Model-agnostic quality assessment for LLM-generated code via dynamic internal representation selection

The rapid spread of information via social media has made text-based fake news detection critically important due to its societal impact.This paper presents a novel detection method called Dynamic Representation and Ensemble Selection (DRES) for identifying fake news based solely on text.DRES leverages instance hardness measures to estimate the classification difficulty for each news article across multiple textual feature representations.By dynamically selecting the textual representation and the most competent ensemble of classifiers for each instance, DRES significantly enhances prediction accuracy.Extensive experiments show that DRES achieves notable improvements over state-of-the-art methods, confirming the effectiveness of representation selection based on instance hardness and dynamic ensemble selection in boosting performance.Codes and data are available at:

The escalating urgency of climate change mitigation has intensified corporate focus on comprehensive supply chain emission monitoring, particularly Scope 3 emissions which represent 70-90% of organizational carbon footprints. Traditional emission tracking methods prove inadequate for managing complex, multi-tiered supply networks, creating an urgent need for innovative technological solutions. This research investigates how digital twin architecture transforms Scope 3 emission transparency through dynamic virtual representations of physical supply networks. The study employs a comprehensive literature review and case analysis methodology, examining implementations across automotive, consumer goods, and technology sectors. Our four-stage framework encompasses data acquisition and integration, virtual modeling and simulation, real-time monitoring and analytics, and predictive optimization capabilities. Findings demonstrate that digital twin technology enables unprecedented supply chain visibility, facilitating proactive environmental management rather than reactive reporting. Organizations implementing digital twins report enhanced decision-making through granular emission data, improved regulatory compliance, optimized transportation routes, and dynamic supplier selection based on environmental performance. Strategic benefits encompass strengthening relationships with stakeholders, competitive market advantages, and support for setting science-based targets. However, implementation challenges relate to the complexity surrounding data integration across the almost impossibly varied supplier networks, huge infrastructure investments at the start, resistance from within regarding the transformation by technology, and privacy concerns on that information that is considered sensitive in the supply chain bureaucracy. The overall picture of adoption includes partnering collaborations, phased implementation approaches, standardized data protocols, and sharply vertical governance frameworks. This research offers new insights on technology-enabled solutions for sustainability while offering actionable advice for practitioners. Future lines of inquiry should consider an industry-oriented framework, standardization protocols, and longitudinal environmental impact assessments to empower digital twin applications for sustainable supply chain transformation.

Effective university-industry collaboration has become a major focus for governments in recent years. Universities are increasingly expected to play a greater role in the innovation system and evidence their contribution to economic development. At the same time, the growth in research quality assessment exercises makes it imperative that all opportunities to discover excellent research with the university context are uncovered. This research was conducted at a commercial design consultancy based within an academic institution, PDR (International Centre of Design and Research). The interaction between applied research and commercial consultancy within PDR led to a recognition of research potentially existing within commercially designed output. The Research questions posed within this PhD thesis are: Research Question 1: How best to recognize where commercial outputs can be characterized as research? Research Question 2: Where a research characterization is met, how best to recognize and describe the research quality? The aim of this research is to identify whether research value exists within commercial design outputs produced in an academic context and how the quality of this research could be assessed. This thesis adopts grounded action research inquiry using semi-structured interviews to explore the barriers to, and drivers of, reconciling commercial design activities with research quality in design. The thesis uncovers opportunities to capture value beyond the commercial impact, which university-industry collaboration projects are initiated for. The research led to the identification of a knowledge creation spiral resulting from the knowledge and technology transfer activities taking place within the commercial design projects in a university- industry context. In some instances, this new knowledge can be of excellent research quality if it meets the necessary research criteria. A discovered barrier to the identification of research quality within in commercial outputs was the lack of granular data on the projects thus demonstrating appropriate rigour. It was concluded that there is a need to build a documentation framework to capture the needed information and assess the research potential of commercial design projects from the outset, in order to capture the exchanged and created knowledge. A prototype of a documentation process was developed and tested with two of PDR’s lead designers. The documentation process barriers are limited to technical issues, and its implementation could assist in learning and development, as well as generating new knowledge of research quality.

The active integration of information during learning with dynamic and interactive visualisations

Occupational Application Modernization of process plants has replaced traditional machine control with sophisticated computerized control, with the work required of control room operators changing from predominantly physical to cognitive. As a result, excessive cognitive workload during an abnormal condition is not an uncommon situation. Almost 70% of accidents in process industries are the result of human error, which is largely modulated by high cognitive workload. A context-dependent methodology using pupillometry was adopted in this study for workload assessment. We found that workload has a significant effect on task performance and subsequent success in handling an abnormal condition. Incorporating contextual information along with pupillometry-based measures was useful to explore the dynamics of variations in workload in depth. In addition, the proposed methodology is a more reliable method for a dynamic environment in which the situation evolves with operator interactions with the process. Technical : Background: Operator workload monitoring has become an important aspect of system engineering due to ever increasing cognitive demands in the control room. Higher cognitive workload, beyond capability, is directly associated with human error in plant operations. Even a small increase in workload during peak demand may result in human error, which can endanger the safety of a process plant. Purpose: Subjective methods are of limited use due to the requirement for near real-time evaluation of cognitive workload. Applications such as process plant control room require the operator to reiterate and re-represent the situation based on the contextual cues from environment. Hence, our objective was to develop and test a methodology that incorporates contextual information in workload measurement using pupillometry. Method: Participants (n = 44) performed simulated process control tasks. Pupil size was recorded using an eye-tracking device. Recorded pupil size was analyzed based on a methodology that incorporates contextual information from the human-machine interface. Events that present new information about the current state of the process were categorized as critical events. Pupil size variation was adopted to provide an estimate of cognitive workload between consecutive critical events. Results: Cognitive workload was associated with process behavior and operator actions. An increase in cognitive workload for specific events was observed when participant actions deteriorated process conditions. We also observed that overall task performance was associated with the frequency of certain events and corresponding variations in pupillary behavior. Conclusion: Incorporating contextual information provides more detailed insights into the dynamics of variations in cognitive workload. The mental representation of a current situation may vary depending on expertise level. The proposed methodology assesses such dynamic mental representations in the form of frequency of critical events and corresponding changes in the cognitive workload that ultimately determines the likelihood of an operator successfully completing a task.

P249 Learning the same motor actions by distinct forms of learning (reinforcement vs. error-based) involves different physiological mechanisms

This paper describes the formulation and validation of a human gunner model suitable for inclusion into a general anti-aircraft artillery attrition algorithm. An optimal control theoretic model with a proportional-integralderivative (P-I-D) controller structure is shown to describe the human gunner's input-output characteristics in the most effective manner. Experimental data from manned AAA simulations is used in formulating and validating the P-ID optimal control gunner model. The gunner model includes target trajectory dependent parameters to account for the adaptive characteristics of the human controller, and allows for lower order internal (to the human) representations of the system dynamics (weapon system and target trajectory) to reflect realistic human psychophysical limitations. The model formulation is in terms of the tracking error and a sufficient number of its higher order derivatives.

Learning new motor behaviors or adjusting previously learned actions to account for dynamic changes in our environment requires the operation of multiple distinct motor learning processes, which rely on different neuronal substrates. For instance, humans are capable of acquiring new motor patterns via the formation of internal model representations of the movement dynamics and through positive reinforcement. In this review, we will discuss how changes in human physiological markers, assessed with noninvasive brain stimulation techniques from distinct brain regions, can be utilized to provide insights toward the distinct learning processes underlying motor learning. We will summarize the findings from several behavioral and neurophysiological studies that have made efforts to understand how distinct processes contribute to and interact when learning new motor behaviors. In particular, we will extensively review two types of behavioral processes described in human sensorimotor learning: (1) a recalibration process of a previously learned movement and (2) acquiring an entirely new motor control policy, such as learning to play an instrument. The selected studies will demonstrate in-detail how distinct physiological mechanisms contributions change depending on the time course of learning and the type of behaviors being learned.

A dual-loop model of the human controller in single-axis compensatory tracking tasks is introduced. This model possesses an inner-loop closure that involves feeding back that portion of controlled element output rate that is due to control activity. A novel feature of the model is the explicit appearance of the human's internal representation of the manipulator-controlled element dynamics in the inner loop. The sensor inputs to the human controller are assumed to be system error and control force. The former can be sensed via visual, aural, or tactile displays, whereas the latter is assumed to be sensed in kinesthetic fashion. A set of general adaptive characteristics for the model is hypothesized, including a method for selecting simplified internal models of the manipulator-controlled element dynamics. It is demonstrated that the model can produce controller describing functions that closely approximate those measured in four laboratory tracking tasks in which the controlled element dynamics vary considerably in terms of ease of control. An empirically derived expression for the normalized injected error remnant spectrum is introduced.

This research presents a new paradigm of an interactive locomotor simulator (ILS) in order to investigate how humans learn to interact and modify the locomotor pattern of another individual. Manual gait rehabilitation is imperfect at complete restoration of locomotor function in patients with stroke. Physical therapy poses a unique learning challenge, as it involves simultaneous sensorimotor co-adaptation and co-learning between one or more therapists and a patient. While patient motor learning has been extensively investigated, how a therapist adapts and learns to manipulate the gait pattern of a patient has been largely unexplored. Multiple motor learning theories exist that explain various ways this may occur, but confounds resulting from therapist-patient co-adaptation make it difficult to probe therapist learning in a clinical setting. We used the ILS to control for these confounds and hypothesized that therapist learning follows the internal model hypothesis. The internal model hypothesis postulates that through interaction with novel physical dynamics a human develops the ability to predict the behavior of the system and extrapolate to subsequent interactions. In the ILS, a user interacted with a computerized virtual patient (VP) via a robotic manipulandum in order to modify the gait pattern. Manipulandum endpoint impedances transformed from joint-space rigid body dynamics and stiffness create a force field that makes the user feel as if they are holding onto the ankle of a real patient. The user must learn how to apply the correct forces to the leg of the VP to restore a healthy gait pattern. Examination of user responses to random catch trials, the presence of after effects, and the ability of users to generalize were analyzed to assess internal model formation. Results suggest that humans do create internal models of external locomotor dynamics.

In this work, we address the problem of action recognition in a multi-view environment. Most of the existing approaches utilize pose information for multi-view action recognition. We focus on RGB modality instead and propose an unsupervised representation learning framework, which encodes the scene dynamics in videos captured from multiple viewpoints via predicting actions from unseen views. The framework takes multiple short video clips from different viewpoints and time as input and learns an holistic internal representation which is used to predict a video clip from an unseen viewpoint and time. The ability of the proposed network to render unseen video frames enables it to learn a meaningful and robust representation of the scene dynamics. We evaluate the effectiveness of the learned representation for multi-view video action recognition in a supervised approach. We observe a significant improvement in the performance with RGB modality on NTU-RGB+D dataset, which is the largest dataset for multi-view action recognition. The proposed framework also achieves state-of-the-art results with depth modality, which validates the generalization capability of the approach to other data modalities. The code is publicly available at https://github.com/svyas23/cross-view-action .

A step at a time: Preliterate children’s simulation of narrative movement during story comprehension

Prediction is argued to be a key feature of human cognition, including in syntactic processing. Prediction error has been linked to dynamic changes in syntactic representations in theoretical models of language processing. This mechanism is termed error-based learning. Evidence from syntactic priming research supports error-based learning accounts; however, measuring prediction error itself has not been a research focus. Here we present a study exploring the use of pupillometry as a measure of prediction error during syntactic priming. We found a larger pupil response to the more complex and less expected passive structure. In addition, the pupil response predicted priming while being weakly dependent on changes in expectations over the experiment. We conclude that the pupil response is not only sensitive to syntactic complexity in comprehension, but there is some evidence that its magnitude is related to the adjustment of dynamic mental representations for syntax that lead to syntactic priming.

Dynamical self-organization and efficient representation of space by grid cells.

Optic ataxia is a disorder associated with posterior parietal lobe lesions, in which visually guided reaching errors typically occur for peripheral targets. It has been assumed that these errors are related to a faulty sensorimotor transformation of inputs from the 'ataxic visual field'. However, we show here that the errors observed in the contralesional field in optic ataxia depend on a dynamic gaze-centered internal representation of reach space.