Designing Proactive Thought Partners for Writing
A one-week probe with 16 writers found users configure proactive AI writing partners prospectively, mostly ignore their suggestions to preserve flow, and prefer lightweight, non-directive framing.
Shows how to design proactive AI assistance for higher-level cognitive work, not just autocomplete, based on real usage data rather than intuition.
Chao Zhang · Abe Davis · Chih-Wei Chen · Chin-Chia Hsu — meet the researchers →
Two readings, equal authority
How to choose: The paper’s words is verbatim — use it when you need to quote, or to judge how they write. Plain language is a paraphrase written for comprehension — use it when you want the idea fast. Neither is a summary of the other; they are two doors into the same room.
“Writing involves diverse cognitive activities, from ideation to revision, and writers’ needs vary across individuals and moments. Proactive AI promises to provide the right support at the right time, yet existing proactive tools largely focus on generic textual assistance, such as autocomplete. This paper studies the design space of proactive thought partners: AI agents that proactively offer customizable, higher-level cognitive support during writing. We instantiated this concept in a technology probe and deployed it with 16 participants for one week. The probe allows users to create partners by configuring their roles and proactivity. As users write, relevant partners take the initiative at appropriate moments to offer suggestions. Our findings show that participants configured proactive support through prospective planning, used suggestions for both idea generation and self-monitoring, and valued lightweight visual representations alongside non-directive rhetorical framing for non-intrusive interventions. We derive implications for designing proactive writing assistants around customization, timing, engagement, and representation.”
The researchers built a writing tool where users create customizable AI 'thought partners' that jump in proactively with suggestions while someone writes, instead of waiting to be asked. After deploying it for a week with 16 real users, they found people set the partners up in advance based on anticipated needs, usually ignored suggestions to keep their writing momentum, and preferred help that appeared subtly and was phrased as an open question rather than a command.
What this paper defines
Every definition below is the paper’s own sentence, with its locator. The plain gloss is a reading aid and is marked as one.
Proactive thought partners
“We study proactive AI assistants that support these higher-level, personal cognitive needs in writing, which we call proactive thought partners (Fig. 1).”1. Introduction
In plain terms: AI agents that proactively offer higher-level cognitive support, not just autocomplete, while someone writes.
Proactivity
“we define proactivity as the system actively inferring user needs and offering suggestions without requiring explicit user prompting.”2.2. Proactive AI Assistants
In plain terms: The system decides on its own, without being asked, when to offer help.
Event Trigger
“Event Trigger(s) define an observable user action that may indicate an opportunity for proactive support, such as pausing, completing a sentence, or selecting text.”4.1. Partner Customization
In plain terms: An observable writer action, like pausing, that flags a candidate moment for the AI to consider stepping in.
Contextual Heuristic
“Contextual Heuristic defines the contextual criteria under which the partner should take the initiative following a triggering event.”4.1. Partner Customization
In plain terms: A user-defined rule for what must be true about the writing context before a partner actually intervenes after a trigger fires.
Technology probe
“an instrument deployed to investigate an open-ended design space and return useful or interesting data about users’ practices, needs, and values (Hutchinson et al., 2003).”3. Design Goals
In plain terms: A deliberately open-ended prototype deployed to learn about user practices and needs, not to prove a finished product works.
Substantive session
“We defined a substantive session as one that met two criteria: (1) at least 20 minutes of active writing or production of at least 300 words; and (2) completion of a self-contained piece of writing, rather than randomly connected sentences or artificial text written only for the study.”5.2. Procedure
In plain terms: The study's own definition of a writing session that counts as real, sufficient engagement with the probe.
What they actually did
Each step is a synthesis. Open any step to see the paper’s own sentence it was derived from, with its locator — so nothing here floats free of the source.
- Built the probe as three linked interfaces: an onboarding panel, a partner-configuration panel, and a writing editor.
Trace this step to the paper
“The probe consists of three main interfaces. First, an onboarding panel allows users to log in and describe their writing goals for a writing session. Second, a partner configuration panel allows users to create, configure, and enable proactive thought partners (Fig. 3). Third, a lightweight Markdown editor allows users to write with their enabled partners (Fig. 2).”4. Proactive Thought Partners: A Technology Probe
- Let users define each partner by a role, one or more event triggers, and a contextual heuristic.
Trace this step to the paper
“In the configuration panel, users can create a proactive thought partner by specifying a name and emoji, a role, one or more event triggers, and a contextual heuristic.”4.1. Partner Customization
- Implemented three rule-based event triggers (long pause, sentence end, text selection) that monitor keystrokes in real time.
Trace this step to the paper
“First, we implemented three event triggers, each grounded in a different rationale from prior work on proactive AI... These triggers are rule-based and operate on user keystrokes monitored in real time within the editor.”4.2. Partner Activation
- On a trigger, sent session goal, current text, and recent writing behavior to an LLM decision engine that activates up to two partners whose heuristics match.
Trace this step to the paper
“When an event trigger occurs, the system sends the writer’s session goal, the current text in the editor, the recent writing behaviors, and the list of enabled partners associated with that event trigger to an LLM-based decision engine (Fig. 4)... the decision engine selects at most two enabled partners, if any, to intervene at that moment by checking whether their user-defined contextual heuristics are satisfied.”4.2. Partner Activation
- Had each activated partner generate a two-part suggestion: an acknowledgment of the writer's activity, then a question-framed suggestion.
Trace this step to the paper
“Each suggestion contains two components... the partner offers a thought-provoking suggestion tailored to the current context. We frame this suggestion as a question because prior work suggests that question-style AI suggestions can stimulate high-quality ideas while preserving users’ ownership.”4.3. Suggestion Generation
- Gave users three graduated ways to engage with a suggestion: ignore it, open it for inspiration/discussion, or have the partner execute it directly in the document.
Trace this step to the paper
“Once the floating tags of activated partners appear, users can choose whether and how to engage with the proactive interventions in three ways (DG2).”4.4. Engagement Forms
- Implemented the probe in Next.js against the Gemini API, using gemini-2.5-flash-lite for the low-latency decision engine and gemini-2.5-flash for suggestion generation, discussion, and text editing.
Trace this step to the paper
“For the decision engine that selects which partners should intervene, we use gemini-2.5-flash-lite to minimize latency... For other LLM-based tasks, including suggestion generation, follow-up discussion, and text insertion or revision, we use gemini-2.5-flash.”4.5. Implementation
- Deployed the probe with 16 participants in a one-week study to answer three research questions about partner creation, suggestion engagement, and user experience.
Trace this step to the paper
“We conducted a one-week technology probe study with 16 participants to explore how writers customize, experience, and engage with proactive thought partners in writing practice.”5. User Study
- Ran a 60-minute onboarding with a tutorial, four preconfigured partners, and a warm-up essay task before the diary phase began.
Trace this step to the paper
“After tutorial, they completed a warm-up writing task, in which they wrote a short essay of approximately 300 words on the positive and negative aspects of AI in daily life.”5.2. Procedure
- During the 7-day diary phase, required at least four substantive sessions and collected a structured diary entry (rating scales plus open-ended prompts) after each session.
Trace this step to the paper
“we asked participants to complete at least four substantive writing sessions during the week to ensure sufficient exposure to the probe... participants submitted a structured diary entry at the end of each writing session. The diary included three 10-point scales measuring satisfaction, perceived timeliness, and perceived helpfulness of the proactive support, as well as four open-ended prompts”5.2. Procedure
- Closed with a UMUX-LITE usability questionnaire and a semi-structured exit interview comparing the probe to reactive chatbots and predictive-text tools.
Trace this step to the paper
“In the final phase, participants first completed the short-form Usability Metric for User Experience (UMUX-LITE) (Lewis et al., 2013). We then conducted a semi-structured interview to probe participants’ week-long experiences and elicit their visions for future proactive thought partners for writing.”5.2. Procedure
- Had two researchers thematically code interview and diary text into a shared codebook, then group codes into higher-level themes, triangulated against interaction logs.
Trace this step to the paper
“Two researchers conducted a thematic analysis of the qualitative data (Braun and Clarke, 2013). They reviewed the interview transcripts and diary entries to identify recurring concepts, then met to compare, discuss, and consolidate the codes into a codebook.”5.3. Analysis
Exactly what was run, and how
| Model | Developer | Temp | Effort / reasoning | Deployment | Other settings |
|---|---|---|---|---|---|
| gemini-2.5-flash-lite | not reported | not reported | API | — | |
| gemini-2.5-flash | not reported | not reported | API | — |
Source for gemini-2.5-flash-lite settings
“For the decision engine that selects which partners should intervene, we use gemini-2.5-flash-lite to minimize latency.”4.5. Implementation
Source for gemini-2.5-flash settings
“For other LLM-based tasks, including suggestion generation, follow-up discussion, and text insertion or revision, we use gemini-2.5-flash.”4.5. Implementation
What they reported — and what they left out
The paper names two Gemini models mapped to distinct pipeline roles (gemini-2.5-flash-lite for the low-latency decision engine, gemini-2.5-flash for suggestion generation, discussion, and editing) and says prompts are in the supplementary materials, but the main text reports no temperature, sampling parameters, context window, or system-prompt content.
The numbers they report
Participants completed roughly four sessions each over the week.
66 total sessions across 16 participants; M = 4.13 sessions/participant, SD = 0.50
See it in the paper
“Across the 16 participants, we collected 66 writing sessions, excluding onboarding sessions. Participants completed an average of 4.13 (𝑆𝐷 = 0.50) sessions each.”6.1. Probe Usage and Usability
Sessions averaged about 390 words and 17 minutes.
M = 390.34 words, SD = 151.41; M = 17.09 minutes, SD = 7.47
See it in the paper
“Each session produced an average of 390.34 (𝑆𝐷 = 151.41) words and lasted an average of 17.09 (𝑆𝐷 = 7.47) minutes.”6.1. Probe Usage and Usability
Sessions spanned personal and academic writing most, journalistic writing least.
personal N=19, academic N=19, professional N=9, creative N=8, technical N=8, journalistic N=3
See it in the paper
“these sessions included personal writing ( 𝑁 = 19), academic writing ( 𝑁 = 19), professional writing ( 𝑁 = 9), creative writing ( 𝑁 = 8), technical writing ( 𝑁 = 8), and journalistic writing ( 𝑁 = 3).”6.1. Probe Usage and Usability
The probe scored in the 'Excellent' usability range.
SUS (via UMUX-LITE) = 82.81, SD = 11.97, ~90th-95th percentile
See it in the paper
“The System Usability Scale (SUS) scores calculated from UMUX-LITE collected during the exit interviews averaged 82.81 (𝑆𝐷 = 11.97). This score corresponds to the “Excellent” category defined by Bangor et al. (2008, 2009), placing it in approximately the 90th-95th percentile.”6.1. Probe Usage and Usability
Daily satisfaction with the probe was high.
8.16/10, SD = 1.76
See it in the paper
“This is consistent with daily survey measures of overall satisfaction, which averaged 8.16/10 (𝑆𝐷 = 1.76).”6.1. Probe Usage and Usability
Satisfaction rose significantly across the week as participants gained experience.
b = 0.33, SE = 0.16, p = .036 (linear mixed-effects model, participant as random effect)
See it in the paper
“A linear mixed-effects model with participant as a random effect showed that satisfaction increased significantly with session index (𝑏 = 0.33, 𝑆𝐸 = 0.16, 𝑝 = .036)”6.1. Probe Usage and Usability
Participants created a few partners each, most used only once.
54 total partners; M = 3.38/participant, SD = 1.96; 35 (64.81%) used in one session only, 19 (35.19%) used across multiple sessions
See it in the paper
“Participants created 54 partners, with an average of 3.38 (𝑆𝐷 = 1.96) partners created per participant. Among these, 35 (64.81%) were enabled in only one writing session, while 19 (35.19%) were used across multiple sessions.”6.2. RQ1: Partner Creation
Most partners were built for higher-level cognitive support rather than local textual assistance, concentrated in a few role types.
43/54 (79.63%) higher-level cognitive support vs. 11/54 (20.37%) local textual assistance; information seeking 14 (25.93%), argument development 12 (22.22%), critical reflection 10 (18.52%), ideation support 7 (12.96%)
See it in the paper
“Participants configured 43 of 54 partners (79.63%) for higher-level cognitive support and 11 (20.37%) for local textual assistance such as language and style refinement... 14 partners (25.93%) focused on seeking evidence or domain knowledge (information seeking), 12 (22.22%) on argument development and reasoning (argument development), 10 (18.52%) on perspective-taking and critical reflection (critical reflection), and 7 (12.96%) on ideation and creative development (ideation support).”6.2. RQ1: Partner Creation
Event triggers were configured broadly, with most partners enabled on all three triggers.
all three triggers: 31/54 (57.41%); pause 40/54 (74.07%); sentence end 40/54 (74.07%); text selection 44/54 (81.48%)
See it in the paper
“event triggers were generally configured broadly: 31 of 54 partners (57.41%) were enabled with all three triggers. Specifically, pauses were enabled for 40 of 54 partners (74.07%), sentence endings for 40 partners (74.07%), and text selections for 44 partners (81.48%).”6.2. RQ1: Partner Creation
Contextual heuristics clustered into three overlapping cue types (draft-state, writing-activity, anticipated-need).
draft-state 35/54 (64.81%); writing-activity 38/54 (70.37%); anticipated-need 36/54 (66.67%)
See it in the paper
“draft-state cues described what was currently present in the text... 35/54 (64.81%). Second, writing-activity cues described what the writer was currently doing in the writing process... 38/54 (70.37%). Third, anticipated-need cues described the support the writer might need at that moment or next... 36/54 (66.67%).”6.2. RQ1: Partner Creation
Fewer than half of surfaced interventions were opened.
1,100 total interventions; 459/1,100 (41.73%) opened
See it in the paper
“The probe provided 1,100 proactive interventions across all participants and their writing sessions (Table 3). Among these, participants opened 41.73% of the surfaced interventions (459/1,100).”6.3. RQ2: Suggestion Engagement
Ignoring was the dominant response, especially after pause-triggered interventions.
ignored overall 641/1,100 (58.27%); pause-triggered ignore rate 64.39%
See it in the paper
“the most common response was to ignore the displayed tag and continue writing without clicking to view the suggestion, accounting for 58.27% of interventions (641/1,100)... Interventions surfaced after a pause were especially likely to be ignored (64.39%)”6.3. RQ2: Suggestion Engagement
Execution was less common than ignoring but had a high acceptance rate, and was most likely after text-selection triggers.
244/1,100 (22.18%) executed; 206/244 (84.43%) of generated text accepted; execute rate by trigger: selection 54.48%, sentence end 28.81%, pause 20.61%
See it in the paper
“Executing represented a more committed form of engagement and accounted for 22.18% of interventions (244/1,100), with 84.43% (206/244) of the generated text accepted. Interventions triggered by text selection were more likely to lead to executing (54.48%) than those triggered by a pause (20.61%) or sentence end (28.81%).”6.3. RQ2: Suggestion Engagement
Participants rated the proactive support as both helpful and reasonably well-timed.
helpfulness 8.11/10, SD = 1.87; timeliness 7.85/10, SD = 1.86
See it in the paper
“participants gave an average rating of 8.11/10 (𝑆𝐷 = 1.87) for helpfulness and 7.85/10 (𝑆𝐷 = 1.86) for timeliness.”6.4. RQ3: User Experience
The sample skewed toward experienced, non-native-English, female participants in their late 30s.
ages 26-53, M = 37.47, SD = 8.57 (one undisclosed); 11 female, 5 male; 5 native English speakers, 11 non-native speakers
See it in the paper
“Participants ranged in age from 26 to 53 years ( 𝑀 = 37.47, 𝑆𝐷 = 8.57; one participant preferred not to disclose their age), with 11 identifying as female and 5 as male... 5 were native speakers and 11 were non-native speakers.”5.1. Participants
What they assert, beside what they showed
Left is the claim in the paper’s own words. Right is the data offered for it. Where the two do not fully meet, a gold band names the distance.
Proactivity shifted writing support from a user-initiated 'pull' model to a system-initiated 'push' model that felt more continuous than chatbot use.
“This shift came from the context-aware “push mechanism” afforded by proactivity”
“it’s a push mechanism rather than a pull mechanism. . . I don’t need to again click and put the information in a different thing and ask. It’s sort of automated on what I want to ask since it’s a partner.”
6.4. RQ3: User Experience (KF7)Participants configured proactive partners as a form of prospective planning, anticipating needs before writing began.
“Participants configured proactive thought partners through a form of prospective planning.”
“I try to envision what the end product would look like and then, if I had a human writer, I imagine what that person’s expertise should be.”
6.2. RQ1: Partner Creation (KF1)Ignoring interventions functioned to preserve writers' flow, and this was especially true after pause-triggered interventions.
“Ignoring preserved writing flow and occurred more often after pauses”
“the most common response was to ignore the displayed tag and continue writing without clicking to view the suggestion, accounting for 58.27% of interventions (641/1,100). Participants described keep typing as a natural, low-cost behavior to dismiss the interventions.”
6.3. RQ2: Suggestion Engagement (KF5)Peripheral placement and fading visual presentation made the proactive suggestions feel non-intrusive.
“Peripheral and fading visual representations made proactive support feel non-intrusive”
“come on the right tab and not on the main writing screen. . . they’re very small which means that they don’t take up a lot of your space even on the right side.”
6.4. RQ3: User Experience (KF9)The probe's UMUX-LITE-derived SUS score places it in the 'Excellent' usability category, roughly the 90th-95th percentile.
“This score corresponds to the “Excellent” category defined by Bangor et al. (2008, 2009), placing it in approximately the 90th-95th percentile.”
“The System Usability Scale (SUS) scores calculated from UMUX-LITE collected during the exit interviews averaged 82.81 (𝑆𝐷 = 11.97).”
6.1. Probe Usage and UsabilityExecuting a suggestion (having the AI write directly) was selective and depended on writers already being ready to externalize a settled intention.
“Executing was selective and depended on readiness to externalize text”
“Interventions triggered by text selection were more likely to lead to executing (54.48%) than those triggered by a pause (20.61%) or sentence end (28.81%).”
6.3. RQ2: Suggestion Engagement (KF6)The study's design implications (DI1-DI9) generalize as guidance for future proactive writing systems broadly.
“[DI3] Treat observable interaction events as candidate moments for intervention rather than sufficient evidence of need.”
“the same event could correspond to different writing states (KF3, KF8). A pause might indicate writer’s block, focused reflection, or a moment when the writer already knows what to write next”
7.2. Designing Timing as Contextual AlignmentHow they frame it, and what they want next
Their framing
The authors repeatedly frame their probe as exploratory and non-evaluative, meant to open up a design space rather than prove a finished product's effectiveness, and they caveat that findings describe how people appropriated a nascent technology rather than causal effects on writing outcomes. They present their four-dimension design-implications table (Table 4: customization, timing, engagement, representation) as forward-looking guidance distilled from what participants said and did during the deployment.
Register: Confident and declarative about descriptive findings (usage statistics, coded themes, participant quotes) but consistently hedged about causal or generalizable claims, closing with a dedicated Limitations section that explicitly walks back scope on sample diversity, causality, and system-specificity.
Where they hedge
“It is important to note that a technology probe study is not an evaluative study of a finished product’s efficacy.”5. User Study
“our findings do not establish causal effects on writing quality, productivity, learning, or long-term agency.”9. Limitations
“current LLMs may not reliably interpret writing context well enough to infer writers’ momentary intentions or anticipated support needs.”9. Limitations
“suggestion quality and timing depended on a particular LLM-based implementation. Different models, prompts, contextual representations, or orchestration strategies may produce different intervention patterns and user experiences.”9. Limitations
What they say it means
- Future systems should help writers translate anticipated goals and difficulties into partner configurations, i.e. support prospective planning.
the paper’s words
“[DI1] Support prospective planning by helping writers anticipate the goals, difficulties, and forms of assistance relevant to an upcoming writing task, and translate these expectations into partner configurations.”Table 4 / 7.1
- Systems should treat events like pauses as candidate moments only, adding a layer of contextual inference before deciding whether to intervene.
the paper’s words
“[DI4] Infer intervention opportunities from multiple contextual cues that capture what is present in the draft, what the writer is currently doing, and what support they may need next.”Table 4 / 7.2
- Ignoring a proactive suggestion should be effortless and require no explicit dismissal action.
the paper’s words
“[DI5] Make ignoring a first-class response to system initiative. Proactive suggestions should be easy to bypass without requiring explicit dismissal or disrupting the writing flow.”Table 4 / 7.3
- Direct text execution should be a selective escalation reserved for settled intentions, not the default response for reflection-oriented suggestions.
the paper’s words
“[DI7] Reserve direct text generation for moments when writers are ready to externalize a sufficiently settled intention.”Table 4 / 7.3
- Proactive suggestions should be represented as lightweight, peripheral, and non-directive so they can recede when unwanted.
the paper’s words
“[DI8] Represent proactive suggestions as lightweight, peripheral bids for attention. Suggestions should remain noticeable without occupying the primary writing space and should recede when writers choose not to engage.”Table 4 / 7.4
What they call for next
- Calls for privacy-by-design practices given the continuous behavioral monitoring proactive writing support requires.
the paper’s words
“Future systems should therefore adopt privacy-by-design practices (Spiekermann, 2012), including data minimization, local or ephemeral processing of behavioral traces, strict retention limits, clear disclosure of what is monitored, and granular controls for pausing monitoring or deleting collected data.”8. Ethical Considerations
- Calls for longer or larger comparative studies against reactive AI assistants and conventional tools to establish causal effects.
the paper’s words
“Longer deployments or a larger scale of comparative, quantitative lab studies against reactive AI assistants and conventional writing tools can be valuable in the future.”9. Limitations
- Calls for models and benchmarks that test whether AI can recognize the paper's three contextual cue types and pick appropriate support.
the paper’s words
“Future research could develop models and benchmarks that evaluate whether AI systems can recognize these cues, distinguish among different writing states that produce similar interaction events, and determine which forms of support are appropriate in context.”9. Limitations
- Calls for evaluating proactive thought partners with more demographically and linguistically diverse participants across broader, higher-stakes writing contexts.
the paper’s words
“Future work could address these limitations by evaluating proactive thought partners with a more demographically and linguistically diverse participant population, including writers with varying levels of AI experience, and across a broader range of professional and high-stakes writing contexts.”9. Limitations
Limitations they state
“our findings do not establish causal effects on writing quality, productivity, learning, or long-term agency.”9. Limitations
“our study can only offer preliminary insight into how proactive support may influence writers’ sense of ownership.”9. Limitations
“their expectations, comfort with AI, and willingness to engage with proactive support may differ from those of writers with less AI experience or different linguistic backgrounds.”9. Limitations
“our privacy instruction to avoid confidential workplace content likely excluded the high-stakes writing these participants do most.”9. Limitations
“current LLMs may not reliably interpret writing context well enough to infer writers’ momentary intentions or anticipated support needs.”9. Limitations
“suggestion quality and timing depended on a particular LLM-based implementation. Different models, prompts, contextual representations, or orchestration strategies may produce different intervention patterns and user experiences.”9. Limitations
Moves worth stealing
Opens by naming a specific gap between two existing literatures (proactive-but-shallow vs. cognitive-but-reactive) and coins a term to occupy exactly that gap, which the abstract and intro then repeat verbatim to anchor the contribution.
“Our work explores the design space in this gap that combines proactivity and cognitive support in writing, which we term “proactive thought partners.””
Uses numbered Design Goals (DG1-DG3) cross-referenced throughout the system description, then closes the loop with a matching numbered Design Implications table (DI1-DI9) — giving a traceable thread from goals to features to findings to implications.
“Specifically, we designed the probe around three goals. [DG1] Support user-configurable proactive partners:”
Grounds nearly every design decision in a one-line rationale tied to prior work before describing the implementation, so each feature reads as motivated rather than arbitrary.
“First, we implemented three event triggers, each grounded in a different rationale from prior work on proactive AI”
Structures Findings around bolded, numbered Key Findings (KF1-KF10) nested under research questions, each closed with an italicized Takeaway paragraph a reader could skim in isolation.
“RQ1 Takeaway Participants configured proactive thought partners through a form of prospective planning.”
Substantiates qualitative claims by quoting multiple different named participants making the same point in their own words, rather than relying on a single anecdote.
“P10 noted that suggestions “come on the right tab and not on the main writing screen... they’re very small” P16 similarly explained, “it’s in the set panel right so it’s not blocked my main””
Includes a freestanding Ethical Considerations section, separate from Limitations, that anticipates privacy objections to its own continuous-monitoring data-collection method.
“Proactive writing support raises privacy concerns because providing timely assistance requires continuous observation of both the evolving document and the writer’s interaction traces.”
Where else this leads
Published alongside it
The nearest publications in time, across all three labs.
- Formalizing Fermat's Last Theorem Anthropic
2026-09-04 - Research acceleration: The view inside OpenAI OpenAI
2026-09-06 - Visual General Intelligence: A White Paper Google DeepMind
2026-08-26 - How Claude is accelerating protein design and analytical chemistry Anthropic
2026-08-18
What this page was built from
Extracted from the full paper text (Sections 1-10 plus references, arXiv-formatted); the paper states its prompts are 'provided in the supplementary materials' but no supplementary/appendix content was present in this text file, and an arXiv identifier (2609.01588) appears in the header though no arXiv URL string is given in the text itself.