arXiv subjects

Satvik Lolla

Publications and source records attributed to Satvik Lolla.

4 recordsLinked to original sources

arXiv2024

Assessing biomedical knowledge robustness in large language models by query-efficient sampling attacks

The increasing depth of parametric domain knowledge in large language models (LLMs) is fueling their rapid deployment in real-world applications. Understanding model vulnerabilities in high-stakes and knowledge-intensive tasks is essential for quantifying the trustworthiness of model predictions and regulating their use. The recent discovery of named entities as adversarial examples (i.e. adversarial entities) in natural language processing tasks raises questions about their potential impact on the knowledge robustness of pre-trained and finetuned LLMs in high-stakes and specialized domains. We examined the use of type-consistent entity substitution as a template for collecting adversarial entities for billion-parameter LLMs with biomedical knowledge. To this end, we developed an embedding-space attack based on powerscaled distance-weighted sampling to assess the robustness of their biomedical knowledge with a low query budget and controllable coverage. Our method has favorable query efficiency and scaling over alternative approaches based on random sampling and blackbox gradient-guided search, which we demonstrated for adversarial distractor generation in biomedical question answering. Subsequent failure mode analysis uncovered two regimes of adversarial entities on the attack surface with distinct characteristics and we showed that entity substitution attacks can manipulate token-wise Shapley value explanations, which become deceptive in this setting. Our approach complements standard evaluations for high-capacity models and the results highlight the brittleness of domain knowledge in LLMs.

cs.CL↗

arXiv2023

Uncertainty-aware Language Modeling for Selective Question Answering

We present an automatic large language model (LLM) conversion approach that produces uncertainty-aware LLMs capable of estimating uncertainty with every prediction. Our approach is model- and data-agnostic, is computationally-efficient, and does not rely on external models or systems. We evaluate converted models on the selective question answering setting -- to answer as many questions as possible while maintaining a given accuracy, forgoing providing predictions when necessary. As part of our results, we test BERT and Llama 2 model variants on the SQuAD extractive QA task and the TruthfulQA generative QA task. We show that using the uncertainty estimates provided by our approach to selectively answer questions leads to significantly higher accuracy over directly using model probabilities.

cs.CL↗

arXiv2021

A Semi-Supervised Approach for Automatic Crystal Structure Classification

The structural solution problem can be a daunting and time consuming task. Especially in the presence of impurity phases, current methods such as indexing become more unstable. In this work, we apply the novel approach of semi-supervised learning towards the problem of identifying the Bravais lattice and the space group of inorganic crystals. Our semi-supervised generative deep learning model can train on both labeled data -- diffraction patterns with the associated crystal structure -- and unlabeled data, diffraction patterns that lack this information. This approach allows our models to take advantage of the troves of unlabeled data that current supervised learning approaches cannot, which should result in models that can more accurately generalize to real data. In this work, we classify powder diffraction patterns into all 14 Bravais lattices and 144 space groups (we limit the number due to sparse coverage in crystal structure databases), which covers more crystal classes than other studies. Our models also drastically outperform current deep learning approaches for both space group and Bravais Lattice classification using less training data.

cond-mat.mtrl-sci↗

arXiv2019

Tuning the Catalytic Properties of Monolayer MoS2 through doping and sulfur vacancies

Fuel cells in vehicles are the leading cause of carbon monoxide emissions. CO is one of the most dangerous gases in the atmosphere, as it binds to the hemoglobin in blood cells 200 times easier than O2. As the amount of CO in the blood stream increases, the level of oxygen decreases, which can lead to many neurological problems. To reduce the amount of CO in the atmosphere, scientists have focused on the adsorption of oxygen. The best substrates used today are platinum and palladium monolayers, which are very expensive. Because of this, researchers have searched for cheap materials, such as MoS2, that are able to adsorb oxygen. However, sulfur is a chemically inert site for the oxygen, which greatly decreases the catalytic potential of monolayer MoS2 sheets. Therefore, we carried out first-principles calculations to study the effect of substitutional doping and creating sulfur vacancies on the catalytic properties of MoS2. We calculated the adsorption energy of O on doped MoS2 sheets with vacancies, and compared it to the adsorption energy of O on a Pd monolayer. We found that doping MoS2 with Ir, Rh, Co and Fe significantly decreased the adsorption energy, to below -4 eV, indicating that doped MoS2 is a more effective catalyst than Pd. Incorporating sulfur vacancies into the doped MoS2 sheet was extremely effective, and decreased the adsorption energy below -6 eV. Our results show that iridium is the best catalyst as it has the lowest adsorption energy before and after sulfur vacancies were induced. We concluded that a combination of doping and creating vacancies in monolayer MoS2 sheets can greatly impact the catalytic behavior and make it a more effective, less expensive catalyst than Pt and Pd.

cond-mat.mtrl-sci↗