BeatriXR
Comprehensive and adaptive feedforward support for guidance in virtual reality
A toolkit for authoring direct feedforward in VR — configurable Triggering, Previewing and Exiting blocks set from in-headset panels, with an optional LLM advisor that proposes configurations.

Feedforward shows a user what an action will do before they commit to it. In VR these previews are usually hand-coded and tightly coupled to one application, which makes them hard to reuse and the design space hard to explore. BeatriXR organises feedforward into reusable blocks — how a preview is triggered, what it shows, and how it ends — authored from a settings panel rather than in code. A language-model advisor can suggest configurations as a mixed-initiative aid, leaving the final choice to the designer.

Unity + MRTK 3 11 XR experts evaluated it EICS 2026 · 32 pages
The gap

Good feedforward is hard to build, and hard to get right

Designers must coordinate many interdependent choices — when a preview triggers, how the action and outcome are shown, how it transitions back — usually coded ad hoc inside a game engine. Even with a toolkit, the space of combinations is large enough that good configurations are easy to miss.

Without a toolkit

Feedforward built ad hoc

  • Hand-coded per app and tightly coupled to application logic
  • Triggers, previews and transitions are hard to reuse
  • The large design space is easy to under-explore
TOOLKIT
With BeatriXR

Configurable building blocks

  • One toolkit covering Muresan et al.’s feedforward design space
  • Trigger / Preview / Exit set from a UI — no programming
  • An LLM advisor helps explore alternatives

Direct feedforward = showing the user, with contextual previews, exactly what an interaction will do. BeatriXR focuses on the visual side of that, deeply.

Contributions

What the paper delivers

C1

A modular feedforward toolkit

Create, visualise and customise direct feedforward with avatars and interface panels, mapped onto an established design space and the triggering / previewing / exiting phases.

C2

An LLM decision-support layer

A generative model proposes configuration alternatives to help designers explore the space — a mixed-initiative aid, not an autopilot.

C3

A working proof of concept

Built in Unity with MRTK 3, demonstrated in an immersive hazardous-materials training scenario; code is open on GitHub.

C4

An exploratory expert study

XR experts assessed the toolkit, its configuration interface, and how useful LLM-generated suggestions were for feedforward design.

Try it · interactive

Configure a feedforward preview

A browser recreation of what BeatriXR’s in-headset settings panel controls. Adjust how a preview demonstrates the action “operate the control” — the avatar, viewpoint, rendering, trigger and exit — and see the configuration described below. Switching to AI advisor loads a configuration in line with what experts in the study favoured.

Feedforward preview · operate the control
Toggle the settings · try the AI advisor
Preview
AvatarFull body
ViewThird-person
Ghosted avatar
Duplicate target
Dislocate preview
Signifiers
Exit
TransitionInstant
State

Illustrative recreation — a stylized scene and avatar, not the paper’s real VR renders or LLM. It mirrors BeatriXR’s actual settings (avatar type, view, ghosting, duplication, dislocation, signifiers, trigger and exit). The AI preset reflects what experts in the study tended to prefer; see the PDF for the real system.

The design space

Three phases, many settings

BeatriXR operationalises an existing feedforward design space (Muresan et al.) into three reusable phases. A designer sets each one from a panel; because the settings combine, the space of possible previews is large — which is what the reusable blocks and the LLM advisor are there to help manage.

1 · Triggering — when it appears

What makes a preview start playing.

  • Gaze, Position, Timer, Action or Persistent
  • Optional signifiers hint where to look or act

2 · Previewing — what is shown

How the action is demonstrated.

  • Avatar type: full / partial / real user / none
  • Perspective: first- or third-person
  • Rendering: ghosted or opaque
  • Level of detail; target = avatar, action or outcome
  • Duplication and dislocation of the preview

3 · Exiting — how it ends

When and how the preview goes away.

  • Untrigger: look away, move away, or automatic
  • Transition: rewind, instant, or fade

The dimensions and option values above are taken from the paper’s design space; it builds on Muresan et al.’s feedforward taxonomy. See the PDF for the full treatment.

How it works

Three things the user touches

At runtime BeatriXR is three elements working together — a guidance UI, avatar demonstrations, and an optional AI module.

UI

Task & Settings panels

The Task Panel shows progress and alternative interactions as a stack of cards. The Settings Panel exposes the Trigger / Preview / Exit options as arrow-cyclers and toggles, with an explanation of each.

progressindirect cues
AVATARS

Demonstrating avatars

A disembodied avatar acts out the interaction using inverse kinematics, with no prior scene knowledge. A compare mode shows two interaction options side by side in separate panels.

singlecompareIK
AI

LLM advisor

A local language model reads the current settings and task sequence and proposes alternative configurations. The expert can adopt the overrides or keep customising — control stays human.

design-timemixed-initiative
How they evaluated it

Two rounds with XR experts

An exploratory, expert-driven evaluation — first formative, then a fuller session on the final UI and the LLM suggestions.

DESIGN

Two sessions

A formative mockup study, then an exploratory study of the finished interface and four LLM advisors.

FORMATIVE

6 experts, 3 tasks

Their feedforward configurations aligned well with reference rankings (means ~12–13 / 15). Feedback turned radio buttons into toggles.

TASKS

Three interactions

Touch a virtual mouse, laser-grab a beaker, and grab to turn a sink — chosen to span common VR interaction styles.

LLMs

4 models rated

GPT-OSS-120B, Mistral Small 3.2, DeepSeek R1 and GPT-5 Mini, scored on quality, usefulness and resulting feedforward.

What they found

Read first time, mostly right

Results are exploratory and qualitative — expert perceptions, not task-performance measures — but consistently positive about the panel’s structure.

COMPREHENSION

0% correct

Across the three tasks (76 / 91 / 88% by task), experts read the settings panel correctly on first exposure — understanding rose as they got used to it.

EXPERTS

0 reviewers

Most with professional or academic XR experience. Self-rated comprehensibility climbed from 5.8 to about 6.4 out of 7 over the tasks.

LLM ADVICE

All 4 were plausible

Every model produced reasonable suggestions; alignment with expert reasoning was only partial. DeepSeek R1 stood out for clear, context-aware output.

Experts liked clustering settings into Trigger / Preview / Exit, but flagged icon ambiguity and information density. They preferred partial, ghosted avatars and manual triggers, and rejected previews that take control away from the user.

Takeaways

Five things to carry forward Tap to expand.

Under the hood

From task file to live preview

Seven modules connected by APIs, so any piece can be swapped as long as it honours the contract.

ENGINE

Unity + MRTK 3

The Mixed Reality Toolkit handles complex VR interactions with minimal setup — mostly assigning components to scene elements.

AVATAR

Rocketbox + IK

A Rocketbox model with per-interaction animations, adapted on the fly by inverse kinematics and rendered opaque or transparent, full or partial.

LLM

Local & deterministic

Served via LM Studio with temperature 0, top-k 1, top-p 0 so identical inputs give identical suggestions.

Honest limits

What this study does not show yet

L1

Perceptions, not performance

The evaluation is expert feedback only. It cannot yet quantify effects on development speed, usability, or how well end users actually learn.

L2

The LLM role is preliminary

No systematic comparison against rule-based alternatives, and runtime adaptation is not yet validated — treat the advisor as an ideation aid.

L3

Visual only

Audio and haptic modifiers were deliberately left out to go deep on visual cues, so multimodal coverage is incomplete.

L4

Richness cuts both ways

The breadth of options can overwhelm. Presets, templates and adaptive interfaces are likely needed for broader audiences.

In one breath

Make feedforward something you compose, then explore

BeatriXR operationalises a feedforward design space into reusable Trigger / Preview / Exit blocks, lets domain experts author VR guidance from a panel instead of code, and adds an LLM advisor to widen the options they consider — while leaving the final call to them. Early expert feedback is encouraging on both the interface and the AI assistance, especially for novices.