The making of knowledge engines in natural language processing has been shaped by two seemingly distinct paradigms: one grounded in structure, the other driven by massively available unstructured data. The structured paradigm leverages predefined symbolic interactions, such as knowledge graphs, as priors and designs models to capture them. In contrast, the unstructured paradigm centers on scaling transformer architectures with increasingly vast data and model sizes, as seen in modern large language models. Despite their divergence, this thesis seeks to establish conceptual connections bridging these paradigms. Two complementary forces, structure and destructure, emerge across both paradigms: structure organizes seen symbolic interactions, while destructure, through periodic embedding resets, improves model plasticity and generalization to unseen scenarios. These connections form a new recipe for developing general knowledge engines that can support transparent, controllable, and adaptable intelligent systems.

Cities today generate enormous streams of data from sensors, cameras, and connected infrastructure. While this information offers unprecedented opportunities to improve urban life, most existing systems struggle with scale, latency, and fragmented insights. This work introduces a framework that blends physics-informed machine learning, multimodal data fusion, and knowledge graph representation with adaptive, rule-based intelligence powered by large language models (LLMs). Physics-informed methods ground learning in real-world constraints, ensuring predictions remain meaningful and consistent with physical dynamics. Knowledge graphs act as the semantic backbone, integrating heterogeneous sensor data into a connected, queryable structure. At the edge, LLMs generate context-aware rules that adapt filtering and decision-making in real time, enabling efficient operation even under constrained resources. Together, these elements form a foundation for digital twin systems that go beyond passive monitoring to provide actionable insights. By uniting physics-based reasoning, semantic data fusion, and adaptive rule generation, this approach opens new possibilities for creating responsive, trustworthy, and sustainable smart infrastructures.

Transformers have demonstrated strong performance across a wide range of sequence modeling tasks, but their quadratic attention complexity limits scalability to long sequences. Linear models such as Mamba and sliding-window attention (SWA) address this by mixing tokens through recurrent or localized operations with fixed-size memory, achieving efficient inference. However, these methods risk degrading performance on long sequences due to their inability to retain detailed information from distant tokens. We propose SCOUT (Segment Compression for Optimized Utility in Transformers), a hybrid architecture that compresses tokens locally within fixed-size segments and applies attention only over these compressed representations. Each token embedding is first enriched via a linear local mixer, Mamba or SWA, that integrates recent context. Then, instead of attending to all previous tokens, each token sparsely attends to a small number of compressed checkpoint tokens that summarize the input history. This design retains much of the expressivity of full attention while substantially reducing the computational and memory cost. By attending to compressed history rather than all previous tokens, SCOUT incurs slightly higher memory than purely linear models, but its growth rate remains sub-quadratic and far more scalable than that of full Transformers. We analyze SCOUT's computational and memory efficiency and evaluate it empirically on long-context language modeling and reasoning tasks. SCOUT with both Mamba and SWA mixers outperforms strong long-sequence baselines under the same computational budget, matches full-attention Transformers on language modeling and common-sense reasoning tasks at 400M and 1.3B scales. Moreover, our SCOUT achieves higher end-to-end throughput than SOTA models, while delivering comparable results on long sequence benchmarks.

Transformers achieve state-of-the-art results across many tasks, but their uniform application of quadratic self-attention to every token at every layer makes them computationally expensive. We introduce DTRNet (Dynamic Token Routing Network), an improved Transformer architecture that allows tokens to dynamically skip the quadratic cost of cross-token mixing while still receiving lightweight linear updates. By preserving the MLP module and reducing the attention cost for most tokens to linear, DTRNet ensures that every token is explicitly updated while significantly lowering overall computation. This design offers an efficient and effective alternative to standard dense attention. Once trained, DTRNet blocks routes only ~10% of tokens through attention at each layer while maintaining performance comparable to a full Transformer. It consistently outperforms routing-based layer skipping methods such as MoD and D-LLM in both accuracy and memory at matched FLOPs, while routing fewer tokens to full attention. Its efficiency gains, scales with sequence length, offering significant reduction in FLOPs for long-context inputs. By decoupling token updates from attention mixing, DTRNet substantially reduces the quadratic share of computation, providing a simple, efficient, and scalable alternative to Transformers.

The success of the generative model has gained unprecedented attention in the music generation area. Transformer-based architectures have set new benchmarks for model performance. However, their practical adoption is hindered by some critical challenges: the demand for massive computational resources and inference time, due to their large number of parameters. These obstacles make them infeasible to deploy on edge devices, such as smartphones and wearables, with limited computational resources. In this work, we present TinyMusician, a lightweight music generation model distilled from MusicGen (a State-of-the-art music generation model). The experimental results demonstrate that TinyMusician retains 93% of the MusicGen-Small performance with 55% less model size. TinyMusician is the first mobile-deployable music generation model that eliminates cloud dependency while maintaining high audio fidelity and efficient resource usage. Music, reflecting culture, social classes, ethnic identities, and historical eras, has woven itself into humanity's shared heritage through centuries of evolution (Toynbee, 2012).

Latency-critical applications tend to show low utilization of functional units due to frequent cache misses and mispredictions during speculative execution in high-performance superscalar processors. However, due to significant impact on single-thread performance, Simultaneous Multithreading (SMT) technology is rarely used with heavy threads of latency-critical applications. In this paper, we explore utilization of SMT technology to support fine-grained parallelization of latency-critical applications. Following the advancements in the development of Large Language Models (LLMs), we introduce Aira, an AI-powered Parallelization Adviser. To implement Aira, we extend AI Coding Agent in Cursor IDE with additional tools connected through Model Context Protocol, enabling end-to-end AI Agent for parallelization. Additional connected tools enable LLM-guided hotspot detection, collection of dynamic dependencies with Dynamic Binary Instrumentation, SMT-aware performance simulation to estimate performance gains. We apply Aira with Relic parallel framework for fine-grained task parallelism on SMT cores to parallelize latency-critical benchmarks representing real-world applications used in industry. We show 17% geomean performance gain from parallelization of latency-critical benchmarks using Aira with Relic framework.

Large language models (LLMs) have demonstrated exceptional proficiency in language understanding. However, when LLMs align their outputs with deceptive and / or misleading prompts, the generated responses could deviate from the de facto information. Such observations are known as fawning hallucinations, where the model prioritizes alignment with the input's implied perspective over accuracy and truthfulness. In this work, we analyze fawning hallucinations in various natural language processing tasks and tailor the so-termed contrastive decoding method for fawning-hallucination mitigation. Specifically, we design two paradigms to generate corresponding deceptive and / or misleading inputs for the consistent fawning hallucinations induction. Then, we propose the collaborative contrastive decoding (CCD) to handle the fawning hallucinations across di ff erent tasks in LLMs. By contrasting the deviation in output distribution between induced and transformed neutral inputs, the proposed CCD can reduce reliance on deceptive and / or misleading information without requiring additional training. Extensive experiments demonstrate that the proposed CCD can e ff ectively mitigate fawning hallucinations and improve the factuality of the generated responses over various tasks. Introduction Large language models (LLMs), exemplified by the Chat-GPT series [1], have demonstrated their remarkable capabilities across a wide range of natural language processing (NLP) tasks. These tasks include text translation [2, 3], summarization [4, 5], and a ffective computing [6, 7, 8, 9], showcasing the versatility and impact of artificial intelligence (AI). Despite the impressive performance, LLMs are criticized for the potential to generate fabricated, inaccurate, or incorrect information. This phenomenon, known as "hallucination", hinders the further practical application of LLMs.

Students routinely use ChatGPT and the like now to help them with their homework, such as writing an essay. It takes less effort to complete and is easier to do than by hand. It can even produce as good if not better output than the student's own work. However, there is a growing concern that over-reliance on using GenAI in this way will stifle the development of learning writing and critical thinking skills. How might this trend be reversed? What if students were required to make more effort when using GenAI to do their homework? It might be more challenging, but the additional effort involved could result in them learning more and having a greater sense of achievement. This tension can be viewed as a form of effort paradox; where effort is both viewed as something to be avoided but at the same time is valued. Is it possible to let students learn sometimes with less and other times more effort? Students are already adept at the former but what about the latter? Could we design new kinds of AI tools that deliberately require more effort to use to deepen the learning experience? In this paper, I begin to outline what form these might take, for example, asking students to use a combination of GenAI tools with traditional learning approaches (e.g. note-taking while reading). I also discuss how else to design tools to think with that augments human cognition; where students learn more the skills of metacognition and reflection.

Text-to-Image (TTI) models are powerful creative tools but risk amplifying harmful social biases. We frame representational societal bias assessment as an image curation and evaluation task and introduce a pilot benchmark of occupational portrayals spanning five socially salient roles (CEO, Nurse, Software Engineer, Teacher, Athlete). Using five state-of-the-art models: closed-source (DALLE 3, Gemini Imagen 4.0) and open-source (FLUX.1-dev, Stable Diffusion XL Turbo, Grok-2 Image), we compare neutral baseline prompts against fairness-aware controlled prompts designed to encourage demographic diversity. All outputs are annotated for gender (male, female) and race (Asian, Black, White), enabling structured distributional analysis. Results show that prompting can substantially shift demographic representations, but with highly model-specific effects: some systems diversify effectively, others overcorrect into unrealistic uniformity, and some show little responsiveness. These findings highlight both the promise and the limitations of prompting as a fairness intervention, underscoring the need for complementary model-level strategies. We release all code and data for transparency and reproducibility https://github.com/maximus-powers/img-gen-bias-analysis.

Text embedding models are essential for various natural language processing tasks, enabling the effective encoding of semantic information into dense vector representations. These models are typically optimized using triplets of (query, positive, negative) data pairs for contrastive learning, where the negative samples play a critical role in enhancing the model's ability to discern subtle semantic distinctions. In this work, we introduce a Multi-Granularity Hard-negative (MGH) synthesis framework that leverages large language models (LLMs) to generate diverse negative samples with varying levels of similarity with the query. This approach facilitates a coarse-to-fine curriculum learning strategy during supervised training, allowing the embedding model to progressively learn more nuanced semantic representations. Meanwhile, we propose an Anchor Token Aware (ATA) pooling method that assigns higher weights to anchor tokens based on aggregation patterns observed in LLMs, improving text embedding accuracy without increasing model complexity. Comprehensive experiments on the MTEB benchmark demonstrate that our methods achieve state-of-the-art performance, surpassing existing synthesis strategies both with synthetic data and when combined with public retrieval datasets.