Projects

Uncertainty Quantification in Large Language Models

Mechanistic Uncertainty Quantification. Proposed a label-free, causal intervention-based uncertainty method for LLMs, measuring the minimum internal steering needed to change a model's output, which detects incorrect answers across 3 models and 5 QA benchmarks and complements conventional output-based uncertainty measures.

Causal robustness margin method: Jacobian Lens readout, consensus, and intervention steering an LLM's response

Uncertainty-Aware Multi-Agent Orchestration. Built an efficient, uncertainty-aware multi-agent LLM orchestration framework that routes between specialized agents using lightweight information-theoretic compression trained via group relative policy optimization (GRPO), with applications in clinical diagnosis and code generation.

Trustworthy Vision-Language Models

Causality in Vision-Language Models. Introduced CausalVLBench, a benchmark spanning three causal reasoning tasks across three annotated physical-system image datasets, and used it to evaluate 9 state-of-the-art LVLMs under zero-shot, few-shot, and chain-of-thought prompting, revealing systematic weaknesses in visual causal reasoning.

CausalVLBench's three causal reasoning tasks: structure inference, intervention target prediction, and counterfactual prediction

Privacy & Unlearning in Vision-Language Models. Designed CAGUL, a resource-efficient unlearning method that uses cross-modal attention to identify a query's least-attended visual tokens and encodes unlearning information there, matching or outperforming fine-tuning-based baselines without altering pretrained parameters or incurring retraining costs.

CAGUL architecture: cross-modal attention guided visual token selection and encoding for VLM unlearning

Causal Generative Modeling

Causal Representation Learning. Proposed ICM-VAE, a VAE framework that models causal mechanisms with nonlinear flow-based functions and a causal disentanglement prior grounded in independent causal mechanisms, outperforming state-of-the-art baselines by 15 to 20% in disentanglement metrics.

ICM-VAE framework: structural causal flow mapping noise encodings to causal representations

Diffusion-based Counterfactual Generative Modeling. Designed CausalDiffAE, a diffusion-based framework that encodes data into causally meaningful latent variables and generates counterfactuals through a DDIM-based do-intervention procedure, producing disentangled, high-quality counterfactual images that outperformed baseline methods.

CausalDiffAE architecture: causal diffusion training and counterfactual generation via intervened SCM

Causal Generative Modeling via Foundation Models. Built FM-CGM, a zero-shot pipeline pairing a reasoning VLM with Stable Diffusion XL and a new Causal Semantic Guidance mechanism to extract, manipulate, and generate causal concepts in images, outperforming DDIM and DDPM inversion baselines on edit accuracy and minimality.

FM-CGM pipeline: concept extraction, manipulation, and counterfactual generation using foundation models