arXiv subjects

Aakriti Shah

Publications and source records attributed to Aakriti Shah.

3 recordsLinked to original sources

arXiv2025

What the "Spotless" Mind Remembers: How Knowledge Entanglement Shapes What Leaks After Unlearning in LLMs

Unlearning in large language models (LLMs) is usually evaluated as whether an "unlearned" fact can be recovered. We instead ask whether a fact's structural entanglement with the rest of a model's knowledge predicts whether it leaks after unlearning, whether this relationship changes systematically, and whether it is causal. Across varied unlearning algorithms (WHP and GA+KL), two domains (both fictional Harry Potter and non-fictional U.S. Senators, 2000-2010), and four models (2.7B-13B parameters), we find that before unlearning, more entangled facts are recalled more often (r = +0.39 to +0.51). WHP weakens this relationship but stays positive (r = +0.16 to +0.33); GA+KL inverts it in every domain and model size (r = -0.14 to -0.25). To our knowledge, this is the first report of this specific reversal in the unlearning literature. To confirm this is causal beyond correlation, we directly manipulate a prompt's entanglement score, holding its content, target model fixed, and show recall moves in the predicted direction and reverses sign under GA+KL in the same direction as the correlational analysis. This manipulation is our central evidence that unlearning acts on the underlying knowledge structure, not just its output. Building on this, we train a predictive model that estimates a prompt's post-unlearning factuality/hallucination profile before unlearning is run, creating a unique way of model auditing in triaging which prompts are likely to leak.

cs.CL↗

arXiv2025

Sample-Efficient Language Model for Hinglish Conversational AI

This paper presents our process for developing a sample-efficient language model for a conversational Hinglish chatbot. Hinglish, a code-mixed language that combines Hindi and English, presents a unique computational challenge due to inconsistent spelling, lack of standardization, and limited quality of conversational data. This work evaluates multiple pre-trained cross-lingual language models, including Gemma3-4B and Qwen2.5-7B, and employs fine-tuning techniques to improve performance on Hinglish conversational tasks. The proposed approach integrates synthetically generated dialogues with insights from existing Hinglish datasets to address data scarcity. Experimental results demonstrate that models with fewer parameters, when appropriately fine-tuned on high-quality code-mixed data, can achieve competitive performance for Hinglish conversation generation while maintaining computational efficiency.

cs.CL↗

arXiv2023

Adversary ML Resilience in Autonomous Driving Through Human Centered Perception Mechanisms

Physical adversarial attacks on road signs are continuously exploiting vulnerabilities in modern day autonomous vehicles (AVs) and impeding their ability to correctly classify what type of road sign they encounter. Current models cannot generalize input data well, resulting in overfitting or underfitting. In overfitting, the model memorizes the input data but cannot generalize to new scenarios. In underfitting, the model does not learn enough of the input data to accurately classify these road signs. This paper explores the resilience of autonomous driving systems against three main physical adversarial attacks (tape, graffiti, illumination), specifically targeting object classifiers. Several machine learning models were developed and evaluated on two distinct datasets: road signs (stop signs, speed limit signs, traffic lights, and pedestrian crosswalk signs) and geometric shapes (octagons, circles, squares, and triangles). The study compared algorithm performance under different conditions, including clean and adversarial training and testing on these datasets. To build robustness against attacks, defense techniques like adversarial training and transfer learning were implemented. Results demonstrated transfer learning models played a crucial role in performance by allowing knowledge gained from shape training to improve generalizability of road sign classification, despite the datasets being completely different. The paper suggests future research directions, including human-in-the-loop validation, security analysis, real-world testing, and explainable AI for transparency. This study aims to contribute to improving security and robustness of object classifiers in autonomous vehicles and mitigating adversarial example impacts on driving systems.

cs.CV↗