Images Amplify Misinformation Sharing in Vision-Language Models
Images are known to increase the perceived credibility of information in humans, but do they have the same effect on AI systems? This work investigated whether vision-language models are more likely to share misinformation when news is accompanied by visual content.
To answer this question, four state-of-the-art vision-language models were evaluated on a dataset of fact-checked political news, comparing their willingness to share the same content with and without accompanying images. The study also explored how induced personality traits influence sharing behavior, drawing inspiration from psychological research on misinformation susceptibility.
The results revealed a significant parallel with human cognition: images consistently increased the propensity of models to share news, with a substantially stronger effect for false information than for true information. Personality traits further modulated this behavior, with profiles associated with manipulative or antisocial tendencies showing the highest rates of misinformation sharing. These findings suggest that multimodal AI systems inherit some of the same vulnerabilities that make visual misinformation effective in human communication.
Reference: [Plebe et al., 2026].
Propensity to Trust in Large Language Models
Trust is essential for collaboration, but little is known about how language models decide when to rely on other agents. This work investigated whether LLMs exhibit a stable propensity to trust, a concept from psychology describing an individual's baseline tendency to delegate responsibility under uncertainty.
Nineteen language models were evaluated using two complementary approaches. First, models completed a standard psychological questionnaire designed to measure trust propensity in humans. Second, they participated in language-based collaborative simulations in which they had to decide whether to entrust tasks to other agents and update their beliefs based on observed performance.
The results revealed a striking discrepancy. Questionnaire responses suggested that most models exhibit similarly high levels of trust, likely reflecting alignment objectives and socially desirable answering patterns. In contrast, the behavioral simulations uncovered substantial differences in how models actually allocate trust. Some models consistently over-trusted collaborators, others remained overly cautious, while the most capable systems adjusted their trust dynamically according to evidence about competence and reliability.
These findings suggest that LLMs' propensity to trust cannot be understood through self-reported attitudes alone. Instead, trust emerges from the interaction between stable behavioral tendencies and the ability to integrate evidence about collaborators over time. Reference: [Plebe, 2026].
Human-Inspired Autonomous Driving
Autonomous driving research increasingly draws inspiration from human cognition, yet the notion of a "human-inspired" driving agent is often used without a clear definition. This work examined how concepts from cognitive science, neuroscience, and psychology have been translated into autonomous driving systems.
By reviewing the literature across perception, decision making, navigation, control, and learning, the survey identified a wide spectrum of human-inspired approaches, ranging from behavioral imitation and cognitive functions to brain architectures and neural mechanisms. To organize this diverse landscape, the work proposed a taxonomy based on both the level of biological abstraction and the stage at which the inspiration is incorporated into the artificial system.
The analysis highlights the opportunities and limitations of different forms of human inspiration, showing how cognitive and neurobiological principles can complement traditional engineering approaches. More broadly, the survey argues that understanding how humans drive may provide valuable guidance for the development of future autonomous vehicles, particularly as the field moves toward more adaptive, interpretable, and socially aware driving systems. Reference: [Plebe et al., 2024].
Neuro-inspired Circular Latent Space for Rotation Estimation
Estimating the orientation of an object from visual input is a fundamental problem in computer vision and robotics. This work investigated how prior knowledge about the structure of rotational transformations can be embedded directly into a neural network architecture.
Inspired by neural circuits involved in orientation representation in the ellipsoid body of Drosophila, the proposed model organizes its latent representation as a circular manifold. This design reflects the cyclic nature of rotations: a full revolution returns an object to its original state. Within this latent space, the rotation operator becomes a translation along the circular latent manifold, establishing a direct relationship between transformations in the latent representation and angular changes in the physical world.
The resulting model accurately estimates the relative rotation between two poses while generalizing to object categories not encountered during training. More broadly, the work demonstrates how biologically inspired inductive biases can improve the representation of geometric transformations in neural networks. Reference: [Plebe & Da Lio, 2023].
Distributed Cognition for Human-Vehicle Collaboration
Most autonomous driving systems treat the human driver and the artificial agent as separate decision-makers. This work explored an alternative perspective inspired by distributed cognition, viewing intelligence as emerging from the interaction between multiple cognitive entities rather than residing entirely within a single agent.
Building on this idea, a collaborative driving architecture was developed in which control is shared between the human driver and the autonomous system. The interaction follows the "rider–horse paradigm" from distributed cognition: rather than issuing explicit commands, the human communicates intentions through natural driving actions such as steering and pedal inputs, while the autonomous agent interprets these signals and contributes its own driving capabilities.
Experiments in a driving simulator demonstrated how this shared-control framework enables continuous cooperation between human and artificial driver, allowing the vehicle to combine human intentions with autonomous decision-making in real time. References: [Plebe et al., 2022], [Da Lio et al., 2022].
Reinforcement Learning for Safe Interaction with Distracted Pedestrians
Autonomous vehicles must operate safely in the presence of vulnerable and unpredictable road users. This work investigated how reinforcement learning (RL) can support decision-making in situations involving distracted pedestrians, whose future actions may be difficult to anticipate.
Rather than learning an entire driving policy from scratch, the proposed approach focused on a higher-level behavioral function that determines a safe traveling speed given the current context. A neural network was trained through RL to estimate the maximum speed that still allows the vehicle to react safely if a pedestrian unexpectedly enters its path. This learned safety layer acts as a behavioral constraint that can be integrated into an existing autonomous driving system, complementing rather than replacing conventional control and planning modules.
Experiments in simulation and on a real vehicle demonstrated that the learned safe-speed function transfers effectively from simulation to reality and enables safer interactions with distracted pedestrians in uncertain traffic scenarios. Reference: [Rosati Papini et al., 2021].
Neuro-inspired Occupancy Grid Mapping of Driving Scenarios
Occupancy grids are a fundamental representation for autonomous navigation, providing an abstract description of the surrounding environment. However, conventional occupancy grids allocate the same spatial resolution to all regions of the scene, regardless of their relevance to driving decisions.
This work investigated a neuro-inspired alternative based on cortical magnification, a property of the human visual system that dedicates a larger portion of neural resources to the central and nearby regions of the visual field. Inspired by this principle, the proposed occupancy grid representation allocates more cells to areas close to the ego vehicle, where accurate obstacle localization is most critical, while representing distant regions with progressively lower resolution.
Neural networks were trained to generate these non-uniform occupancy grids directly from monocular camera images. The resulting representation preserves more information about nearby vehicles and obstacles without increasing the overall size of the grid, providing a more efficient encoding for autonomous driving tasks. Reference: [Plebe et al., 2021].
Cognitive-inspired Visual Perception for Autonomous Driving
Autonomous driving systems must extract meaningful information from high-dimensional visual inputs while operating under strict computational constraints. This work investigated whether principles from cognitive theories of perception could guide the learning of compact and interpretable visual representations for driving tasks.
The approach combined variational autoencoders with ideas from predictive perception, which views internal representations not as passive descriptions of the environment but as tools for anticipating future states and supporting action. This perspective motivated a training procedure that encouraged the emergence of latent representations aligned with fundamental driving concepts.
Experiments showed that the resulting models could learn highly compact visual representations, requiring as few as 16 latent units to encode the two most relevant semantic categories in the driving environment: other vehicles and lane markings. The work demonstrates how insights from cognitive science can inspire efficient machine learning architectures for autonomous systems. Reference: [Plebe & Da Lio, 2020].
Interior Lighting Design with Genetic Algorithms
Designing the lighting of interior spaces requires balancing .multiple competing objectives, including visual quality, uniform illumination, and energy efficiency. This work explored how evolutionary optimization can support the automated design of lighting layouts in architectural environments.
A detailed 3D model of the target space was created in Blender and evaluated using physically-based ray tracing to simulate the interaction of light with materials and geometry. The resulting illumination metrics provided the fitness functions for a multi-objective optimization process based on a variant of the NSGA-II genetic algorithm. By iteratively exploring alternative lighting configurations, the system optimized both the placement and intensity of artificial light sources while minimizing energy consumption.
The project demonstrated how evolutionary algorithms can effectively tackle complex architectural design problems, generating Pareto-optimal lighting solutions that balance illumination quality and energy efficiency.
Reference: [Plebe & Pavone, 2017]; Video presentation: [LINK].
Simulating Fire Outbreaks in Industrial Environments
Industrial facilities such as oil refineries contain dense networks of interconnected components, where spatial relationships strongly influence the propagation and consequences of fire events. This work investigated how computational geometry techniques can support real-time risk assessment in such environments.
Starting from a detailed laser scan of a real oil refinery, the acquired point cloud was processed in Blender to generate a high-fidelity 3D model of the plant. The resulting environment combined flame dynamics with geometric algorithms for minimum-distance computation between polyhedral objects, enabling real-time estimation of the thermal exposure of nearby equipment. By continuously evaluating the spatial relationship between flames and plant components, the system identified vulnerable areas and assessed the risk of cascading failures as the fire evolved.
The resulting framework was later integrated into a broader platform for industrial risk assessment and emergency management.
Reference: [Plebe & Grasso, 2016]; Video presentation: [LINK].
