IIT Delhi CSE / AI Research

Parth Verma

Computer Science undergraduate at IIT Delhi working across AI research, multimodal learning, graph neural force fields, and financial machine learning.

Download CV View Projects

Profile

AI research and applied machine learning.

I am pursuing a B.Tech in Computer Science and Engineering at IIT Delhi. My work spans deep learning systems, graph neural force fields, multimodal models, representation learning, and quantitative research.

Research

First author of an ICML 2026 paper on efficient merging of graph neural force fields.

Industry

Quantitative Research Intern at WorldQuant, working on global financial datasets and ML signals.

Applied ML

Built systems across multimodal learning, neural translation, computer vision, and graph mining.

Research

Selected Research

ICML 2026 Accepted publication

First Author / ICML 2026

GFFMERGE: Efficient Merging of Graph Neural Force Fields and Beyond

Parth Verma et al. / ICML 2026

Developed GFFMerge, a framework for merging specialized graph neural force fields without full joint retraining. The method aligns embeddings for compatible linear blocks and uses targeted fine-tuning to recover performance across molecular and materials benchmarks.

Achieved performance close to jointly fine-tuned gold-standard models.
Reduced training cost with 5-27x speedups across benchmark settings.
Explored model merging beyond standard neural network domains.

Projects

Model Systems and Applied ML

Deep Learning / Multimodal Systems

Multimodal Model Systems

Implemented CLIP-style image-text learning with ViT backbones for retrieval, linear probing, and representation analysis.
Evaluated retrieval through visual t-SNE diagnostics, reaching 90.12% image-to-text and 91.20% text-to-image Recall@1.
Built Qwen-based VQA with 88.74% exact match and generative VAE + latent diffusion pipelines with 3.96/9.62 FID.

GitHub repository

Computer Vision / Neural Machine Translation

Deep Learning Systems: Vision and Neural Machine Translation

Built ResNet-18 classifier with BN/IN/BIN/LN/GN, augmentation tuning, Grad-CAM, and 76.80% to 83.20% accuracy gain.
Developed English-to-Indic NMT with GloVe/BERT encoders, attention decoding, BPE tokenization, and beam search.
Added scheduled teacher forcing and cross-lingual Hindi-to-Marathi transfer for robust local Indic language translation.

GitHub repository

Vision-Language Modeling

Multimodal Transformer for Visual Question Answering

Built a multimodal VQA model with ResNet101, text encoder, and cross-attention, achieving 78.7% accuracy on CLEVR.
Enhanced model accuracy via backbone fine-tuning, Focal Loss, and BERT-based embeddings with differential learning rates.

GitHub repository

Data Mining / Graph Mining

Frequent Pattern and Graph Mining for Classification

Compared runtimes of Apriori and FP-tree algorithms for frequent itemset mining under varying minimum support thresholds.
Implemented gSpan, FSG, and Gaston for frequent subgraph mining, designing a pipeline for discriminative feature extraction.

GitHub repository

Representation Learning / CNNs

Representation Learning and CNN-Based Image Classification

Built a CNN for bird species classification, achieving 97.3% validation accuracy with augmentation and regularization.
Applied optimized CNN methods, using class activation maps to interpret and visualize the model's focus on key visual features.

GitHub repository

Experience

Internships

WorldQuant Quantitative research

University College London Research internship

May 2025 - July 2025

Quantitative Research Intern / WorldQuant

Built global financial event masks, developed ML models for future return prediction, and created a Numba-powered signal processing library with more than 60 intraday-data functions.

Received a full-time return offer based on internship performance.

May 2024 - July 2024

Research Intern / University College London

Collected and preprocessed financial tweets, built a TF-IDF sentiment dictionary, and improved sentiment classification accuracy from 70% to 78% by integrating the dictionary with VADER.

Co-authored a review paper on stock price prediction using sentiment analysis.

Contact

Feel free to reach out.

parthverma.work12@gmail.com LinkedIn GitHub CV