Smart glasses, MR headsets, and ambient copilots are turning the AI assistant from something you query into something that is simply present. Being present raises a product question that did not exist before: when should the assistant speak first? Too early and it interrupts. Too late and it adds nothing. Never, and it may as well not be there.
A request-response assistant never has to answer this. The user's question is itself the permission to speak, and the design problem reduces to answering well. An always-on assistant has no such signal. It has to generate the permission internally, on every timestep, from whatever it can perceive about what the user is doing.
Update, 20 September 2026. The survey has been accepted by Transactions on Machine Learning Research. The accepted version adds a machine-scored pilot (§8.4), and the section on evidence below has been revised to match. There is also a six-minute video summary.
What the paper says
The survey organizes the whole problem around one decision rule: act only when the expected benefit of acting, minus the expected cost of interrupting, exceeds a threshold. Written out, that is intervene <=> E[benefit] - E[cost of interrupting] > theta.
The rule is not new. Horvitz's 1999 alerting model states it in essentially this form. The survey's claim is about what happened to the two terms since. By 2026 the benefit term has grown enormously, thanks to better models and better perception, while the cost term has all but vanished from modern implementations. A rule with one term estimated carefully and the other treated as a constant is not really a cost-benefit rule; it is a confidence threshold wearing a cost-benefit costume.
That asymmetry produced two literatures that do not cite each other.
The first runs from 1999 to 2022 and quantified the cost of interruption in real detail. Horvitz modeled it as an expectation over an inferred attentional state. Iqbal and Horvitz measured, in 2007, a cognitive recovery arc of 20 to 25 minutes following a single alert, and found that 27% of interrupted tasks were never resumed at all. Those are not soft numbers. A recovery arc measured in tens of minutes means a single badly placed alert can cost more than the entire session of work it was meant to help, and a quarter of abandoned tasks is a failure mode with no visible error message attached. This line stalled anyway, because the systems of that era had almost no benefit to offer in return. The best case was reading an email a few seconds sooner. When the benefit term is that small, careful accounting on the cost side has nowhere to go: the rule says stay quiet almost always, and there is little left to design.
The second line runs from 2024 to 2026 and finally has executors that can do real work. YETI, ProAssist, VisionClaw, and PRISM all run multi-step tasks. But they rebuilt the decision layer from scratch. Citation between the two lines is close to absent: where it occurs at all, it is a passing related-work mention that carries none of the formalism into the model. The cost term instead gets rediscovered piecemeal, in fragments, by systems that do not know they are rediscovering it. PRISM prices false positives with Bayesian risk and pushes its false-positive rate from 50.22% down to 22.94%. Precision Proactivity measures cognitive load but never wires it into the gate. EgoSocial names social violations without quantifying them. WakeAnchor and ProAgent account for compute overhead rather than user cost. Two specific quantities, attentional state and the cost of staying silent, have gone untouched since 1999.
What happens when the cost term is missing has been demonstrated experimentally. One team gave a proactive agent a one-sided reward that penalized false positives without rewarding timeliness. GPT-4o's recall collapsed from 98.11% to 56.76%. The agent did exactly what it was asked to do. With only one side of the ledger scored, silence is always the safe move, and the model found that out faster than its designers did. That both terms are load-bearing is something the field relearned the expensive way.
How strong is the evidence
The survey argues that evaluation is the gating layer: until intervention quality is measurable, there is no way to define an RL reward for proactivity at all. It sorts measurement quality into four tiers. L0 is explicit accept-versus-dismiss signal, which is cheap and attributable but badly biased. L3 is the real target, an honest measure of whether this particular intervention was well timed. The entire field trains on L0, because L0 is the only tier that is operationally available, and L0 cannot distinguish "I want this" from "I want this right now."
The evidence for that gap is PRISM itself. Gate on acceptance alone and its false-positive rate rebounds to 62.50%. Users will happily accept a useful suggestion delivered at a terrible moment. The acceptance click confirms that the content was worth something; it says nothing about what the interruption cost to collect it. Optimizing against that signal drives a system toward suggestions that are hard to refuse rather than suggestions that are well placed.
No existing benchmark answers all three questions at once: was it welcome, did it catch the right thing, was the cost priced in? Pro²Bench scores both false positives and false negatives as 0 on quality, so interrupting someone with something irrelevant and watching a critical task fail in silence draw identical penalties. The distinction that timing quality lives in disappears completely. EgoSocial reports macro F1 of 51.65 to 56.08 on recognizing social context, but only 12.50% to 17.67% on intervention timing. The models read the room; they just cannot pick the moment.
The survey flags its own weaknesses too. Of the 50 references, 21 are unreviewed arXiv preprints from 2024 to 2026, which is the standard hazard of an emerging area. Where prose and tables disagree the tables govern; where formulas and released code disagree the discrepancy is noted separately. As a survey, its main contribution is still gathering the scattered cost terms into one usable ledger. Its one new experiment is the pilot described below, and both its scale and its conclusions are limited.
The human-annotation part of the benchmark design is still a proposal, not a result: a single scenario family, 200 to 300 windows, 3 annotators per window, seeded from real intervention moments taken from instructor utterances in HoloAssist. The key design choice is replacing binary labels with three-way windows: must intervene, may intervene, must not intervene. That middle band is exactly where cost-benefit reasoning does any work, and binary annotation erases it. Annotator disagreement is preserved rather than adjudicated, since how disruptive an interruption feels genuinely varies from person to person. Averaging that away would manufacture a ground truth that does not exist. Whether a benchmark built this way separates systems that currently look identical is an open question, and the survey does not claim to have answered it.
The accepted version adds a machine-scored pilot (§8.4) that answers a narrower question first: can the metrics be run at all? Labels are seeded by rule from HoloAssist's public annotations: wrong actions become must-intervene windows, instructor-initiated follow-ups become may-intervene windows, and stretches where the user is speaking become must-not windows, for 1,758 sessions and 131.5 hours. Among five reference policies, the one that speaks every second has the highest timing recall (0.962) and the lowest cost-weighted score at every exchange rate in the sweep (−367,752 at λ = 1). A simulated detector set to catch 70% of mistakes, behind a cost-sensitive gate, reaches 0.703 recall with a 0.09% violation rate against 0.82%, and scores +5,932 at λ = 1. From λ = 2 upward, never speaking wins, which is why the exchange rate is reported as a curve and not fixed. The pilot shows the metrics are executable and separate the policies. It does not show the labels are right: they are seeded by rule, not judged by people, the detector is simulated, and inter-annotator agreement is the first deliverable of the next phase.
What it means in practice
More and more of the agents orchestrated through FIM One carry proactive triggers: scheduled checks, anomaly alerts, task pushes. "When to interrupt the user" is not a matter of taste on our side. It decomposes, term by term, along the rule above: what is this notification worth if it lands, what does landing it cost right now, and where is the line.
Two lessons transfer directly. First, the reward has to be two-sided. Penalizing false positives alone trains the assistant into silence, and someone has already paid the tuition on that experiment. Second, cost is state-dependent. The same reminder is not worth the same amount to a person in a meeting and a person staring out a window. PRISM fixes its cost ratio for the entire deployment, which is precisely where the central insight of the 1999 paper remains absent today. A single global threshold is the easy thing to ship and the thing that will be wrong most often.
The honest summary is that nobody currently has a way to check this properly. Until intervention timing can be measured on its own terms, teams shipping proactive features are tuning against acceptance rates and hoping the gap does not matter.