Shivank Garg

As a final-year undergraduate at IIT Roorkee, I work on building self-improving AI agents, with broad interests in LLM and VLM alignment. I am currently at Bespoke Labs, focusing on agentic optimization and reinforcement learning. Previously, I worked on video generation models at Adobe Research and conducted research at Sony AI.

Beyond my own research, I actively mentor students on AI research projects at IIT Roorkee and through Algoverse AI Research.

Feel free to reach out if you would like to collaborate on any research project or discuss any of my research.

Email  /  CV  /  Github  /  Google Scholar

Research

I'm interested in LLM and VLM Alignment, post-training RL based techniques for generative models, and diffusion models and their applications . Some papers are highlighted.

We study fairness in Large Language Models (LLMs) by examining chain-of-thought prompting. Our results show that biased outputs do not always stem from biased reasoning, distinguishing model behavior from human decision-making.

We introduce Implicit Preference Optimization (IPO), a novel framework that enhances Large Language Models (LLMs) through self-improvement. IPO leverages the model's intrinsic capability to generate continuous reward signals, which are then used to create preference pairs for fine-tuning with Direct Preference Optimization (DPO), eliminating the need for external preference models.

We introduce a technique for concept ablation, specifically targeting the removal of copyrighted logos and trademarks from images to protect intellectual property. We demonstrate the limitations of existing methods, particularly their vulnerability to jailbreaking via paraphrased prompts, and propose a more robust solution.

We introduced Attention Shift, an inference-time technique for content safety in text-to-image diffusion models utilising attention map manipulation of diffusion models to generate safe images from diffusion models while maintaining semantic consistency.

We present a data-centric approach to improve the robustness of traffic light detection models in adverse weather. We introduce a novel data curation and augmentation strategy that simulates challenging conditions like snow and fog, demonstrating a significant improvement in model generalization and performance on out-of-distribution data.

We introduce a challenging dataset and show how large vision-language models can reveal subtle generative artifacts in synthetic images, significantly improving artifact-based detection while highlighting limitations on artifact-free deepfakes.

We investigate how post-training methods shape reasoning in large language models, showing that models fine-tuned with reinforcement learning exhibit stronger awareness of learned behaviors and better generalization, yet often have weak alignment between their internal reasoning traces and final outputs.

Mentorship

Papers from projects that I have mentored.

We introduce MiSCHiEF, a minimal pair benchmark in safety and cultural contexts for fine-grained evaluation of image-caption alignment, revealing persistent misalignment challenges in current vision-language models.

We investigate the mathematical reasoning capabilities of LLMs. Systematically evaluating their performance on complex math problems when provided with hints, and further assessed their robustness under various adversarial conditions to probe the limits of their problem-solving abilities.

A survey on different positional embeddings used in Vision Transformers and their generalisation to variable resolution images.

We introduce LoRA-Mini, an efficient parameter-efficient fine-tuning (PEFT) method. By decomposing the standard LoRA adaptation matrices and applying selective training of these matrices, LoRA-Mini significantly reduces the number of trainable parameters and lowers memory consumption during fine-tuning.

We introduce a novel hybrid framework for adaptive urban planning utilising LLMs and genetic algorithms. The agentic system is designed to generate sustainable and balanced city layouts by optimizing for multiple socio-economic and environmental factors.

We introduce a novel confidence-based Membership Inference Attack (MIA). Our method, LIRA, utilizes a model's prediction confidence to infer whether a data point was included in its training set.

Volunteering

Reviewer for TMLR, ICLR, ACL, EMNLP, AAAI

GPU and SLURM Maintainer for GPU Clusters

The webpage is template is stolen from Jon Barron's website.