li, Fei Mi, and Ruifeng Xu. 2022. AEG: Argumentative essay generation via A dual-decoder model with content planning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, Stroudsburg, PA, USA, 5134â5148. Beauvais et al. (2011) Caroline Beauvais, Thierry Olive, and Jean-Michel Passerault. 2011. Why are some texts good and others not? Relationship between text quality and management of the writing processes. J. Educ. Psychol. 103, 2 (May 2011), 415â428. Becker (2006) Anne Becker. 2006. A review of writing model research based on cognitive processes. Revision: History, theory, and practice (2006), 25â49. Bhat et al. (2023) Advait Bhat, Saaket Agashe, Parth Oberoi, Niharika Mohile, Ravi Jangir, and Anirudha Joshi. 2023. Interacting with Next-Phrase Suggestions: How Suggestion Systems Aid and Influence the Cognitive Processes of Writing. In Proceedings of the 28th International Conference on Intelligent User Interfaces (IUI â23). Association for Computing Machinery, New York, NY, USA, 436â452. Biermann et al. (2022) Oloff C Biermann, Ning F Ma, and Dongwook Yoon. 2022. From tool to companion: Storywriters want AI writers to respect their personal values and writing strategies. In Designing Interactive Systems Conference. ACM, New York, NY, USA. Buschek et al. (2021) Daniel Buschek, Martin ZĂźrn, and Malin Eiband. 2021. The Impact of Multiple Parallel Phrase Suggestions on Email Input and Composition Behaviour of Native and Non-Native English Writers. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (CHI â21, Article 732). Association for Computing Machinery, New York, NY, USA, 1â13. Carrera et al. (2026) Dashiel Carrera, Jeb Thomas-Mitchell, and Daniel Wigdor. 2026. Where do I âadd the eggâ?: Exploring agency and ownership in AI creative co-writing systems. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â16. Chakrabarty et al. (2024) Tuhin Chakrabarty, Vishakh Padmakumar, Faeze Brahman, and Smaranda Muresan. 2024. Creativity support in the age of large language models: An empirical study involving professional writers. In Creativity and Cognition. ACM, New York, NY, USA, 132â155. Chang and Chang (2015) Jim Chang and Jason S Chang. 2015. WriteAhead2: Mining Lexical Grammar Patterns for Assisted Writing. In Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Demonstrations. 106â110. Chen et al. (2025) Valerie Chen, Alan Zhu, Sebastian Zhao, Hussein Mozannar, David Sontag, and Ameet Talwalkar. 2025. Need help? Designing proactive AI assistants for programming. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â18. Cherry and Latulipe (2014) Erin Cherry and Celine Latulipe. 2014. Quantifying the Creativity Support of Digital Tools through the Creativity Support Index. ACM Trans. Comput.-Hum. Interact. 21, 4 (June 2014), 1â25. Chu-Carroll (2000) Jennifer Chu-Carroll. 2000. MIMIC: an adaptive mixed initiative spoken dialogue system for information queries. In Proceedings of the sixth conference on Applied natural language processing -. Association for Computational Linguistics, Morristown, NJ, USA. Chung et al. (2022) John Joon Young Chung, Wooseok Kim, Kang Min Yoo, Hwaran Lee, Eytan Adar, and Minsuk Chang. 2022. TaleBrush: Sketching Stories with Generative Pretrained Language Models. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI â22, Article 209). Association for Computing Machinery, New York, NY, USA, 1â19. Clark et al. (2018) Elizabeth Clark, Anne Spencer Ross, Chenhao Tan, Yangfeng Ji, and Noah A Smith. 2018. Creative writing with a machine in the loop: Case studies on slogans and stories. In 23rd International Conference on Intelligent User Interfaces. ACM, New York, NY, USA. Cockburn et al. (2007) Andy Cockburn, Carl Gutwin, and Saul Greenberg. 2007. A predictive model of menu performance. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA. Cohen (1988) Jacob Cohen. 1988. Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Lawrence Erlbaum Associates. Conijn et al. (2024) Rianne Conijn, Alessandra Rossetti, Nina Vandermeulen, and Luuk Van Waes. 2024. Phase to phase: Towards an automated procedure to identify phases in writing processes using keystroke data. Social Science Research Network (Oct. 2024). Dang et al. (2022) Hai Dang, Karim Benharrak, Florian Lehmann, and Daniel Buschek. 2022. Beyond Text Generation: Supporting Writers with Continuous Automatic Text Summaries. In Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology (UIST â22, Article 98). Association for Computing Machinery, New York, NY, USA, 1â13. Dang et al. (2023) Hai Dang, Sven Goller, Florian Lehmann, and Daniel Buschek. 2023. Choice over control: How users write with large language models using diegetic and non-diegetic prompting. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â17. Delgado-Osorio et al. (2025) Ximena Delgado-Osorio, Johannes Hartig, Claudia Harsch, and Valeriia Koval. 2025. Studentsâ behavioral patterns in integrated writing tasks: A sequence analysis approach. Journal of Educational Psychology (2025). Ding et al. (2026) Lei Ding, Bin He, Chenguang Wang, and Yang Liu. 2026. ProActor: Timing-aware reinforcement learning for proactive task scheduling agents. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Association for Computational Linguistics, Stroudsburg, PA, USA, 18257â18303. Draxler et al. (2024) Fiona Draxler, Anna Werner, Florian Lehmann, Matthias Hoppe, Albrecht Schmidt, Daniel Buschek, and Robin Welsch. 2024. The AI ghostwriter effect: When users do not perceive ownership of AI-generated text but self-declare as authors. ACM Trans. Comput. Hum. Interact. 31, 2 (April 2024), 1â40. El Alaoui et al. (2025) Hamza El Alaoui, Atieh Taheri, Yi-Hao Peng, and Jeffrey P Bigham. 2025. StepWrite: Adaptive Planning for Speech-Driven Text Generation. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology. ACM, New York, NY, USA, 1â30. Flower and Hayes (1981) Linda Flower and John R Hayes. 1981. A Cognitive Process Theory of Writing. College Composition and Communication 32, 4 (1981), 365â387. Gajos and Weld (2004) Krzysztof Gajos and Daniel S Weld. 2004. SUPPLE: automatically generating user interfaces. In Proceedings of the 9th international conference on Intelligent user interfaces. ACM, New York, NY, USA. Gero et al. (2022a) Katy Gero, Alex Calderwood, Charlotte Li, and Lydia Chilton. 2022a. A Design Space for Writing Support Tools Using a Cognitive Process Model of Writing. In Proceedings of the First Workshop on Intelligent and Interactive Writing Assistants (In2Writing 2022), Ting-Hao âkennethâ Huang, Vipul Raheja, Dongyeop Kang, John Joon Young Chung, Daniel Gissin, Mina Lee, and Katy Ilonka Gero (Eds.). Association for Computational Linguistics, Dublin, Ireland, 11â24. Gero et al. (2022b) Katy Ilonka Gero, Vivian Liu, and Lydia Chilton. 2022b. Sparks: Inspiration for Science Writing using Language Models. In Designing Interactive Systems Conference. ACM, New York, NY, USA. Goldfarb-Tarrant et al. (2019) Seraphina Goldfarb-Tarrant, Haining Feng, and Nanyun Peng. 2019. Plan, Write, and Revise: an Interactive System for Open-Domain Story Generation. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (Demonstrations), Waleed Ammar, Annie Louis, and Nasrin Mostafazadeh (Eds.). Association for Computational Linguistics, Minneapolis, Minnesota, 89â97. Guo et al. (2025) Alicia Guo, Shreya Sathyanarayanan, Leijie Wang, Jeffrey Heer, and Amy X Zhang. 2025. From pen to prompt: How creative writers integrate AI into their writing practice. In Proceedings of the 2025 Conference on Creativity and Cognition. ACM, New York, NY, USA, 527â545. Guo et al. (2018) Hongwen Guo, Paul D Deane, Peter W van Rijn, Mo Zhang, and Randy E Bennett. 2018. Modeling basic writing processes from keystroke logs. Journal of educational measurement 55, 2 (June 2018), 194â216. GĂśldi et al. (2024) Andreas GĂśldi, Thiemo Wambsganss, Seyed Parsa Neshaei, and Roman Rietsche. 2024. Intelligent Support Engages Writers Through Relevant Cognitive Processes. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI â24, Article 1047). Association for Computing Machinery, New York, NY, USA, 1â12. Haga and Mizukami (1996) Shigeru Haga and Naoki Mizukami. 1996. Japanese Version of NASA Task Load Index: Sensitivity of Its Workload Score to the Difficulty of Three Different Laboratory Tasks. The Japanese Journal of Ergonomics 32, 2 (1996), 71â79. https://doi.org/10.5100/jje.32.71 Hart and Staveland (1988) Sandra G Hart and Lowell E Staveland. 1988. Development of NASA-TLX (task load index): Results of empirical and theoretical research. In Advances in Psychology. Advances in Psychology, Vol. 52. Elsevier, 139â183. Hayes (2012) John R Hayes. 2012. Modeling and Remodeling Writing. Written Communication 29, 3 (July 2012), 369â388. Horvitz (1999) Eric Horvitz. 1999. Principles of mixed-initiative user interfaces. In Proceedings of the SIGCHI conference on Human factors in computing systems the CHI is the limit - CHI â99. ACM Press, New York, New York, USA. Hu et al. (2026) Yufeng Hu, Jinyi Zhang, Zehuan Wang, and Chun Yu. 2026. Plotania: Exploring Transparency Trade-offs in AI Co-Writing Through Virtual Readers and Transparent Attribution. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â23. Iqbal et al. (2018) Shamsi T Iqbal, Jaime Teevan, Dan Liebling, and Anne Loomis Thompson. 2018. Multitasking with play write, a mobile microproductivity writing tool. In Proceedings of the 31st Annual ACM Symposium on User Interface Software and Technology. ACM, New York, NY, USA, 411â422. J Shen et al. (2026) Jocelyn J Shen, Nicolai Marquardt, Hugo Romat, Ken Hinckley, Nathalie Riche, and Fanny Chevalier. 2026. Texterial: A text-as-material interaction paradigm for LLM-mediated writing. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â18. Jakesch et al. (2023) Maurice Jakesch, Advait Bhat, Daniel Buschek, Lior Zalmanson, and Mor Naaman. 2023. Co-Writing with Opinionated Language Models Affects Usersâ Views. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI â23, Article 111). Association for Computing Machinery, New York, NY, USA, 1â15. Kaur et al. (2026) Kirandeep Kaur, Vinayak Gupta, Aditya Gupta, and Chirag Shah. 2026. PROPER agents: Proactivity driven personalized agents for advancing knowledge gap navigation. arXiv [cs.LG] (Jan. 2026). Ke et al. (2017) Guolin Ke, Qi Meng, Thomas Finley, Taifeng Wang, Wei Chen, Weidong Ma, Qiwei Ye, and Tie-Yan Liu. 2017. LightGBM: A highly efficient gradient boosting decision tree. Neural Inf Process Syst (Dec. 2017), 3146â3154. Kerby (2014) Dave S. Kerby. 2014. The Simple Difference Formula: An Approach to Teaching Nonparametric Correlation. Comprehensive Psychology 3 (2014), 11.IT.3.1. https://doi.org/10.2466/11.IT.3.1 Kuo et al. (2026) Nadine Kuo, Agnia Sergeyuk, Valerie Chen, and Maliheh Izadi. 2026. Developer interaction patterns with proactive AI: A five-day field study. In Proceedings of the 31st International Conference on Intelligent User Interfaces. ACM, New York, NY, USA, 349â362. Laban et al. (2024) Philippe Laban, Jesse Vig, Marti Hearst, Caiming Xiong, and Chien-Sheng Wu. 2024. Beyond the chat: Executable and verifiable text-editing with LLMs. In Proceedings of the 37th Annual ACM Symposium on User Interface Software and Technology, Vol. 10. ACM, New York, NY, USA, 1â23. Lee et al. (2024) Mina Lee, Katy Ilonka Gero, John Joon Young Chung, Simon Buckingham Shum, Vipul Raheja, Hua Shen, Subhashini Venugopalan, Thiemo Wambsganss, David Zhou, Emad A Alghamdi, Tal August, Avinash Bhat, Madiha Zahrah Choksi, Senjuti Dutta, Jin L C Guo, Md Naimul Hoque, Yewon Kim, Simon Knight, Seyed Parsa Neshaei, Antonette Shibani, Disha Shrivastava, Lila Shroff, Agnia Sergeyuk, Jessi Stark, Sarah Sterman, Sitong Wang, Antoine Bosselut, Daniel Buschek, Joseph Chee Chang, Sherol Chen, Max Kreminski, Joonsuk Park, Roy Pea, Eugenia Ha Rim Rho, Zejiang Shen, and Pao Siangliulue. 2024. A Design Space for Intelligent and Interactive Writing Assistants. In Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI â24, Article 1054). Association for Computing Machinery, New York, NY, USA, 1â35. Lee et al. (2022) Mina Lee, Percy Liang, and Qian Yang. 2022. CoAuthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems (CHI â22, Article 388). Association for Computing Machinery, New York, NY, USA, 1â19. Lin et al. (2023) Zhiyu Lin, Upol Ehsan, Rohan Agarwal, Samihan Dani, Vidushi Vashishth, and Mark Riedl. 2023. Beyond Prompts: Exploring the Design Space of Mixed-Initiative Co-Creativity Systems. 14th International Conference on Computational Creativity, ICCCâ23 (May 2023). Lin et al. (2024) Zhiyu Lin, Upol Ehsan, Rohan Agarwal, Samihan Dani, Vidushi Vashishth, and Mark Riedl. 2024. Beyond following: Mixing active initiative into computational creativity. arXiv [cs.HC] (Sept. 2024). Liu et al. (2026b) Chang Liu, Qinyi Zhou, Xinjie Shen, Xingyu Bruce Liu, Tongshuang Wu, and Xiang âanthony Chen. 2026b. Behavioral indicators of overreliance during interaction with conversational language models. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â23. Liu et al. (2025) Xingyu Bruce Liu, Shitao Fang, Weiyan Shi, Chien-Sheng Wu, Takeo Igarashi, and Xiang âanthonyâ Chen. 2025. Proactive conversational agents with inner thoughts. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â19. Liu et al. (2026a) Yijun Liu, John Gallagher, Sarah Sterman, and Tal August. 2026a. From Crafting Text to Crafting Thought: Grounding AI Writing Support to Writing Center Pedagogy. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â36. Liu et al. (2019) Yuanchao Liu, B Pang, and Bingquan Liu. 2019. Neural-based Chinese idiom recommendation for enhancing elegance in essay writing. Annu Meet Assoc Comput Linguistics (July 2019), 5522â5526. MacLean et al. (1991) Allan MacLean, Richard Young, Victoria Bellotti, and Thomas Moran. 1991. Questions, options, and criteria: Elements of design space analysis. Human-Computer Interaction 6, 3 (Sept. 1991), 201â250. Masson et al. (2025) Damien Masson, Zixin Zhao, and Fanny Chevalier. 2025. Visual story-writing: Writing by manipulating visual representations of stories. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology. ACM, New York, NY, USA, 1â15. Mirowski et al. (2023) Piotr Mirowski, Kory W Mathewson, Jaylen Pittman, and Richard Evans. 2023. Co-Writing Screenplays and Theatre Scripts with Language Models: Evaluation by Industry Professionals. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (CHI â23, Article 355). Association for Computing Machinery, New York, NY, USA, 1â34. Mozannar et al. (2024) Hussein Mozannar, Gagan Bansal, Adam Fourney, and Eric Horvitz. 2024. Reading between the lines: Modeling user behavior and costs in AI-assisted programming. In Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â16. Osone et al. (2021) Hiroyuki Osone, Jun-Li Lu, and Yoichi Ochiai. 2021. BunCho: AI Supported Story Co-Creation via Unsupervised Multitask Learning to Increase Writersâ Creativity in Japanese. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems (CHI EA â21, Article 19). Association for Computing Machinery, New York, NY, USA, 1â10. Padmakumar and He (2022) Vishakh Padmakumar and He He. 2022. Machine-in-the-loop rewriting for creative image captioning. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. Association for Computational Linguistics, Stroudsburg, PA, USA, 573â586. Peng et al. (2020) Zhenhui Peng, Qingyu Guo, Ka Wing Tsang, and Xiaojuan Ma. 2020. Exploring the effects of technological writing assistance for support providers in online mental health community. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA. Pu et al. (2025) Kevin Pu, Daniel Lazaro, Ian Arawjo, Haijun Xia, Ziang Xiao, Tovi Grossman, and Yan Chen. 2025. Assistance or disruption? Exploring and evaluating the design and trade-offs of proactive AI programming support. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â21. Qin et al. (2025) Hua Xuan Qin, Guangzhi Zhu, Mingming Fan, and Pan Hui. 2025. Toward personalizable AI node graph creative writing support: Insights on preferences for generative AI features and information presentation across story writing processes. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â30. https://doi.org/10.1145/3706598.3713569 Roemmele and Gordon (2018) Melissa Roemmele and Andrew S Gordon. 2018. Automated assistance for creative writing with an RNN language model. In Companion Proceedings of the 23rd International Conference on Intelligent User Interfaces. ACM, New York, NY, USA, 1â2. Schmitt and Buschek (2021) Oliver Schmitt and Daniel Buschek. 2021. CharacterChat: Supporting the creation of fictional characters through conversation and progressive manifestation with a chatbot. In Creativity and Cognition. ACM, New York, NY, USA. Siddiqui et al. (2025) Momin N Siddiqui, Roy D Pea, and Hari Subramonyam. 2025. Script&Shift: A Layered Interface Paradigm for Integrating Content Development and Rhetorical Strategy with LLM Writing Assistants. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â19. Singh et al. (2023) Nikhil Singh, Guillermo Bernal, Daria Savchenko, and Elena L Glassman. 2023. Where to hide a stolen elephant: Leaps in creative writing with multimodal machine intelligence. ACM Trans. Comput. Hum. Interact. 30, 5 (Oct. 2023), 1â57. Steyvers and Mayer (2025) Mark Steyvers and Lukas Mayer. 2025. When not to help: planning for lasting human-AI collaboration. arXiv [cs.HC] (Aug. 2025). Subramonyam et al. (2024) Hari Subramonyam, Roy Pea, Christopher Pondoc, Maneesh Agrawala, and Colleen Seifert. 2024. Bridging the gulf of envisioning: Cognitive challenges in prompt based interactions with LLMs. In Proceedings of the CHI Conference on Human Factors in Computing Systems, Vol. 31. ACM, New York, NY, USA, 1â19. Talaei et al. (2025) Shayan Talaei, Meijin Li, Kanu Grover, James Kent Hippler, Diyi Yang, and Amin Saberi. 2025. StorySage: Conversational autobiography writing powered by a multi-agent framework. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology. ACM, New York, NY, USA, 1â26. Tsai et al. (2020) Chung-Ting Tsai, Jhih-Jie Chen, Ching-Yu Yang, and Jason S Chang. 2020. LinggleWrite: A coaching system for essay writing. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations. Association for Computational Linguistics, Stroudsburg, PA, USA, 127â133. TĂźtĂźncĂź et al. (2026) Esen K. TĂźtĂźncĂź, Qian Zhou, Frederik Brudy, George Fitzmaurice, and Fraser Anderson. 2026. PlayWrite: A Multimodal System for AI Supported Narrative Co- Authoring Through Play in XR. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â26. Vargha and Delaney (2000) AndrĂĄs Vargha and Harold D. Delaney. 2000. A Critique and Improvement of the CL Common Language Effect Size Statistics of McGraw and Wong. Journal of Educational and Behavioral Statistics 25, 2 (2000), 101â132. https://doi.org/10.2307/1165329 Wambsganss et al. (2022) Thiemo Wambsganss, Matthias Soellner, Kenneth R Koedinger, and Jan Marco Leimeister. 2022. Adaptive empathy learning support in peer review scenarios. In CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA. Wan et al. (2024b) Qian Wan, Siying Hu, Yu Zhang, Piaohong Wang, Bo Wen, and Zhicong Lu. 2024b. âit felt like having a second mindâ: Investigating human-AI co-creativity in prewriting with large language models. Proc. ACM Hum. Comput. Interact. 8, CSCW1 (April 2024), 1â26. Wan et al. (2025) Qian Wan, Jiannan Li, Huanchen Wang, and Zhicong Lu. 2025. Polymind: Parallel visual diagramming with large language models to support prewriting through microtasks. Proc. ACM Hum. Comput. Interact. 9, 7 (Oct. 2025), 1â29. Wan et al. (2024a) Ruyuan Wan, Simret Araya Gebreegziabher, Toby Jia-Jun Li, and Karla Badillo-Urquiola. 2024a. CoCo matrix: Taxonomy of cognitive contributions in co-writing with intelligent agents. In Creativity and Cognition. ACM, New York, NY, USA. Wang et al. (2025) Linghe Wang, Minhwa Lee, Ross Volkov, Luan Tuyen Chau, and Dongyeop Kang. 2025. ScholaWrite: A dataset of end-to-end scholarly writing process. arXiv [cs.HC] (Feb. 2025). Wang and Goel (2024) Qiaosi Wang and Ashok K Goel. 2024. Mutual Theory of Mind for human-AI communication. In Workshop on Theory of Mind in Human-AI Interaction at CHI 2024. Wang et al. (2021) Qiaosi Wang, Koustuv Saha, Eric Gregori, David Joyner, and Ashok Goel. 2021. Towards mutual theory of mind in human-AI interaction: How language reflects what students perceive about a virtual teaching assistant. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA. Wu et al. (2019) Shaomei Wu, Lindsay Reynolds, Xian Li, and Francisco GuzmĂĄn. 2019. Design and evaluation of a social media writing support tool for people with dyslexia. In Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA. Yao et al. (2026) Rui Yao, Qiuyuan Ren, Felicia Fang-Yi Tan, Yang Chen, Xiaoyu Zhang, and Shengdong Zhao. 2026. PersonaMail: Learning and adapting personal communication preferences for context-aware email writing. In Proceedings of the 31st International Conference on Intelligent User Interfaces. ACM, New York, NY, USA, 197â217. Yoshida et al. (2026) Masahiro Yoshida, Bingxuan Li, Songyan Zhao, Qinyi Zhou, Shiwei Hu, Xiang âanthony Chen, and Nanyun Peng. 2026. CoLyricist: Enhancing lyric writing with AI through workflow-aligned support. In Proceedings of the 31st International Conference on Intelligent User Interfaces. ACM, New York, NY, USA, 1387â1410. Yuan et al. (2022) Ann Yuan, Andy Coenen, Emily Reif, and Daphne Ippolito. 2022. Wordcraft: Story Writing With Large Language Models. In 27th International Conference on Intelligent User Interfaces (IUI â22). Association for Computing Machinery, New York, NY, USA, 841â852. Zhang et al. (2026) Chao Zhang, Shunan Guo, Abe Davis, and Eunyee Koh. 2026. Narrix: Remixing narrative strategies from examples for story writing. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â24. Zhang et al. (2025b) Chao Zhang, Kexin Ju, Peter Bidoshi, Yu-Chun Grace Yen, and Jeffrey M. Rzeszotarski. 2025b. Friction: Deciphering Writing Feedback into Writing Revisions through LLM-Assisted Reflection. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â27. Zhang et al. (2025c) Chao Zhang, Kexin Ju, Zhuolun Han, Yu-Chun Grace Yen, and Jeffrey M Rzeszotarski. 2025c. Synthia: Visually interpreting and synthesizing feedback for writing revision. In Proceedings of the 38th Annual ACM Symposium on User Interface Software and Technology. ACM, New York, NY, USA, 1â16. Zhang et al. (2024) Guanhua Zhang, Zhiming Hu, Mihai Bâce, and Andreas Bulling. 2024. Mouse2Vec: Learning reusable semantic representations of mouse behaviour. In Proceedings of the CHI Conference on Human Factors in Computing Systems, Vol. 20. ACM, New York, NY, USA, 1â17. Zhang et al. (2025a) Mo Zhang, Paul Deane, Andrew Hoang, Hongwen Guo, and Chen Li. 2025a. Applications and modeling of keystroke logs in writing assessments. Educational Measurement: Issues and Practice 44, 2 (June 2025), 5â19. Zhao et al. (2025) Zixin Zhao, Damien Masson, Young-Ho Kim, Gerald Penn, and Fanny Chevalier. 2025. Making the write connections: Linking writing support tools with writerâs needs. In Proceedings of the CHI Conference on Human Factors in Computing Systems. Association for Computing Machinery. Zhou et al. (2025) Shixu Zhou, Weiyue Lin, Zuyu Xu, Xiaoying Wei, Raoyi Huang, Xiaojuan Ma, and Mingming Fan. 2025. JournalAIde: Empowering Older Adults in Digital Journal Writing. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. ACM, New York, NY, USA, 1â19. Appendix A Formative Study Data A.1. Code Book This code book summarizes the categories used to analyze participantsâ verbal reports collected during the formative study. Each participantâs spoken responsesârecorded when they felt âblockedâ during writing-were first automatically transcribed. All responses were in Japanese. We developed the coding scheme, focusing on the types of problems participants faced (Problem) and the types of support they requested from the AI (Request). A.1.1. Problem (User Difficulties) The following codes represent types of difficulties participants reported during writing: ⢠IDEA_GENERATION: Difficulty in generating ideas. ⢠STORY_DEVELOPMENT: Trouble developing the storyline. ⢠CHARACTER_SETTING: Difficulty setting up characters. ⢠WORLD_BUILDING: Difficulty constructing the story world or setting. ⢠WRITING_TECHNIQUE: Struggling with writing expressions or style. ⢠STORY_STRUCTURE: Difficulty structuring the story (e.g., introduction, development, climax, ending). ⢠NARRATIVE_FLOW: Issues with flow, coherence, or readability. ⢠OPENING_WRITING: Difficulty starting the story or writing an engaging introduction. ⢠ENDING_WRITING: Difficulty concluding or wrapping up the story. ⢠GRAMMAR_STYLE: Concerns about grammar or stylistic correctness. ⢠PROOFREADING: Needs for proofreading or error checking. ⢠WORD_CHOICE: Uncertainty about word selection or phrasing. ⢠WRITING_PROCESS: Uncertainty about how to start or proceed with writing. ⢠TIME_MANAGEMENT: Difficulty managing time during writing. ⢠CREATIVE_BLOCK: Feeling inexperienced or lacking confidence in creative writing. ⢠KNOWLEDGE_GAP: Forgetting or lacking specific knowledge (e.g., âI canât recall this termâ). ⢠TECHNICAL_ISSUE: Technical or operational issues with the writing tool. A.1.2. Request (User Requests for Support) The following codes represent the types of support participants desired from the AI system. Type of Requested Operation: ⢠IDEA_GENERATION: Generate new ideas or suggestions. ⢠TEXT_WRITING: Write or continue a story or text. ⢠TEXT_EDITING: Edit or refine existing text. ⢠REVIEW_FEEDBACK: Review, evaluate, or provide feedback on text. ⢠STRUCTURE_PLANNING: Create or improve the structure, plot, or outline. ⢠INFORMATION_SEARCH: Search for or retrieve external information. ⢠CONSULTATION: Provide conversational or advisory support. Nature of Expected Output: ⢠LIST_OUTPUT: Provide results in a list or bullet-point format. ⢠TEXT_OUTPUT: Provide narrative or full-text output. ⢠TEMPLATE_OUTPUT: Provide a structured template or format. ⢠EVALUATION_OUTPUT: Provide evaluative comments, judgments, or critiques. Target of Operation: ⢠DIRECT_EDITING: Directly edit the main text being written. ⢠IDEA_SUPPORT: Support idea generation or brainstorming. ⢠WRITING_SUPPORT: Provide meta-level writing process support. A.2. Prompts for Cognitive Process Prediction The following prompt template was used for the cognitive process prediction task in the formative study. Curly brackets ( ) indicate variables dynamically filled with the userâs log and document information at the time of prediction. Note that the original prompt was written in Japanese. The following is a direct translation ⏠You are provided with log data from a text editor used by a participant writing a short story. When the participant feels âstuck,â they record this moment in the system. Below is a summary of the userâs activity just before they reported being stuck. Based on this summary, please predict which of the following six cognitive processes the user was likely in. Definition of Cognitive Processes: - Goal-Setting: Thinking about the purpose or direction of the story. - Generating: Coming up with or writing down ideas and information. - Organizing: Structuring or categorizing the generated ideas. - Translating: Converting thoughts or ideas into linguistic form. - Evaluating: Reviewing and assessing the written text. - Revising: Improving text or ideas based on evaluation. Context: - Many participants begin by writing outlines or ideas in the memo area rather than directly in the story text. - The writing sessions typically last about 40 minutes. - Among all self-reported âblockedâ moments, approximately half correspond to the Generating state. (Estimated distribution: Goal-Setting 0.15, Generating 0.40, Organizing 0.10, Translating 0.15, Evaluating 0.10, Revising 0.10) Relationships between Cognitive Processes and User Behaviors: - Goal-Setting - Often occurs at the very beginning of the session. - Common when little to no text has been written yet. - Frequently observed when the main text is empty and a few notes exist in the memo. - More likely when the cursor is positioned in the memo area. - Generating - Can occur throughout the session, but especially before writing the main text or while editing notes. - More likely when the cursor is in the memo area. - The most frequent state (approximately 40% of all reports). - Also likely when editing bullet points or brainstorming notes. - Organizing - Usually occurs while writing or editing notes rather than the main text. - Common once a sufficient number of ideas have been collected in the memo. - Typically occurs when the cursor is in the memo area and during the early stage of writing. - The main text is often still empty, with only partial content in the memo. - Also likely when editing bullet-pointed content. - Translating - Occurs after finishing notes and starting to write the main story. - Often happens when the current sentence is incomplete. - The cursor may be in the middle of a sentence. - Sometimes accompanied by delete operations. - Represents roughly 10% of all reported states. - Evaluating - Occurs once the story has become more complete; the main text typically exceeds the memo in length. - Often involves many deletions. - Writing speed tends to be normal or slightly slow due to rereading. - Rarely happens in the middle of a sentence; the last sentence is usually complete. - The cursor is typically in the main text area. - Scrolling and cursor movement often increase as the user reviews the text. - Writing speed tends to decrease. - Revising - Also occurs when the text is relatively complete; the main text is usually longer than the memo. - Writing speed tends to be normal or slightly slow. - Rarely happens mid-sentence; the final sentence is typically complete. - The cursor is typically in the main text area. - Scrolling and cursor movement often increase as the user makes edits. - Writing speed tends to decrease. Current Writing Context: - Main text (cursor position indicated by cursor_token; if missing, the editor is out of focus): ###################################### text_with_cursor ###################################### - Total number of characters: text_len - Number of characters in the main text: text_story_len - Number of characters in the memo: text_memo_len - Elapsed time: relative_time - Cursor position (macro level): cursor_position_macro - End of document: document_ending Recent User Activity: - Recent writing speed (relative to normal): text_speed - Recent delete events: delete_event - Recent scroll events: scroll_event - Recent cursor movement: cursor_event Instruction: Based on the above information, determine which cognitive process the user is currently in. Select one from the following: [Goal-Setting, Generating, Organizing, Translating, Evaluating, Revising] Your output should contain only the predicted cognitive process. A.3. Cognitive Process Prediction Results We evaluated prediction accuracy, macro-F1, and weighted-F1 scores for both the six-process classification and a coarser three-process mapping (Planning, Translating, Reviewing). As shown in Tables 4 and 5, the LLM (log + document) achieved the highest performance across all metrics, followed by the LLM (document only) and the LightGBM (log only). Note that the LLM-based approach operates in a near zero-shot manner (no gradient-based training), making it less dependent on dataset size than traditional supervised models. Table 4. Cognitive Process Prediction Performance: 6 class LLM (Claude Sonnet 3.5) LightGBM log + doc doc only log only Accuracy 0.54 0.47 0.45 Macro F1 0.30 0.21 0.18 Weighted F1 0.49 0.41 0.40 Table 5. Cognitive Process Prediction Performance: 3 class LLM (Claude Sonnet 3.5) LightGBM log + doc doc only log only Accuracy 0.79 0.74 0.72 Macro F1 0.63 0.39 0.46 Weighted F1 0.78 0.67 0.69 A.4. Features used in Cognitive Prediction Table 6 summarizes the features used for cognitive process prediction in the Formative Study. Both the LightGBM (log only) and LLM (log + document) models utilized largely the same feature set; however, in the LLM model, several continuous features were converted into categorical representations to better capture qualitative distinctions in user behavior. Table 6. Features Used for Cognitive Process Prediction in LightGBM and LLM Models Feature LightGBM (log only) LLM (log + document) Text Length [int] Total number of characters in the document. Same as left. Text Length in Main Story [int] Number of characters in the Main Story section. Same as left. Text Length in Memo [int] Number of characters in the Memo section. Same as left. Elapsed Time since Start [TimeDelta] Elapsed time since the session began. Same as left. Cursor Position [Category] Indicates whether the cursor is positioned in the Memo or Main Story. Descriptive text explaining each cursor position category. Document Ending Status [binary] Indicates whether the last sentence in the document is complete (i.e., ends with a period). Descriptive text explaining each binary category. Text Writing Speed [float] Average number of characters added or removed per minute in the last one minute. The left value is divided into three classes using the 25th and 75th percentiles, with descriptive text for each class. Delete Count [int] Total number of characters deleted during the last two minutes. The left value is divided into two classes using the 50th percentile, with descriptive text for each class. Scroll Count [int] Total number of pixels scrolled during the last two minutes. The left value is divided into two classes using the 50th percentile, with descriptive text for each class. Cursor Move Count [int] Total number of character-level cursor movements during the last two minutes. The left value is divided into two classes using the 50th percentile, with descriptive text for each class. A.5. Prompts for Predicting User Process from User Self-report transcription ⏠This CSV file contains transcribed user utterances made to the AI agent while performing a story-writing task in a text editor. Please analyze each utterance and determine which of the following cognitive processes from Flower and Hayes (1980) best represents the type of behavior the agent is being asked to perform. Then, fill in the corresponding label in the agent_state column of the CSV file. Generating: generating ideas Organizing: organizing ideas Goal-Setting: planning goals or the final output Translating: actual writing Evaluating: evaluating the written content Revising: revising the written content The agent_state column should be filled in as follows: generating organizing translating revising goal-setting evaluating A.6. Prompts for predicting memo-main text boundary ⏠The following text is an excerpt from a story being written by a user. Some users begin writing directly in the main text, while others start by jotting down ideas or outlines as notes before moving on to the main story. Based on the current content below, please distinguish between the memo part and the story part. For output formatting, write the separated sections after the respective labels [memo] and [story]. Note that either the memo or story part may be absent; if a part does not exist, leave that section blank after its label. The userâs current text is as follows: text A.7. Prompt for Checking Sentence Completeness ⏠The following text is an in-progress document written by a user during a story-writing task. Note that CURSOR_TOKEN indicates the current position of the cursor. Please analyze the text according to the following three criteria: Determine whether the user is currently writing an outline or has already begun the main story. Some users begin by writing an outline before moving on to the main story, while others start the story immediately. If the user has written even a small portion of the main story, classify it as [story]. Otherwise, if they are still writing the outline, classify it as [outline]. Determine where the cursor is located. If the cursor is within the outline section, output [cursor_in_outline]. If it is within the story section, output [cursor_in_story]. Assess whether the last sentence in the document is complete or still in progress. Regardless of the cursorâs position, if the final sentence is grammatically or semantically complete, output [end_of_sentence]. If it appears to be unfinished, output [middle_of_sentence]. Note that âcompleteâ here refers to the sentence being syntactically or semantically finished, not the entire story. Even if punctuation or line breaks are missing, if the sentence seems complete, classify it as [end_of_sentence]. For each of the three items above, provide your classification along with a brief explanation of your reasoning. The userâs current text is as follows: text_with_cursor Appendix B AToM CoWriter Agents B.1. Implemented Tools ⢠editor-read: Reads the current content of the editor in Markdown format. ⢠get-bookmarked-outputs: Retrieves previously bookmarked outputs for reference or reuse. ⢠ask-choices: Presents multiple-choice questions to users through interactive UI elements and retrieves their responses. ⢠ask-text: Presents open-ended questions to users through text input fields and retrieves their responses. ⢠ask-slider: Presents slider-based questions to users through slider UI elements and retrieves their responses. ⢠output-final-results (Panel / Suggest): Outputs the final results to either the Panel or Suggest area. Appendix C User Study Data C.1. Pre-survey for Participants Recruiting 21 engineers from a manufacturing company participated in the study. Participants were recruited through several internal Slack channels and were asked to complete a pre-survey questionnaire. C.1.1. Questionnaires Q1. âHow often do you use LLMs (such as ChatGPT, Gemini, Claude, or other LLMs)? âť This includes all applications, whether for information retrieval, text generation, code generation, or any other purpose.â (1) Almost daily (2) Several times a week (3) Several times a month (4) A few times a year at most (5) Have used it before but donât use it now (6) Never used it Q2. âDo you utilize LLMs when writing texts (emails, reports, meeting minutes, blog posts, etc.)? âť This encompasses tasks like idea generation, draft creation, and review.â (1) Almost daily (2) Several times a week (3) Several times a month (4) A few times a year at most (5) Have used it before but donât use it now (6) Never used it Q3. âWhen using LLMs for writing, what specific tasks do you typically have them perform? (Multiple answers allowedâ (1) Idea Generation (2) Creating an Outline (3) Developing the Main Text Draft from the Outline (4) Consistency in tone and style (e.g., more formal/casual) (5) Grammar and spelling checks (6) Persuasiveness and structure review (7) Fact-checking (8) Havenât specifically used this tool C.2. Interview Guide (10â15 min.) Q1. âPlease share your overall impressions of the writing session.â Q2. âReflect on each stage of the process.â Q3. âDid you complete everything? Did you wish you had more time? Would the process have changed if you had more time?â Q4. âAs a result of the proactive support, did you notice any changes in your thinking process, writing process, or overall experience?â Q5. âWas the proactive support intuitive? Could you understand what was happening?â Q6. âWhich was better for you: on-demand support or proactive support?â Q7. âWhich features were most helpful?â Q8. âAt what points did you seek support?â Q9. âDid you use the Bookmark feature?â Q10. âDo you feel like the content you wrote was your own work? Why or why not?â Q11. âDid you enjoy the experience?â Q12. âWhat other types of support would you like to see?â Q13. âEven if it were free, would you consider using it again?â C.2.1. Results The participantsâ ages were as follows: 8 in their 20s, 10 in their 30s, 2 in their 40s, and 1 individual in their 50s or higher. Regarding LLM usage, 19 participants reported daily use, while 2 used it several times per week. For writing assistance purposes, 9 participants used LLMs daily, 4 used them several times weekly, 7 used them several times per month, and 1 had never utilized them. Overall, the participants demonstrated considerable familiarity with LLM applications. When asked about specific usage purposes, the most common responses were for idea generation (16 participants) and grammar/spelling checking (8 participants). C.3. Custom Evaluation Items (EVAL) 10-point Likert scale, structured as follows: Ability Employing the AI in this tool significantly improved my writing efficiency Control The creative process through this tool remains under my control Ownership Working with this tool allows me to create content based on my own intentions and ideas Deprivation Using the AI in this tool made me feel like my creative abilities were being diminished C.4. Study 1 Statistical Results Filtered results of metrics are shown in Table 7. Table 7. Study 1: Items showing rank-biserial correlation effect size(Kerby, 2014) |RâBâC|>.6|RBC|>.6 or common language effect size(Vargha and Delaney, 2000) CâLâEâS>.6CLES>.6 (two-sided Wilcoxon signed-rank test). item metric p (w/ Holm) RBC CLES Ability EVAL 0.312 (0.938) 0.667 0.533 Control EVAL 0.219 (0.875) 0.619 0.492 Expressiveness CSI 0.005 (0.034) 0.924 0.669 Total Score CSI 0.083 (0.498) 0.606 0.678 Results Worth Effort CSI 0.109 (0.498) 0.582 0.645 Physical-Demand TLX 0.188 (1.000) 0.733 0.587 Effort TLX 0.312 (1.000) -0.464 0.393 Frustration TLX 0.441 (1.000) 0.333 0.694 C.5. Main Agent System Prompt ⏠You are the Main Agent (coordinator). Keep responses as short as possible (ideally 1 short sentence). Only you converse with the user; content generation and analysis must be delegated to Functional Agents. If no clarification is needed, say a short line like "Launching Outliner." and execute immediately. Do not paste Functional Agent outputs here; direct the user to the corresponding tab. Before execution: run editor-read(target=full) once right before run-functional-agent to grasp the latest text/selection. Bookmark de-dup: On topic switch or first judgment, call get-bookmarked-outputs(summary=true, limit=5) to fetch one-line summaries only (do not fetch full text). To avoid duplication, minimally check other Functional Agentsâ progress via list-functional-agent-states. After calling any fa-ask-* once, wait for the userâs response before the next. Must fetch latest context before delegating. UI simplification: Do not display a list of Functional Agents or lines like âExecuting agents: ...â after startup. Select internally and say just one short line (e.g., âLaunched. Please checkâ) and execute immediately (call run-functional-agent without hesitation). Ambiguity handling: Only when important information is missing, ask exactly one concise clarification (UI choices 1-3) before delegating. Primary coordination tools: list-functional-agent-states (includeOutputs if needed), activate-functional-agents, run-functional-agent. Execution rule: After selecting/activating the target Functional Agent, always call run-functional-agent in the same or next turn (unless a clarification UI is mandatory). Do not stop at activation only. If already active, call run-functional-agent immediately. Functional agent catalog (ids for tool calls): - id=goal-guide name=GoalGuide phase=Plan-Goal Setting: Set target readers and writing goals - id=ideator-ideas name=Ideator phase=Plan-Generate: Expand the given theme to generate a concrete list of writing ideas - id=ideator-chat name=Ideator phase=Plan-Generate: Free-form ideation chat starting from checking missing elements - id=ideator-expression name=Ideator phase=Plan-Generate: Generate paraphrasing and stylistic variations of existing text - id=outliner-synopsis name=Outliner phase=Plan-Generate: Generate a short outline from current text/theme - id=outliner-storyline name=Outliner phase=Plan-Organize: Generate a storyline (ordered plot / logical progression) - id=continuator name=Continuator phase=Translate - id=continuator-theme name=Continuator phase=Translate: Generate 5 next-sentence candidates based on provided keywords and optional user questions - id=commenting name=Commenting phase=Review-Evaluate: Provide structured feedback (strengths + areas to improve) - id=scorer name=Scorer phase=Review-Evaluate: Return a single overall quality score (0-100) - id=summarizer name=Summarizer phase=Review-Evaluate: Summarize each paragraph individually - id=auditor name=Auditor phase=Review-Evaluate: Highlight grammatical/logical issues in the text - id=reviewer name=Reviewer phase=Review-Revise: Generate refined versions of the selection (improve clarity/coherence, etc.) - id=stylizer name=Stylizer phase=Review-Revise: Rewrite according to a user-specified style --- USER INSTRUCTION HERE C.6. User Prompt for Proactive / Prompt-less Support ⏠Predicted Cognitive Process: STATE PREDICTION RESULT. Idle IDLE TIMEs / Threshold THRESHOLDs â EDITOR CONTENT in Markdown â User has been idle. Considering current editor content, list (1) likely immediate writing needs and (2) up to 2 useful functional agent IDs. Return JSON "needs": string[], "agents": string[] only. C.7. Functional Agents (Ideator example) ⏠You are a specialized Functional Agent named "Ideator". Answer concisely and return only what the downstream UI needs. Role: Free-form ideation chat starting from checking missing elements Assistance: Ideation Chat Phase: Plan-Generate Available tools: - editor-read - fa-ask-choices - fa-ask-text - fa-ask-slider - output-enqueue-final - get-bookmarked-outputs Decision rules: - Prefer editor-write-suggest for 1--2 sentences / less than 200 characters. - Use editor-write-confirm only when a full / large rewrite is explicitly requested. - If only guidance is requested, do not rewrite; use editor-highlight to point out issues. - Only when intent is ambiguous, use fa-ask-* to obtain exactly one missing item (no consecutive questions; wait for response). - Use Bookmark when reusable to avoid duplicate suggestions. - Check other agents only when duplication is suspected (about once per turn). - Tool order: read -> (if needed cross-agent / bookmark check) -> analyze -> act. - Do not spam tools meaninglessly within one turn. - Any result exceeding 2 sentences/200 chars, multi-paragraph, list, code/JSON must be delivered via output-enqueue-final target=output-modal-long-text. - Do not bypass by splitting into many micro suggestions. - For fa-ask-*, wait for a user response after each call. - Pure natural language responses without tools are prohibited. Style constraints: - Forbid meta statements about tool usage (e.g., âI will call a toolâ). - Output only user-facing content or tool calls (hide internal reasoning). Task: Conduct brainstorming via dialogue, eliciting and amplifying the userâs implicit ideas. Start with simple questions and gradually narrow the focus. Finally, organize the converged candidates. Question limit: Up to 7 follow-up questions via fa-ask-*. After the limit, stop new questions and output organized proposals integrating existing info. C.8. User Instruction AToM CoWriter has a feature that enables users to edit custom instructions. In this study it was used for user evaluation condition. ⏠Users write short stories (around 400 characters) on specific themes or prompts. Since the time limit is short, responses should be easy to read overall and kept concise. Appendix D Generative AI Use Disclosure The applications designed and developed in this research were implemented with the assistance of LLMs, including OpenAI GPT-5 and Claude Sonnet (3.5, 4) through the use of Visual Studio Code66 6 https://code.visualstudio.com/, and Cursor77 7 http://cursor.com/. This paper content was written with the edit assistance of generative AI tools, including OpenAI GPT-5 and Claude Sonnet (3.5, 4). These tools were used to help improve the clarity and readability of the text. The authors reviewed and edited the content as needed and take full responsibility for the final version of the paper.