Knowledge distillation compresses a larger neural model (teacher) into smaller, faster student models by training them to match teacher outputs. However, the internal computational transformations that occur during this process remain poorly understood. We apply techniques from mechanistic interpretability to analyze how internal circuits, representations, and activation patterns differ between teacher and student. Focusing on GPT2-small and its distilled counterpart DistilGPT2, we find that student models reorganize, compress, and discard teacher components, often resulting in stronger reliance on fewer individual components. To quantify functional alignment beyond output similarity, we introduce an alignment metric based on influence-weighted component similarity, validated across multiple tasks. Our findings reveal that while knowledge distillation preserves broad functional behaviors, it also causes significant shifts in internal computation, with important implications for the robustness and generalization capacity of distilled models.

In class-incremental learning (CIL), effective incremental learning strategies are essential to mitigate task confusion and catastrophic forgetting, especially as the number of tasks $t$ increases. Current exemplar replay strategies impose $\mathcal{O}(t)$ memory/compute complexities. We propose an autoencoder-based hybrid replay (AHR) strategy that leverages our new hybrid autoencoder (HAE) to function as a compressor to alleviate the requirement for large memory, achieving $\mathcal{O}(0.1 t)$ at the worst case with the computing complexity of $\mathcal{O}(t)$ while accomplishing state-of-the-art performance. The decoder later recovers the exemplar data stored in the latent space, rather than in raw format. Additionally, HAE is designed for both discriminative and generative modeling, enabling classification and replay capabilities, respectively. HAE adopts the charged particle system energy minimization equations and repulsive force algorithm for the incremental embedding and distribution of new class centroids in its latent space. Our results demonstrate that AHR consistently outperforms recent baselines across multiple benchmarks while operating with the same memory/compute budgets. The source code is included in the supplementary material and will be open-sourced upon publication.

Mixture of experts (MoE) methods are a key component in most large language model architectures, including the recent series of DeepSeek models. Compared to other MoE implementations, DeepSeekMoE stands out because of two unique features: the deployment of a shared expert strategy and of the normalized sigmoid gating mechanism. Despite the prominent role of DeepSeekMoE in the success of the DeepSeek series of models, there have been only a few attempts to justify theoretically the value of the shared expert strategy, while its normalized sigmoid gating has remained unexplored. To bridge this gap, we undertake a comprehensive theoretical study of these two features of DeepSeekMoE from a statistical perspective. We perform a convergence analysis of the expert estimation task to highlight the gains in sample efficiency for both the shared expert strategy and the normalized sigmoid gating, offering useful insights into the design of expert and gating structures. To verify empirically our theoretical findings, we carry out several experiments on both synthetic data and real-world datasets for (vision) language modeling tasks. Finally, we conduct an extensive empirical analysis of the router behaviors, ranging from router saturation, router change rate, to expert utilization.

Federated learning is a distributed learning framework where clients collaboratively train a global model without sharing their raw data. FedAvg is a popular algorithm for federated learning, but it often suffers from slow convergence due to the heterogeneity of local datasets and anisotropy in the parameter space. In this work, we formalize the central server optimization procedure through the lens of mirror descent and propose a novel framework, called FedDuA, which adaptively selects the global learning rate based on both inter-client and coordinate-wise heterogeneity in the local updates. We prove that our proposed doubly adaptive step-size rule is minimax optimal and provide a convergence analysis for convex objectives. Although the proposed method does not require additional communication or computational cost on clients, extensive numerical experiments show that our proposed framework outperforms baselines in various settings and is robust to the choice of hyperparameters.

We present Social Digital Twinner, an innovative social simulation tool for exploring plausible effects of what-if scenarios in complex adaptive social systems. The architecture is composed of three seamlessly integrated parts: a data infrastructure featuring real-world data and a multi-dimensionally representative synthetic population of citizens, an LLM-enabled agent-based simulation engine, and a user interface that enable intuitive, natural language interactions with the simulation engine and the artificial agents (i.e. citizens). Social Digital Twinner facilitates real-time engagement and empowers stakeholders to collaboratively design, test, and refine intervention measures. The approach is promoting a data-driven and evidence-based approach to societal problem-solving. We demonstrate the tool's interactive capabilities by addressing the critical issue of youth school dropouts in Kragero, Norway, showcasing its ability to create and execute a dedicated social digital twin using natural language.

Many-shot jailbreaking (MSJ) is an adversarial technique that exploits the long context windows of modern LLMs to circumvent model safety training by including in the prompt many examples of a "fake" assistant responding inappropriately before the final request. With enough examples, the model's in-context learning abilities override its safety training, and it responds as if it were the "fake" assistant. In this work, we probe the effectiveness of different fine-tuning and input sanitization approaches on mitigating MSJ attacks, alone and in combination. We find incremental mitigation effectiveness for each, and show that the combined techniques significantly reduce the effectiveness of MSJ attacks, while retaining model performance in benign in-context learning and conversational tasks. We suggest that our approach could meaningfully ameliorate this vulnerability if incorporated into model safety post-training.

This research offers a unique evaluation of how AI systems interpret the digital language of Generation Alpha (Gen Alpha, born 2010-2024). As the first cohort raised alongside AI, Gen Alpha faces new forms of online risk due to immersive digital engagement and a growing mismatch between their evolving communication and existing safety tools. Their distinct language, shaped by gaming, memes, and AI-driven trends, often conceals harmful interactions from both human moderators and automated systems. We assess four leading AI models (GPT-4, Claude, Gemini, and Llama 3) on their ability to detect masked harassment and manipulation within Gen Alpha discourse. Using a dataset of 100 recent expressions from gaming platforms, social media, and video content, the study reveals critical comprehension failures with direct implications for online safety. This work contributes: (1) a first-of-its-kind dataset capturing Gen Alpha expressions; (2) a framework to improve AI moderation systems for youth protection; (3) a multi-perspective evaluation including AI systems, human moderators, and parents, with direct input from Gen Alpha co-researchers; and (4) an analysis of how linguistic divergence increases youth vulnerability. Findings highlight the urgent need to redesign safety systems attuned to youth communication, especially given Gen Alpha reluctance to seek help when adults fail to understand their digital world. This study combines the insight of a Gen Alpha researcher with systematic academic analysis to address critical digital safety challenges.

Neural networks have become ubiquitous with guitar distortion effects modelling in recent years. Despite their ability to yield perceptually convincing models, they are susceptible to frequency aliasing when driven by high frequency and high gain inputs. Nonlinear activation functions create both the desired harmonic distortion and unwanted aliasing distortion as the bandwidth of the signal is expanded beyond the Nyquist frequency. Here, we present a method for reducing aliasing in neural models via a teacher-student fine tuning approach, where the teacher is a pre-trained model with its weights frozen, and the student is a copy of this with learnable parameters. The student is fine-tuned against an aliasing-free dataset generated by passing sinusoids through the original model and removing non-harmonic components from the output spectra. Our results show that this method significantly suppresses aliasing for both long-short-term-memory networks (LSTM) and temporal convolutional networks (TCN). In the majority of our case studies, the reduction in aliasing was greater than that achieved by two times oversampling. One side-effect of the proposed method is that harmonic distortion components are also affected. This adverse effect was found to be model-dependent, with the LSTM models giving the best balance between anti-aliasing and preserving the perceived similarity to an analog reference device.

Graph Neural Networks (GNNs) are powerful tools for learning on structured data, yet the relationship between their expressivity and predictive performance remains unclear. We introduce a family of premetrics that capture different degrees of structural similarity between graphs and relate these similarities to generalization, and consequently, the performance of expressive GNNs. By considering a setting where graph labels are correlated with structural features, we derive generalization bounds that depend on the distance between training and test graphs, model complexity, and training set size. These bounds reveal that more expressive GNNs may generalize worse unless their increased complexity is balanced by a sufficiently large training set or reduced distance between training and test graphs. Our findings relate expressivity and generalization, offering theoretical insights supported by empirical results.

Timely prediction of students at high risk of dropout is critical for early intervention and improving educational outcomes. However, in offline educational settings, poor data quality, limited scale, and high heterogeneity often hinder the application of advanced machine learning models. Furthermore, while educational theories provide valuable insights into dropout phenomena, the lack of quantifiable metrics for key indicators limits their use in data-driven modeling. Through data analysis and a review of educational literature, we identified abrupt changes in student behavior as key early signals of dropout risk. To address this, we propose the Dual-Modal Multiscale Sliding Window (DMSW) Model, which integrates academic performance and behavioral data to dynamically capture behavior patterns using minimal data. The DMSW model improves prediction accuracy by 15% compared to traditional methods, enabling educators to identify high-risk students earlier, provide timely support, and foster a more inclusive learning environment. Our analysis highlights key behavior patterns, offering practical insights for preventive strategies and tailored support. These findings bridge the gap between theory and practice in dropout prediction, giving educators an innovative tool to enhance student retention and outcomes.