Fluency Metrics for Real-Time Human-AI Interaction

5 minute read

Most benchmarks for AI assistants measure what the system does: whether it completed the task, how accurately, and how fast. These are the right questions to ask of a tool. I would argue they are incomplete questions to ask of a collaborator.

Consider a real-time voice assistant working with a person on a task. The assistant waits until the person has finished speaking, processes the request for a few seconds, and returns a complete and correct answer. By every standard measure, the interaction succeeded. But the person sat idle while waiting, perhaps glanced at their phone, perhaps lost the thread of what they were doing. Repeat this a few dozen times in a session, and the person may be less inclined to use the system again, with none of this showing up in a success rate, an accuracy score, or a latency number.

What is missing is a measure of how well the agent's actions mesh with the human's in time. This is a question the human-robot interaction (HRI) community has spent more than a decade answering, and I would argue that the answers transfer to real-time AI systems more directly than one might expect.

What is fluency?

Guy Hoffman codified a set of objective metrics for collaborative fluency [1]. Human Idle Time (H-IDLE) is the fraction of the task during which the human is waiting. Robot Idle Time (R-IDLE) is the same for the robot. Functional Delay (F-DEL) is the time between the end of one agent's action and the start of the other's, when the second depends on the first. Concurrent Activity (C-ACT) is the fraction of the task during which both are active at once. Alongside these objective measures sits a subjective fluency scale, in which the human rates how smoothly the collaboration felt.

What strikes me about these metrics is that they are defined in terms of temporal activity patterns, not in terms of what the activity is. Nothing in them refers to grasping, handing over, or moving through space. A human and an agent trading turns in a conversation, or working in parallel on a document, produce the same kind of timeline as a human and a robot assembling a part. That is what makes the framework a natural candidate for real-time AI.

Why not just measure latency?

Latency is a property of the agent alone. Fluency is a property of the pair. Two systems with identical response latency can produce very different human idle time, depending on whether the agent acknowledges a request while it works, breaks a long answer into increments, or begins when the human's intent is clear rather than when the human stops speaking. Latency tells you how long the agent took. It does not tell you what the human was doing in the meantime, or whether they knew what to expect.

We saw the difference this makes in our cross-training work [2]. The human and robot switched roles during training, so that each observed how the other preferred the task to be done. The resulting improvement in temporal coordination — 71% more concurrent motion and 41% less human idle time compared to interactive reinforcement learning — was accompanied by higher ratings of trust and of perceived alignment with their preferences. Participants did notice and valued the coordination itself.

Perhaps the most interesting finding in this literature is that anticipation matters beyond efficiency. Hoffman and Breazeal [3] found that a robot acting on a prediction of its partner's next need was perceived as a more fluent teammate, even where the gains in task time were modest. For a voice assistant, responding promptly once intent is clear, or offering the next step as the current one wraps up, is the same phenomenon in a different medium.

What changes when the partner is an AI rather than a robot?

The transfer is not entirely mechanical. A few definitions have to be rethought.

First, activity boundaries. For a robot, activity is motion. For a voice agent, the natural analogue is speech: human activity runs from voice onset to voice offset, and AI activity from the onset of synthesized speech to its end. Text and agentic settings are murkier. Is a developer reading a diff idle? Is the agent idle while it waits for a test suite to finish?

I would argue that in knowledge work the quantity we care about is not idleness but being blocked: the human waiting on the agent with nothing useful to do in the meantime. An agent that keeps the human informed while it works, or hands back something partial to read, is reducing blocked time even when its total response time is unchanged.

Second, concurrency in a single channel. While two people can work simultaneously on separate activities, overlapping speech in a conversation can be disruptive. C-ACT in a voice setting may therefore capture interruptions rather than productive concurrency. In agentic settings, where the AI can work in the background while the human does something else, concurrency regains its original meaning.

Why does this matter?

In the previous post I argued that the objective of a collaborative agent is not only to complete the current task, but to maximize the probability that the human wants to collaborate with it again. Fluency is one of the few process measures that plausibly speaks to that objective. A person who was never left idle, never interrupted at the wrong moment, and never had to wait without knowing why has had a different experience than one who received the same final output on a worse timeline. The output looks identical on a leaderboard. The willingness to come back does not.

The HRI community linked fluency to trust and preference on physical tasks over many years of studies. The deeper question is whether temporal coordination metrics developed for physical collaboration can predict how people experience real-time AI interaction, and ultimately whether they want to collaborate with the system again.

References

[1] G. Hoffman. "Evaluating Fluency in Human-Robot Collaboration." IEEE Transactions on Human-Machine Systems, 2019.

[2] S. Nikolaidis and J. Shah. "Human-Robot Cross-Training: Computational Formulation, Modeling and Evaluation of a Human Team Training Strategy." HRI, 2013, pp. 33–40.

[3] G. Hoffman and C. Breazeal. "Effects of Anticipatory Action on Human-Robot Teamwork: Efficiency, Fluency, and Perception of Team." HRI, 2007.

Partner World Models for Collaborative Agents

5 minute read

Yann LeCun said that “I don’t understand how you can even think of building an agentic system without the agentic system having the ability of predicting the consequences of its actions.” [1]

I find that this argument holds even more strongly in the case of human-AI -- or human-robot -- collaboration, where what an agent's action is worth is a function of what it makes the human partner do next.

This requires having a model of human behavior that is not simply a generic forecasting model of what a partner will do next. Instead, it is an action-conditioned predictive model of how relevant latent states and observable behaviors evolve in response to the agent's actions. I would argue that this is a world model in a technical sense, where the states are compact representations of human decision making.

Why do we need a partner model?

With such a model, the agent can explore counterfactuals that it cannot afford to run. The agent can imagine how the human would react in different scenarios that it does not actually need to enter.

For instance, if a surgical assistant pre-positions an instrument before it is asked, will the surgeon take it, or will the interruption break their sequence?

Or consider a coding agent. If the agent writes 200 lines of code in one shot, will the developer read them carefully or skim and merge? A patch that gets merged because it was too large to review could actually be a worse outcome than the same patch, split into three reviewable commits, that is accepted more slowly but understood much better. 

But why not skip the partner model and learn the policy directly?

In the limit of infinite data, a sufficiently expressive policy could presumably learn to take the best action given the interaction history. I would argue, however, that explicitly modeling the partner provides a better inductive structure for generalization, adaptation, and counterfactual reasoning from finite interaction data.

First, human collaboration itself appears to rely heavily on structured representations of other people. Humans that engage in joint activity represent both a shared objective and the distinct roles and perspectives that participants occupy within it, reasoning about one another's goals, beliefs, and knowledge in order to coordinate [2].

The representational benefit can be understood through role reversal. A child who has only ever played one side of a collaborative game can, on request, immediately perform the complementary role -- the one they have only watched -- without retraining [3]. There is no obvious equivalent mechanism for a policy that simply maps its own interaction history to actions.

Second, the goal of human-AI collaboration should not be simply to complete the current task as efficiently as possible. It should also be to maximize the probability that the human wishes to collaborate with the agent again in the future.

This is inherently a long-horizon property of an interaction. A partner model can turn this into a denser signal, where deviations from expected behavior may serve as leading indicators. For instance, a developer may hesitate before accepting a pull request, repeatedly request explanations, or take over a subtask that they previously delegated to the agent. 

Third, human behavior is not stationary with respect to the agent. Humans adapt to their collaborators. The agent's actions therefore affect not only what the human does next, but also what the human learns about the agent, what conventions emerge between them, and how the human behaves in future interactions.

A partner model allows the agent to reason about these coupled dynamics rather than treating the human as an exogenous source of observations.

Such a model can also provide a quantitative measure of uncertainty over human behavior dynamics. If the agent is unsure how the human is going to respond, that is actionable information -- a reason to slow down, explain its action, or ask a question. It is hard to induce a comparably interpretable introspective quantity directly into a value function.

How can we learn such a model?

Learning a partner model creates a chicken-and-egg problem. To interact effectively with the human, the agent needs to anticipate how the human will react to its actions, which in turn requires interacting with the human in the first place.

More than a decade ago, we touched on this problem by proposing an iterative approach in a simple human-robot collaboration setting. The human and robot would cross-train on the task by switching roles [4].

The idea was that role switching allowed the robot to build a model of the human by observing how the human would react to actions that the robot itself would later execute. At the same time, it gave the human a way to demonstrate to the robot how they preferred the robot to behave. 

In a study with 36 participants, cross-training outperformed interactive reinforcement learning on perceived trust and on objective measures of team fluency, with 71% more concurrent motion and 41% less human idle time.

For a coding agent, switching roles would obviously not be physical. Instead, it could take the form of structured interaction that reveals both how the developer responds to the agent and how the developer wants the agent to behave.

One possibility would be a structured training curriculum that exposes the agent and the human to low-risk, informative interactions. For instance, the agent might first be asked to explain an existing function, then propose a one-line modification, then generate a small patch, then refactor a module, and only later undertake a large multi-file change.

At each stage, the agent would observe not only whether the task succeeds, but also how the developer responds. Does the developer inspect the output closely? Do they request an explanation? Modify the proposal? Take control? Accept the change directly? Ask the agent to proceed more autonomously?

Overall, I would argue for a shift in how we think about learning in collaborative agents: from learning how to act to learning a model of the partner rich enough to reason about how to act together.

References

[1] Y. LeCun, talk, https://www.youtube.com/watch?v=kYkIdXwW2AE

[2] M. Tomasello. The Evolution of Agency: Behavioral Organization from Lizards to Humans. MIT Press, 2022.

[3] Fletcher, G. E., Warneken, F., & Tomasello, M. (2012). Differences in cognitive processes underlying the collaborative activities of children and chimpanzees. Cognitive Development.

[4] S. Nikolaidis and J. Shah. “Human-Robot Cross-Training: Computational Formulation, Modeling and Evaluation of a Human Team Training Strategy.” HRI, 2013, pp. 33-40.