Education

Ph.D. in Computer Science University of Massachusetts Amherst
(Expected May 2031)

M.S. in Computer Science University of Massachusetts Amherst
(May 2026)

B.Tech. in Computer Science and Engineering Vellore Institute of Technology, Chennai (May 2024)

About Me

I’m an incoming PhD student in Computer Science at UMass Amherst working on interpretable and trustworthy machine learning, with a focus on graph learning and evaluation via controlled interventions/ablations. I also work on molecular property prediction with graph-based models.

In my free time, I love to play video games and paint (need to do that more often though).

Research Interests

Trustworthy / Interpretable ML
Graph Neural Networks (GNNs) and LLM-based representation learning
Molecular graphs for bioinformatics / property prediction

Ongoing Work

Highlighting Modes of Information in Graph Learning Evaluation (September 2025 - Present)

I am developing an evaluation framework that decomposes performance of GNNs on graphs into four modes of information: node features, edge features, graph features, and graph topology/structure, using controlled interventions to test what signal a model is actually using.
Implemented controlled ablations (feature shuffling, label permutations, synthetic generators controlling structure vs feature informativeness).
Unified 10-fold stratified CV harness with per-epoch logs, learning curves, and confidence intervals across Baseline GNN models (GCN/GAT/GraphSAGE/GIN) and structure agnostic baseline (MLP).
Analysis spans over 6 datasets across domains (social + molecular benchmarks).

Causal Reasoning Benchmarking for LLMs (June 2025 - Present)

Investigating CLADDER-style causal reasoning evaluations by applying small, targeted interventions to causal queries (e.g., polarity flips, variable/value swaps, distractor injections) to test whether model behavior is robust to changes that preserve causal structure.
Built an end-to-end pipeline for intervention generation, automated scoring, and visualization, enabling large-scale comparison across multiple LLMs and intervention types.

FactCurriculum - Adaptive Verifiable Factual Training to Reduce Hallucinations in LLMs (Jan 2026 - Present)

Building RLVE-inspired environments to automatically verify “games” for factual knowledge.
Fine-tuning Baseline and SFT open-source LLMS.

Publications and Manuscripts

(Accepted at JCIM 2026) MycoPermeNet-v2: Improved Prediction of Mycomembrane Permeation via Feature Fusion and Noisy Student Self-Distillation.

Link: (Under Construction)

(Preprint, 2025) The Impact of Data Characteristics on GNN Evaluation for Detecting Fake News

Link: arXiv

(ICONAT 2023) Disseminating the Process of Hurricane Path Prediction using Multilayer Perceptron and Support Vector Machine upon Varied Kernel Functions

DOI: 10.1109/ICONAT57137.2023.10080280

Experience

Teaching Assistant - COMPSCI 589 Machine Learning, UMass Amherst (Jan 2026 – Present)

Grading and feedback for ML topics (supervised learning, optimization, evaluation metrics) with consistent rubrics and clear explanations.

Research Assistant - Science for AI Governance (S4AIG), UMass Amherst (May 2025 - Present)

Built an automated benchmark pipeline for causal reasoning in LLMs/LMMs, generating 10K+ in-context/counterfactual/interventional prompt variants and evaluating 7 models at scale.
Implemented controlled interventions (polarity flips, variable/value swaps, distractor injections, consistency checks) and intervention-aware scoring over 50K+ model responses, reducing manual evaluation substantially.
Developed analysis + visualization tooling (robustness profiles, deltas by intervention type, confusion matrices/error taxonomies) and a version-controlled, CI-backed pipeline for reproducible reporting.

Selected Projects

Evaluating the Role of Graph Structure in Fake News Detection

Investigated whether Graph Neural Networks (GNNs) genuinely leverage structural information in benchmark fake news datasets.
Replicated the UPFD framework and implemented GAT, GCN, and MLP models using PyTorch Geometric; performed controlled ablation to isolate structural vs. feature contributions.
Designed and generated custom synthetic ego-graphs with controlled feature/structure separability to validate findings beyond real-world bias.
Found small GNN gains on ego-centric graphs with highly informative node embeddings; ran feature/structure randomization + synthetic studies to show when structure truly helps.

MycoPermeNet-v2 (SAGE Lab)

Implemented Noisy Student self-distillation for Chemprop/DMPNN; built fusion pipeline combining DMPNN embeddings + RDKit descriptors with clean ablation switches.
Owned harness: scaffold-balanced splits, reproducible seeds, checkpointing, sweep execution, calibration/error analysis.

Isha Karn