Every interface designer learns the same rule early: when the system is busy, say so. An unexplained pause reads as a crash, a spinner reads as progress, and the honest signal is the kind one.
A paper submitted on 16 October argues that in one specific and increasingly common situation, this rule is backwards.
The setup
Pratham Arora, Abhinav Lodha and Siddharth Siddharth ran a within-subjects study with 18 participants, comparing three isolated feedback approaches during waits in conversations with an LLM-driven agent in VR:
| Approach | Example |
|---|---|
| Verbal fillers | ”um…”, “hmm…” |
| Visual indicators | loading spinners |
| Embodied gestures | thinking animations |
The framing in the abstract is precise about the gap: previous work has established that feedback reduces the perception of waiting, but “we lack understanding of which types work best and why.” So the design isolates the three channels rather than comparing bundled designs.
The result
Embodied gestures significantly outperformed visual loading indicators across multiple dimensions, making agents feel more lifelike and socially present.
And the finding with teeth: “explicit visual feedback like loading spinners actually degraded perceived humanness and broke immersion by forcing users to shift from social to technical mental models.”
Not merely less effective. Worse than the alternatives, by a mechanism the authors name.
The mechanism is the contribution
The explanation is the part worth keeping, because it predicts things beyond this study. The authors’ reading is that “natural, behavior-based feedback maintains the illusion of conversing with a thinking entity, while technical indicators expose the computational substrate.”
A spinner is not a neutral progress signal. It is a piece of vocabulary from a different register — the register of file transfers and page loads — and dropping it into a conversation reclassifies the thing you are talking to. You were speaking to someone; now you are waiting on a process. Once the user has switched to a technical mental model, every subsequent response gets evaluated as output rather than as speech, and the social presence the rest of the system was building is spent.
A thinking animation does the same informational job — I have received your input, I am not frozen, keep waiting — in the register the interaction is already in. Pausing and looking away is what a person does when considering an answer. It is latency feedback that does not announce itself as latency feedback.
Verbal fillers are the interesting middle case: they are behavioural rather than technical, which should put them on the right side of the divide, but they occupy the audio channel the agent needs for its actual answer, and a filler is a much more specific social performance than a gesture — “um” implies a particular kind of hesitancy.
Why this matters more than it sounds
Waiting is not an edge case for VR agents, it is the steady state. An LLM response — particularly one routed through speech recognition, inference and speech synthesis — takes a conversationally awkward amount of time, and unlike a chat window, a VR agent is a body standing in front of you doing nothing during the gap. The embodiment that makes the agent compelling is exactly what makes the latency visible.
That is also why the finding does not obviously transfer. Nothing here says spinners are bad in general — the mechanism depends on the interaction already being social. In a file manager there is no illusion to break. The claim is narrower and sharper: when your interface has asked the user to treat something as an entity, do not then show them the machine.
There is an uncomfortable corollary the paper does not need to spell out. Making a system’s delays legible is usually considered a form of honesty, and this result says that honesty has a measurable cost in the thing designers of these systems are optimising for. Whether a thinking animation that conceals a 4-second inference call is good design or a small deception depends on what you think the agent is for, and that question is not an empirical one.
Limits
Eighteen participants, within-subjects, three isolated conditions. That is a reasonable size for detecting a large effect in a controlled comparison and it is not a basis for broad claims about agent design. The conditions are deliberately isolated, so the obvious next question — whether a gesture plus a filler beats either, or whether combining them is what breaks the spell — is outside what was tested.
The paper presents “empirically-grounded design guidelines for creating more engaging AI-powered VR agents,” which is a reasonable output for a study of this size, with the usual caveat that guidelines derived from 18 people in a lab are a hypothesis for your own product rather than a conclusion about it.
Sources
- Wait for It: A Component-Level Analysis of Latency Feedback Mechanisms for LLM Agents in VR — arXiv:2610.23336
- arXiv PDF