The research team, led by Yimin Qin of the Department of Mechanical and Aerospace Engineering at UC Davis, tackled a problem that has long frustrated human factors scientists. Trust in automation shapes how operators allocate attention, rely on system outputs, and decide when to intervene, and miscalibrated trust, whether excessive or insufficient, degrades team performance and safety. Yet trust is most often measured with retrospective questionnaires or post-task surveys, tools that offer limited temporal resolution and may be distorted by recall bias. Continuous self-report can capture temporal variation, but it reveals little about the underlying cognitive processes that support trust formation and adjustment.

To bridge this gap, the researchers turned to EEG, a non-invasive technique with millisecond-level temporal resolution that is sensitive to attentional engagement, cognitive control, workload, and decision-related neural activity. Ten participants performed an operationally motivated human-autonomy teaming task emulating satellite-based surveillance operations, in which the goal was to identify ground troop movement with the assistance of a simulated autonomous system. Participants monitored nine simulated satellites through a graphical interface, reviewed the system’s binary classifications of troop movement, assisted in guiding satellite image acquisition, and reported their momentary trust on a continuous slider every 30 seconds. Compensation was tied to the quantity and accuracy of reclassification decisions, introducing tangible risk and reward that kept participants genuinely engaged.

The experiment unfolded over five sessions: an initial training session followed by four testing sessions on separate days. During each testing session, participants interacted with one of four autonomous system variants that crossed two levels of reliability with two levels of explainability. The high-reliability systems classified troop movement with substantially higher accuracy than the low-reliability systems, which were correct approximately two-thirds of the time. Explainability was manipulated through the system’s explanatory style: low-explainability conditions offered terse, machine-like explanations, while high-explainability conditions used more naturalistic language. The four variants, named VEARS, BREES, SPEROS, and LEXAR, were randomly assigned across days, and participants were not informed of the specific characteristics of each system.

EEG data were recorded using a 62-channel g.tec active-electrode headset at a sampling rate of 500 Hz, synchronized with trust reports through the Lab Streaming Layer framework with sub-millisecond accuracy. After preprocessing with bandpass filtering, notch filtering, and independent component analysis, the researchers extracted spectral power features within six frequency bands: delta, theta, alpha, beta, low-gamma, and mid-gamma. Using a sliding-window approach over 2-second windows, spectral power was computed across all 62 channels and aggregated to match the temporal resolution of trust reporting, yielding a high-dimensional feature vector of 372 spectral features at each trust update.

The heart of the study lies in its modeling framework. The researchers proposed a non-linear dynamic model (NDM) that treats trust as a latent dynamic process evolving over time and indirectly reflected in EEG features. The model integrates a low-dimensional linear latent process, representing the temporal evolution of trust through a state-space structure with Kalman filtering and Rauch-Tung-Striebel smoothing, with a non-linear manifold that maps EEG spectral features to the latent state. The manifold is learned through multilayer perceptrons with tanh activation, and a supervised non-linear mapper predicts trust from the inferred manifold state. The NDM was trained jointly on multi-step EEG prediction and trust prediction objectives, with the dynamic and manifold latent dimensions fixed at two to force the model to explain high-dimensional EEG features through a compact state representation.

Two baseline models were implemented for comparison: a linear dynamic model (LDM) that retains the temporal state-space structure but assumes linear relationships between EEG features, latent state, and trust, and a linear static model (LSM) that ignores temporal dynamics entirely and treats trust inference as an instantaneous regression problem with LASSO regularization. All models were evaluated using Monte Carlo cross-validation, with one trial from each autonomous system condition randomly selected for testing and the remaining trials used for training, repeated twenty times. Crucially, the researchers corrected the unit of statistical inference to participant-level averages, conducting all paired tests on the ten independent subject-level averages to avoid pseudoreplication from repeated cross-validation folds.

The results were striking. The NDM achieved significantly higher correlation with self-reported trust than both the LDM (p = 0.0021) and the LSM (p < 0.0001), and significantly lower prediction error than the static model (p = 0.0020). The RMSE reduction relative to the linear dynamic model was numerically favorable but not statistically significant after participant-level aggregation. The advantage of the NDM became most evident when trust exhibited meaningful temporal structure: for participants whose trust oscillated rapidly between high and low values, the NDM closely followed the amplitude and timing of trust changes, while the baseline models produced overly smoothed trajectories that failed to capture fluctuations. For drifting trust patterns, the NDM more accurately reflected the rate of trust evolution and intermediate deviations, though all models generally captured the overall direction of change.

Model interpretation revealed that trust inference relied on distributed neural features rather than a single brain region. Delta and alpha band features exhibited spatially coherent importance patterns over frontal and fronto-central regions, particularly along midline electrodes such as FPz, Fz, and FCz, while gamma-band features displayed numerous high-importance values with greater spatial heterogeneity. When examined across representative trust dynamics, a shared frontal and prefrontal delta-band importance emerged across all three profiles, suggesting that slow frontal neural activity constitutes a core component of trust inference preserved across individuals. Beyond this shared low-frequency signature, distinct spatial-spectral distributions appeared: oscillatory trust dynamics extended more strongly into low- and mid-gamma activity over frontal and parietal regions, potentially reflecting high sensitivity to rapid changes in system behavior, while drifting trajectories showed broader distribution across alpha and low-gamma bands consistent with gradual trust updating and evidence accumulation.

Supervised Uniform Manifold Approximation and Projection (UMAP) applied to the smoothed latent state trajectories revealed that high- and low-trust states occupy distinct regions of the inferred latent space across all three trust dynamics. For oscillatory trust, high- and low-trust samples formed compact and well-separated clusters, suggesting rapid transitions between distinct latent regimes. For drifting trust, the latent embedding exhibited smoothly varying structures with samples distributed along continuous trajectories, consistent with gradual trust evolution. For stable trust, latent states remained within a constrained subspace with high- and low-trust samples separated by relatively short patterns. These findings indicate that the learned latent state captures both trust level and temporal structure, preserving differences in how trust evolves across individuals. The researchers acknowledge limitations, including the modest sample of ten participants in an intensive multi-session EEG design, the exploratory nature of feature-importance interpretations, and the focus on spectral power features rather than functional connectivity or cross-frequency interactions. Nevertheless, the work demonstrates that incorporating non-linear neural structure within a dynamic modeling framework improves EEG-based trust inference and supports the development of cognition-aware adaptive autonomy, pointing toward trust-aware autonomous systems capable of monitoring and responding to evolving human trust states in real time.

Subject of Research: An EEG-based non-linear state-space model continuously infers human trust in autonomous systems during a space-relevant supervisory task, achieving significantly higher correlation with self-reported trust than linear dynamic and static baseline models.

Article Title: An EEG-based non-linear state-space model for trust inference in a space-relevant human-autonomy teaming task

Article References: Qin, Y., Prugsanapan, P., Leary, S., Shih, K. Y., Kong, Z., Hayman, A., & Clark, T. (2026). An EEG-based non-linear state-space model for trust inference in a space-relevant human-autonomy teaming task. Frontiers in Neuroergonomics, 7, Article 1810115. https://doi.org/10.3389/fnrgo.2026.1810115

Image Credits: AI Generated

DOI: 10.3389/fnrgo.2026.1810115

Keywords: EEG, trust inference, human-autonomy teaming, state-space model, non-linear dynamics, neuroergonomics, adaptive autonomy, satellite surveillance, machine learning, latent dynamics, prefrontal cortex, gamma band

The first half of the challenge is hardware. Conventional neural probes are stiff devices made of silicon or metal, and the brain is soft, so every heartbeat, breath and head movement creates relative motion between the implant and the surrounding tissue. Over weeks, that mechanical mismatch triggers an immune response: glial cells wall off the foreign object, signal quality degrades, and neurons that were once cleanly recorded drift out of reach or disappear entirely. The Harvard team built around this by using flexible mesh-like electrode arrays whose mechanical properties are far closer to those of brain tissue itself. Because the probes move with the brain rather than fighting against it, they provoke less of a scarring response and stay in stable physical registration with the neurons around them.

That stability paid off dramatically. Implanting the arrays in the mouse visual cortex and hippocampus, the researchers recorded single-unit action potentials from the same identifiable neurons across sessions spanning five months. Spike sorting and waveform analysis confirmed that the cells tracked over time were genuinely the same neurons: within-neuron waveform correlations stayed high, estimated displacements of the neurons’ positions remained small, and firing-rate statistics were consistent across the recording period. The team compared mice that watched visual movies frequently with mice that saw the same stimuli only occasionally, and found no significant differences in tracking quality between the groups, ruling out the possibility that repeated stimulation itself was distorting the recordings.

With a stable window onto the same cells, the researchers could then measure representational drift directly and rigorously. They presented mice with repeated showings of the same movies and quantified how the population vectors of neural activity changed between recording sessions separated by days or weeks. The drift was real and persistent: correlations between population vectors declined as the time interval between sessions grew, in both the visual cortex and hippocampal CA1. Importantly, the team used deep-learning-based eye and pupil tracking to show that changes in arousal or attention could not account for the drift, and statistical comparisons showed that drift rates did not slow down over the course of the months-long experiment.

The study also probed what drives the drift. When the researchers parametrically scrambled the movies, degrading their higher-order structure while preserving basic statistics, they found that single-neuron tuning correlations and drift rates changed systematically with the level of scrambling, indicating that the statistics of the visual stimulus itself shape how representations wander. They also ran a conditioned task in which a scene in the movie predicted a water reward, and found that reward engagement altered some aspects of drift, such as tuning-curve correlations, while leaving others, including population-vector correlations and latent dynamics, statistically unchanged. Drift was also stronger in hippocampal CA1 than in primary visual cortex, consistent with the hippocampus’s role in rapidly remapping experience.

The second half of the study is where the work becomes most consequential for neurotechnology. Rather than trying to stop the drift, the team modeled it. Using Gaussian-process factor analysis, they extracted low-dimensional latent dynamics, the smooth trajectories that neural populations trace through an abstract state space as they process a stimulus. They found that while the raw latent trajectories from different sessions did not overlap, they were geometrically similar: the same movie drove the population through roughly the same shaped path, just rotated or otherwise transformed in the state space. A classic orthogonal Procrustes alignment, a geometric transformation that finds the best rotation mapping one set of coordinates onto another, was enough to bring the latent dynamics from different months into close correspondence.

Once the latent spaces were aligned, decoding became stable. The researchers trained a decoder on the transformed latent dynamics from one session and found that it maintained high performance across long time spans, accurately predicting which time window of a movie the animal was watching even months later. Decoding from the transformed latent dynamics significantly outperformed decoding from either the raw high-dimensional single-neuron activity or the unaligned latent dynamics, for both support-vector-machine and LSTM decoders, in both frequently and infrequently stimulated mice. In effect, the geometric transformation acts as a translation layer that converts each day’s neural code back into a common reference frame, so the decoder never needs to be retrained.

The team pushed the idea further with a proof-of-concept general model that spanned both time and animals, showing that aligned latent dynamics could support decoding that generalizes across different mice, and even demonstrated frame reconstruction of the visual stimulus using a neural-data-guided diffusion model. They also validated the approach with nonlinear alternatives, including autoencoder-based extraction of latent dynamics and encoder-decoder architectures that align sessions without explicit Procrustes fitting. Simulation experiments underscored why the flexible hardware matters: when the researchers artificially injected probe instability, such as 20 to 30 percent neuron dropout and tuning changes mimicking a drifting stiff probe, the quality of the latent-dynamics alignment degraded significantly. Stable recording of the same neurons is not a luxury but a prerequisite for the alignment mathematics to work.

The implications reach well beyond the mouse visual cortex. Brain–computer interfaces for paralyzed patients have long battled the same twin enemies: probe migration and immune scarring on the hardware side, and slow drift in the neural code on the algorithmic side. Previous work had shown that aligning low-dimensional neural activity spaces can stabilize a brain–computer interface over days, but doing so required a stable substrate of tracked neurons, which rigid implants struggle to provide. By combining tissue-matched flexible electronics with months-long single-neuron tracking and geometric realignment of population dynamics, this study demonstrates a complete framework in which both the sensor and the decoder adapt to a changing brain rather than pretending the brain stands still. If the approach translates to clinical implants, it could mean neural interfaces that stay calibrated for years, turning representational drift from a showstopper into just another coordinate transformation.

Subject of Research: Long-term tracking and decoding of representational drift in the mouse visual cortex using flexible electrode arrays

Article Title: Realignment of representational drift in mouse visual cortex via flexible electrode arrays

Article References: Shen, H., Zhao, S., Marin-Llobet, A., Qin, S., Jiang, S., Tang, X., Lee, M., Zhang, X., Lee, J., Chen, J., & Liu, J. (2026). Realignment of representational drift in mouse visual cortex via flexible electrode arrays. Nature Biomedical Engineering. https://doi.org/10.1038/s41551-026-01780-x

Image Credits: AI Generated

DOI: 10.1038/s41551-026-01780-x

Keywords: representational drift, flexible electrode arrays, brain-computer interface, visual cortex, hippocampus, neural decoding, latent dynamics, manifold alignment, single-neuron tracking, mouse neuroscience, neural probes, Nature Biomedical Engineering