What clues — screen activity, click history, even guesses about your mindset — let an AI predict what you want next?
What input signals do intervention-aware models use to forecast user desires?
This explores what kinds of evidence AI systems watch, such as screens, clicks, activity histories, or inferred mental states, so they can predict what a user wants or when the user will step in to correct an agent.
This explores what evidence AI systems use to predict what a user wants, including the moment a user will take over from an agent. The most direct paper in the collection is about web agents that learn to predict when a person will intervene. Building those predictions into a live agent made users rate it 36.8% more useful Can models predict when users want to intervene in web agents?. The summary available here doesn't say exactly which signals those models read. It gives the result, not the input list. So the more useful answer comes from setting that work beside nearby research that is more explicit about what it watches.
The closest relative is passive screen observation. One system watches what is on a user's screen and infers narrow, short-term goals: what the person is trying to do in the next few minutes, not what they want in general. People preferred its outputs over a standard LLM's 66–86% of the time Can systems infer what users want from watching their screen?. Predicting when a user will intervene is like the flip side of this. One system infers what you want right now, and the other infers when the agent has stopped giving it to you. Both treat ongoing behavior as the signal, not stated instructions.
Zoom out in time and the signal becomes the activity log. LLMs reading long histories of user activity can find 'interest journeys' that last more than a month, described in specific phrases like 'designing hydroponic systems for small spaces.' Traditional recommenders, which only notice that people who liked X also liked Y, miss these entirely Can language models discover what users actually want from activity logs?. Zoom out in sensory range and you get agents that watch continuous video and audio. They build memory graphs organized around people and objects, keeping specific events separate from general facts, so they can learn preferences without ever asking Can agents learn preferences by watching rather than asking?.
The less obvious idea is that tracking behavior alone may not be enough. Research on world models found that predictions of human actions went wrong even when the physical scene was modeled correctly. Accuracy needed explicit, coupled representations of the person's beliefs, wants, and intentions Can world models predict human action from physics alone?. That suggests the best forecasting systems may need an internal model of what the user is thinking, not just a record of what they did. There's a caution, though. Benedict Evans argues that LLMs let platforms 'rent' this kind of understanding of why users act, but research cited alongside that argument finds LLMs make up 35–49% of their claims about user attributes Can LLMs infer user needs better than owned behavioral data?.
The twist to keep in mind is that forecasting desire isn't purely passive. When agents act on their predictions, especially agents that complete purchases, they start shaping the preferences they claim to read. Some argue that preferences then form through back-and-forth between agent and user rather than existing beforehand Do AI shopping agents reshape what consumers actually want?. So a user stepping in to correct an agent may be the most honest signal available: it's the moment the human pushes back against a loop the agent is co-writing.
Sources 7 notes
Fine-tuned models that predict user intervention moments and integrate those predictions into live web agents achieved a 36.8% improvement in user-rated usefulness during a real-world task study, with 61.4–63.4% better intervention prediction accuracy than baselines.
Systems that passively observe user activity to infer just-in-time objectives produce outputs preferred 66–86% of the time over standard LLMs. This approach sidesteps fixed training objectives by inferring narrow, momentary goals from behavior instead.
66% of users pursue valued interest journeys lasting over a month, described in specific phrases like 'designing hydroponic systems for small spaces.' LLM-powered journey discovery bridges the semantic gap that collaborative filtering cannot reach, operating at user-level granularity with persona-level precision.
M3-Agent demonstrates that separating episodic events from semantic knowledge in an entity-centric graph, combined with parallel memorization and control processes, allows agents to infer and act on user preferences without asking. This architecture mirrors human cognitive systems that bind disparate information about individuals across sensory modalities.
Research across eight LLM-based world models shows that tracking only the physical scene leads to wrong action predictions even when the scene looks correct. Mental World Modeling makes beliefs, wants, and intentions explicit state components coupled to physical simulation, and all three elements are required for accurate human decision prediction.
Show all 7 sources
Benedict Evans contends that LLMs can infer deeper user motivations (the "why") than correlation-based recommenders, allowing platforms to rent this capability via API rather than accumulating their own behavioral data. However, research shows LLMs fabricate 35–49% of user attribute claims, undermining confidence in their inferred understanding.
Bilgihan and colleagues argue that transactional AI agents—which execute purchases rather than just recommend—end the consumer's sole cognitive role in preference formation. Preferences emerge through bidirectional feedback loops between agent learning and consumer trust, not independently beforehand.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- User-LLM: Efficient LLM Contextualization with User Embeddings
- The Personalization Mirage: How LLMs Fabricate User Profiles, and Why Self-Monitoring Misleads
- When Are LLM Inferences Acceptable? User Reactions and Control Preferences for Inferred Personal Information
- A Framework for Studying AI Agent Behavior: Evidence from Consumer Choice Experiments
- Shopping by algorithm: How agentic AI deploys human heuristics as a surrogate consumer
- Designing a Dashboard for Transparency and Control of Conversational AI
- Mental World Modeling
- Large Language Models for User Interest Journeys