Global ETD Search

Search theses and dissertations gathered from participating repositories worldwide. Every result links back to the library that holds it. No account is needed.

Results

Showing 1 to 20 of 22 for “"mechanistic interpretability"”.

  1. Practical Methods for the Mechanistic Interpretability of Large Language Models

    Los Large Language Models (LLM) han demostrado una capacidad asombrosa en una amplia gama de tareas relacionadas con el lenguaje, lo que los convierte en herramientas de gran interés. Sin embargo, debido a su elevado número de parámetros e interacciones internas, este tipo de modelos se comporta …

    alicante Repository record for Practical Methods for the Mechanistic Interpretability of Large Language Models (opens in a new tab)

  2. Robustness and Safety in AI Systems: Adaptive Quantile Recalibration for Test-Time Adaptation and Mechanistic Interpretability of LLM Jailbreaking

    … vulnerabilities in large language models through mechanistic interpretability. Even with safety training in place, LLMs continue to be vulnerable to jailbreaking attacks that can produce harmful outputs. This thesis explores circuit discovery techniques to identify specific subnetworks that play a …

    concordia Download PDF for Robustness and Safety in AI Systems: Adaptive Quantile Recalibration for Test-Time Adaptation and Mechanistic Interpretability of LLM Jailbreaking (opens in a new tab)

  3. Understanding and the Black Box Problem in Artificial Intelligence

    … comprehensible to human stakeholders. I focus on mechanistic interpretability, an approach which employs functional decomposition to characterise the internal structure of AI systems. Mechanistic interpretability can deliver detailed explanations, but at significant costs in time and labour. …

    bayreuth Repository record for Understanding and the Black Box Problem in Artificial Intelligence (opens in a new tab)

  4. Stochastic Universal Differential Equations for Epidemiological Modeling: Uncertainty Quantification in Disease Transmission Dynamics

    … probabilistic forecasts while preserving the mechanistic interpretability and computational efficiency of the deterministic baseline. In doing so, a two-phase training algorithm is introduced to ensure stable convergence and the framework is validated against the ensemble output from ExaEpi, …

    etsu Repository record for Stochastic Universal Differential Equations for Epidemiological Modeling: Uncertainty Quantification in Disease Transmission Dynamics (opens in a new tab) Download PDF for Stochastic Universal Differential Equations for Epidemiological Modeling: Uncertainty Quantification in Disease Transmission Dynamics (opens in a new tab)

  5. Integration of Mechanism-Driven Computational Modeling and Public Data Resources for Chemical Toxicity Assessment

    … on animal testing, which limits throughput, mechanistic interpretability, and applicability to human-relevant endpoints. This dissertation presents computational approaches that integrate publicly available bioactivity data with machine learning to support chemical safety evaluation. Oral …

    rowan Repository record for Integration of Mechanism-Driven Computational Modeling and Public Data Resources for Chemical Toxicity Assessment (opens in a new tab)

  6. Low-Resource Ecoacoustic Audio Classification

    … Multimodal Large Language Models (MLLMs) using mechanistic interpretability, revealing fundamental limitations in audio-text reasoning. Collectively, these contributions offer validated methods, unique datasets, and critical insights for robust ecoacoustic monitoring and broader challenges in …

    cuny-grad Repository record for Low-Resource Ecoacoustic Audio Classification (opens in a new tab)

  7. Artificial Intelligence Tools for Comprehensive Environmental Chemical Risk Assessment

    … In addition, many deep learning models lack mechanistic interpretability. Moreover, linking biotransformation reactions to specific enzymes is essential for ultimately predicting environment-specific biotransformation pathways. However, training models for assigning enzymes to orphan …

    zurich

  8. Deep Learning-Based Incident Dementia Prognosis from Wearable ECG and Accelerometer Signals

    … for identifying the minority class. Finally, mechanistic interpretability via linear probing confirms that the model's latent space encodes clinically relevant markers, including resting heart rate, rMSSD, M10, and relative amplitude, without explicit supervision. Together, these results …

    umn Repository record for Deep Learning-Based Incident Dementia Prognosis from Wearable ECG and Accelerometer Signals (opens in a new tab)

  9. Understanding Implicit Content in Language and Language Models

    … contradict it. Using probing techniques from mechanistic interpretability, we find that model representations themselves are swayed by contradictory partner assertions. We then study these failures in LLMs designed for real-world deployment. In a maternal health chatbot, identifying implicit …

    maryland Repository record for Understanding Implicit Content in Language and Language Models (opens in a new tab)

  10. Understanding Cross-Model Layer Correspondences: From Interpretability to Distillation

    This thesis investigates whether language models of different scales within the same family develop structurally similar internal representations, and whether this structure can be leveraged to improve knowledge distillation. To study this, we employ PairLens, a framework that extends the Logit …

    umn Repository record for Understanding Cross-Model Layer Correspondences: From Interpretability to Distillation (opens in a new tab)

  11. Efficient and human-inspired natural language processing methods for multilingual and low-resource settings

    … the dissertation shifts focus to human-inspired interpretability and mechanistic understanding of language models. By integrating psycholinguistic and neurolinguistic probing paradigms, it reveals a persistent gap between model performance and true linguistic competence, with LLMs demonstrating …

    lmu-germany Repository record for Efficient and human-inspired natural language processing methods for multilingual and low-resource settings (opens in a new tab)

  12. A HYBRID FRAMEWORK FOR MODELING SURROUNDING-VEHICLE BEHAVIOR FOR EVALUATING AUTONOMOUS VEHICLE PERFORMANCE IN RISK-CRITICAL DRIVING SCENARIOS

    … behavior is represented with sufficient realism, interpretability, and behavioral richness. Existing modeling paradigms address this need only partially: classical microscopic models provide compact structure, interpretable parameters, and simulation-friendly rollout, but may become behaviorally …

    maryland Repository record for A HYBRID FRAMEWORK FOR MODELING SURROUNDING-VEHICLE BEHAVIOR FOR EVALUATING AUTONOMOUS VEHICLE PERFORMANCE IN RISK-CRITICAL DRIVING SCENARIOS (opens in a new tab)

  13. Interpreting Deep Learning Models and Unlocking New Applications With It

    … heads, using text descriptions. We apply these interpretability insights to (i) mitigate spurious correlations, (ii) enable zero-shot segmentation, and (iii) facilitate text or image-conditioned image retrieval. We also extend our mechanistic interpretability techniques to understand and control …

    maryland Repository record for Interpreting Deep Learning Models and Unlocking New Applications With It (opens in a new tab)

  14. Robust and Interpretable Medical Image Segmentation via Retrieval and Mechanistic Analysis

    … that target both the learning dynamics and the interpretability of medical image segmentation models. In the first part, we propose Joint Retrieval-Augmented Segmentation (J-RAS), a retrieval-guided framework that reduces dependence on large annotated datasets and improves robustness under …

    wilfrid-laurier Repository record for Robust and Interpretable Medical Image Segmentation via Retrieval and Mechanistic Analysis (opens in a new tab) Download PDF for Robust and Interpretable Medical Image Segmentation via Retrieval and Mechanistic Analysis (opens in a new tab)

  15. Linear recurrent models for robust and interpretable long-context modeling

    … development of scalable architectures and the interpretability of their underlying recurrence mechanisms. In the context of scalable architectures, we adapt the Mamba linear recurrent byte-level sequence modeling, tackling the computational challenges posed by extended sequence lengths. By …

    cornell Repository record for Linear recurrent models for robust and interpretable long-context modeling (opens in a new tab)

  16. Sociotechnical controls for mitigating AI risk in the absence of ground truth

    Organizations increasingly rely on AI systems in consequential settings where no single "ground truth" exists. Without a consensus ground truth, traditional AI risk management frameworks, such as evaluating risk via held-out accuracy, may mask underlying issues related to bias or label drift. This …

    texas Repository record for Sociotechnical controls for mitigating AI risk in the absence of ground truth (opens in a new tab)

  17. Locating and interpreting factual association in Speech Language Models

    Recent advances have enabled Speech Language Models (SLMs) to both understand and generate text and speech by representing audio as discrete tokens learned from raw waveforms without supervision. As these multimodal systems become increasingly common in real-world applications, it is crucial to …

    chalmers Repository record for Locating and interpreting factual association in Speech Language Models (opens in a new tab)

Page 1 of 2