Skip to content
Technical University of Munich

Adaptive Human–AI Collaboration Through EEG-Based Cognitive Load Modeling

Type: Master

Date: Immediately

Supervisor: Nils Ness

Artificial intelligence increasingly supports humans in complex tasks, ranging from visual classification and multimodal sensemaking to planning, writing, and high-stakes decision support. While state-of-the-art systems often achieve expert-level performance, they typically operate without awareness of the human collaborator’s cognitive state. Humans, by contrast, continuously adjust their communication and reasoning strategies based on another person’s workload, attention, and mental effort (Sweller, 1988).

This discrepancy creates a critical research gap: How can AI systems become sensitive to human cognitive load, and how might this change the dynamics of human–AI collaboration?

Recent advances in EEG representation learning and neural decoding frameworks such as Thought2Text demonstrate that brain signals can be mapped into meaningful embedding spaces that influence the outputs of large language models (Mishra, 2023). Building on this foundation, this thesis invites a master’s student to investigate how human cognitive load, measured using EEG analysis, can be used to adapt an AI agent’s behavior during collaborative tasks.

Rather than decoding thoughts or semantic content from EEG signals, this research focuses on cognitive load as a control signal: guiding the AI to modulate its explanations, level of detail, pacing, and interaction strategy according to the human’s mental effort. This approach opens new pathways toward AI systems that communicate more like cooperative human partners (Kosch et al., 2021).

Research Context and Scenarios

To systematically explore this concept, the thesis will center on one or two collaborative task scenarios from different cognitive domains where load-adaptive behavior is relevant, such as:

  • Image classification under time pressure
  • Explanatory assistance
  • Error detection or verification tasks

Each task will serve as a context for analyzing how cognitive load fluctuates and how an AI system should adapt to optimize collaboration, communication, and shared task performance.

Objectives and Methodology

  1. Develop an EEG-based cognitive load inference and embedding framework.
    Select and preprocess EEG datasets from problem-solving tasks, analyze neural markers of load (e.g., frequency band ratios, event-related potentials), and design an embedding representation suitable for AI integration. Based on prior multimodal frameworks like Thought2Text (Mishra, 2023), develop an approach for translating EEG features into embeddings capable of conditioning a large language model.
  2. Evaluate adaptive versus non-adaptive AI collaboration through simulation tasks.
    Integrate the EEG-based cognitive load embedding into an LLM using approaches such as control tokens, concatenated vectors, or cross-attention. Simulate collaborative tasks under varying cognitive load levels to compare adaptive and non-adaptive AI behaviors in terms of explanation style, pacing, reasoning clarity, and overall task performance. Analyze outcomes to identify principles, limitations, and opportunities for designing cognitive-load–aware AI systems that enhance communication and cooperation between humans and machines (Kosch et al., 2021; Sweller, 1988).

Selected References

  • Kirwan, C. B., Vance, A., Jenkins, J. L., & Anderson, B. B. (2023). Embracing brain and behaviour: Designing programs of complementary neurophysiological and behavioural studies. Information Systems Journal, 33(2), 324–349. https://doi.org/10.1111/isj.12402
  • Kosch, T., Hassib, M., Buschek, D., & Schmidt, A. (2021). Investigating cognitive load in human–computer interaction using psychophysiological methods. ACM Transactions on Computer-Human Interaction, 28(3), 1–32.
  • Mishra, A. (2023). Thought2Text: Decoding EEG signals into text with pre-trained language models. GitHub Repository. https://github.com/abhijitmishra/Thought2Text?tab=readme-ov-file
  • Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science, 12(2), 257–285.
  • University of Arkansas, Davis, F., Riedl, R., University of Applied Sciences Upper Austria, University of Linz, Hevner, A., & University of South Florida. (2014). Towards a NeuroIS Research Methodology: Intensifying the Discussion on Methods, Tools, and Measurement.Journal of the Association for Information Systems, 15(10), I–XXXV. https://doi.org/10.17705/1jais.00377
To top