Advaith Malladi

University of North Carolina at Chapel Hill

Advisor: Prof. Shashank Srivastava

International Institute of Information Technology - Hyderabad

• Introduction to NLP
• Advanced NLP
• Information Retrieval and Extraction
• Cognitive Science and AI
• Computational Linguistics
• Computational Psycholinguistics

Research Fellow

Under the guidance of Prof. Abhilasha Ravichander, I am working on understanding how conflicting behaviors such as creativity and factuality interact in LLMs, and training Mixture-of-Experts LLMs to dynamically alternate between such behaviors at a token level.

Undergraduate Researcher

Under the guidance of Prof. Shashank Srivastava, I have worked on generating faithful natural language explanations to explain differences between model pairs (Accepted at EMNLP 2025), training latent space monitors to detect hallucinations and deception in LLMs, and training language models to generate explanations directly optimized for behavioral using RLHF techniques such as GRPO (Accepted at ICML 2026).

Undergraduate Researcher

Under the guidance of Prof. Radhika Mamidi, I have worked on utilizing Natural Language Inference to detect hallucinations in Definition Modelling, Machine Translation, and Paraphrase Generation (SemEval 2024, 2025). I am currently working on the generalizability of latent space monitors in merged language models and on techniques to merge latent space monitors.

Machine Learning Research Intern

I worked on utilizing the planning capabilities of LLMs to generate a dataset of synthetic contact-center call transcripts. I utilized these synthetic call transcripts to train smaller models and evaluated their performance compared to models trained on real human-generated call transcripts.

Machine Learning Intern

I collaborated with the NLP team at Subtl.AI to build the in-house Retrieval Augmented Generation (RAG) pipeline. I also created synthetic datasets using LLMs to evaluate the retrieval step of the RAG pipeline.

Teaching Assistant

• 2025  |  CS4.406 Information Retrieval and Extraction
• 2025  |  CL2.404 Computational Psycholinguistics
• 2024  |  CS7.501 Advanced Natural Language Processing
• 2024  |  CS7.401 Introduction to Natural Language Processing
• 2023  |  CS1.301 Algorithm Analysis and Design

Retrieval Guided Code Generation Leveraging the Planning Capabilities of LLMs

• We used retrieval-augmented generation (RAG) to improve code generation by providing relevant examples.
• We incorporated LLM planning capabilities to generate structured pseudocode before actual code generation.
• We found that retrieval with planning significantly improves code accuracy and coherence compared to zero-shot generation.
• Link: Code Generation

Prompt-tuned vs Fine-tuned models Which Better Account for Brain Language (and Vision) Representations?

• We used decoder language models to compare which representations better account for the brain's representations: fine-tuned or prompt-tuned.
• We extended the same idea to vision encoder models, implementing vision prompt tuning along the way.
• We learned that prompt-tuned text representations are similar to the brain, while fine-tuned vision representations are closer to the brain.
• Link: Fine Tune vs Prompt Tune

Retrieval Augmented Multimodal Factual Verification

• Given a claim, we built a pipeline to retrieve the relevant paragraphs, tables and images from Wikipedia. These serve as Evidence.
• After the retrieval of Evidence, we made use of two kinds of models for the claim verification part: FactBERT+DINO and Bridge Tower (presented in AAAI'23).
• Link: Multimodal Factual Verification

Parameter Efficient Prompt Tuning of GPT-2

• Implemented Parameter-Efficient Prompt Tuning on GPT-2 for Question Answering, Machine Translation and Summarization.
• Link: Parameter Efficient Prompt Tuning

Attention is all you Need

• Implemented the Encoder-Decoder architecture presented in the "Attention Is All You Need" paper from scratch without using the ready-made encoder-decoder modules.
• Build a Machine Translation System for English to French.
• Link: Eng2French

Decoder Language Model

• A decoder-only transformer based language model built from scratch using PyTorch.
• Link: Transformer Language Model

Evaluating Discourse Coherence in Paragraphs

• A stacked LSTM model which can detect topic coherence and temporal/sequential coherence in a paragraph with an accuracy of 0.8.
• This model was taught 2 kinds of discourse coherence, temporal discourse coherence and topic discourse coherence by inducing different kinds of negative sampling.
• Link: Textual Coherence

Embeddings from Language MOdelling (ELMO)

• Trained a Stacked Bi-LSTM model on Masked Language Modelling to learn contextual word embeddings for downstream tasks.
• Link: ELMO