Should an AI assistant act more or less pushy depending on the task — or even moment to moment?
Do different task types require different proactivity tuning needs?
This explores whether an AI assistant's ideal level of initiative (how often it jumps in, suggests, or acts without being asked) should change with the kind of task, rather than being one fixed setting.
This explores whether the right amount of AI initiative depends on what you're doing, or whether one proactivity setting can serve every task. The short answer: the collection has only one study aimed squarely at this question. What it shows is useful, though, and it points somewhere unexpected. The right level seems to vary within a single task as it unfolds, not only between task types.
The most direct evidence comes from two Slack studies of a group brainstorming bot Do users want to adjust AI proactivity during conversations?. People didn't settle on one preferred level of involvement. Mid-session, they told the bot to contribute less when it crowded the conversation, and pushed it to contribute more when it hung back. Earlier frameworks treated proactivity as a dial you set once, or as a simple on/off switch. These users treated it as something to adjust as the work went on. One way to read this is that 'task type' is too coarse a unit. Even inside brainstorming, the early diverge-everything phase and the later narrowing phase seem to call for different amounts of initiative.
The collection has no direct comparison of proactivity across task types. It does have a strong lateral pattern: the same tuning can push behavior in opposite directions depending on the domain. Preference tuning (RLHF) makes code output less varied but creative writing more varied, because code rewards converging on a correct answer and creative work rewards standing out Does preference tuning always reduce diversity the same way?. Multi-task training shows a similar split. Structured tasks make the model's outputs more predictable, open-ended tasks make them less so, and the order of training matters enough to change results by about 6% Does training order reshape how models handle different task types?. If adjustments that simple behave so differently across domains, a single proactivity setting is unlikely to work equally well for a coding task and a brainstorm. That's an inference, not a tested result.
Agent research hints at a second dimension: proactivity isn't one skill. MyPhoneBench found that finishing the task, protecting privacy along the way, and reusing a user's saved preferences without being asked are statistically separate abilities, and no model led on all three Do phone agents succeed at all three critical tasks equally?. That third ability is a kind of proactivity: acting on what the user already told you. It doesn't come for free with being good at tasks. Tasks like leases, insurance and finance depend heavily on unstated requirements, so there the useful kind of initiative is pointing out what the user didn't think to ask. Meta claims its Muse Wonder agent handles this kind of work, but offers no evidence Does Meta's Muse Wonder truly handle long tasks better?.
The surprise is that the collection's best evidence argues for letting users steer proactivity in the moment, not for presets per task type. It also suggests 'proactive' covers several distinct behaviors that should be measured separately: contributing ideas, recalling preferences, and flagging hidden requirements. A controlled study comparing proactivity needs across task types isn't in the collection yet.
Sources 5 notes
Two Slack studies of a brainstorming bot found participants asked the agent to reduce contribution frequency mid-session and pushed it to contribute more when reactive. This contradicts prior frameworks treating proactivity as binary or static.
RLHF reduces lexical-syntactic diversity in code generation but increases it in creative writing. The direction depends on what each domain incentivizes: code rewards convergence toward correct solutions, while creative writing rewards stylistic distinctiveness.
Omni-Thinker shows structured domains decrease output entropy while creative domains increase it. BWT-guided scheduling—training structured tasks first—yields 6.2% gains over joint training by preventing entropy collapse from damaging open-ended capabilities.
MyPhoneBench demonstrates that task success, privacy-compliant completion, and saved-preference reuse are statistically distinct capabilities with no model dominating all three. Success-only rankings do not predict privacy or preference performance.
Meta's announcement of Muse Wonder presents Active Memory and Active Focus as major agentic capabilities but provides no benchmarks, user studies, or third-party validation. The claimed domains (leases, insurance, finance) are areas where frontier models demonstrably struggle with implicit requirements—a gap the announcement does not measure or address.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Evaluating the Diversity and Quality of LLM Generated Content
- Do Phone-Use Agents Respect Your Privacy?
- Controlling AI Agent Participation in Group Conversations: A Human-Centered Approach
- Introducing Muse Wonder and Muse Wonder Full: Frontier Models for Agents
- Omni-Thinker: Scaling Multi-Task RL in LLMs with Hybrid Reward and Task Scheduling
- Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
- DiscussLLM: Teaching Large Language Models When to Speak
- HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants