Emotive Architectures: The Role of LLMs in Adjusting Work Environments

1. Problem Statement

The paper addresses the growing disconnect between static physical/digital work environments and the dynamic emotional and cognitive needs of remote and hybrid workers.

• The Gap: Traditional "smart" environments rely on rigid, pre-defined logic or simple sensor triggers (e.g., motion sensors turning on lights). They lack the ability to interpret nuanced human states (stress, fatigue, focus) expressed through natural language or subtle behavioral cues.
• The Challenge: Remote work has blurred the boundaries between domestic and professional spaces ("phygital" spaces), creating complex, entangled environments that static architecture cannot adapt to. Current digital tools often fail to replicate the depth of human connection, leading to isolation and reduced well-being.
• The Limitation of Current AI: While AI can process posture or proximity, it often struggles with interpreting nuanced social interactions and emotional context without explicit programming. Furthermore, there is a lack of frameworks that treat the environment as a "co-agent" rather than just a tool.

2. Methodology

The research employs a mixed approach combining theoretical framework development and a practical experimental prototype.

A. Theoretical Framework

The authors define three spatial conditions and their fusion:

1. Physical Environment: Viewed not as a static container but as a relational, sensory system capable of dynamic adaptation (drawing on architects like Cedric Price and Yona Friedman).
2. Virtual Environment: Utilizing immersive technologies (VR/AR/MR) and generative AI (LLMs, DALL·E) for content creation and interaction.
3. Hybrid Fusion: The "entangled coexistence" where digital and physical layers mutually transform one another.

• Core Mechanism: The paper proposes Large Language Models (LLMs) as "semantic mediators." Instead of executing direct commands, LLMs interpret vague, affective, or ambient language (e.g., "I feel overwhelmed") and translate them into spatial interventions (lighting, acoustics, interface changes).

B. Experimental Prototype: The Adaptive Workpod

To validate the framework, the authors developed a prototype called the Adaptive Workpod.

• System Architecture:

  • Input Layer (Multimodal):
    • Audio: Open microphone capturing natural verbal expressions (e.g., "I need to focus").
    • Video: Webcam capturing engagement indicators (gaze direction, posture shifts, facial fatigue).
    • Digital Logs: Optional browser activity tracking to infer focus drift.
  • Processing Layer (LLM Agent):
    • Utilized ChatGPT-4o (via API) in advanced voice-and-vision mode.
    • The LLM processes inputs in real-time, utilizing short-term memory of prior interactions to infer user intent and emotional state.
    • It generates semantic recommendations rather than hard-coded rules.
  • Output Layer (Actuators):
    • Ambient Lighting: Adjustable color temperature and brightness.
    • Audio: Sound masking (white noise), music, or silence.
    • Interface: Screen display modes (e.g., distraction-reducing overlays, full-screen immersion).

• Evaluation Protocol:

  • Participants engaged in four remote work sessions.
  • Metrics: Self-reported focus/stress levels (pre/post), perceived appropriateness of interventions (rated as helpful, intrusive, or irrelevant), and system latency.

3. Key Contributions

1. Conceptual Shift: Proposes a paradigm shift from "Space as Interface" (a controllable surface) to "Space as Atmosphere" (a co-adaptive, expressive field). The environment is redefined as a participant in the user's experience, capable of "listening" and "interpreting."
2. Semantic Mediation: Demonstrates how LLMs can bridge the gap between fuzzy human emotional expression and concrete environmental adjustments, acting as a translator between affect and architecture.
3. Co-Adaptive Framework: Introduces a design framework where the environment learns from user interactions over time, moving beyond static automation to dynamic, context-aware responsiveness.
4. Ethical Integration: Explicitly addresses the ethical implications of emotion AI, emphasizing the need for transparency, user agency, and privacy, particularly in light of the EU's proposed AI Act restrictions on emotion recognition.

4. Results

The Adaptive Workpod prototype was tested against four specific behavioral scenarios, yielding the following results:

| Scenario | User Cue | LLM Interpretation | Intervention | Measured Outcome |
| :--- | :--- | :--- | :--- :--- |
| Drowsiness | "I'm feeling a bit drowsy." | Low energy/risk of mind-wandering. | Switched lights to cooler/brighter white; suggested a standing stretch. | +1 level alertness (5-point scale); posture corrected in <30s; rated "helpful." |
| **Focus Loss** | Gaze off-screen >10s. | Loss of visual focus/fatigue. | On-screen break reminder; dimmed peripheral lighting. | 50% reduction in off-screen gaze (5 min); latency <1s; <10% intrusive rating. |
| Distraction | 2x 5-min social media visits. | Procrastination/distraction. | Blocked social media for 5 mins; played soft white noise. | Return to work domains in <2 mins; +1 focus score; 80% rated "helpful." |
| Stress | "This task is stressing me out." | High stress signal. | Warmed light hue; guided breathing exercise via voice. | -1 stress score (3 min); completed breathing prompt; system recalled intervention in later sessions. |

• Overall Findings: The system successfully demonstrated that LLM-mediated semantic cues could trigger spatial adaptations that improved focus, reduced stress, and increased user well-being without invasive biosensors. The "memory" feature allowed for increasing personalization over time.

5. Significance and Future Outlook

• Architectural Innovation: This research redefines the role of the architect and designer. The designer's role shifts from scripting fixed functions to "choreographing" relationships between language, mood, and environment.
• Human-Centered AI: It validates the potential of using generative AI not just for content creation, but for creating emotionally resonant environments that support human dignity and well-being in the hybrid work era.
• Future Directions: The authors identify the need for:
  • Enhanced actuation capabilities.
  • Expanded user modeling (long-term memory).
  • Integration of physiological sensing (e.g., heart rate) for higher precision.
  • Critical Imperative: Maintaining strict ethical safeguards regarding privacy, data governance, and user consent to prevent the misuse of emotion tracking.

In conclusion, the paper argues that by integrating LLMs as semantic mediators, we can create co-adaptive environments that are not merely reactive but are capable of understanding and responding to the complex emotional and cognitive rhythms of human work.