Anisha Gunjal

I am a Machine Learning Research Engineer at Scale AI, where I research on reasoning and alignment in large language models. My research also focuses on building robust datasets and evaluation benchmarks to promote trustworthiness and safety in LLMs. I recently completed my M.S. in Computer Science at UT Austin, where I was advised by Prof. Greg Durrett at the TAUR Lab. At UT, my research centered on fact verification and evaluation for language model outputs.

Prior to this, I was a Visiting Scholar at the Cognitive Learning for Vision and Robotics Lab (CLVR) at the University of Southern California, advised by Prof. Joseph Lim. I also served as the Lead Machine Learning Engineer for Documents AI at HyperVerge, Inc., where I built scalable OCR and information extraction systems deployed across South and Southeast Asia.

Beyond the realms of research, I love traveling, salsa dancing, and capturing the adventures of my canine companions!

Email  /  CV  /  Scholar  /  Twitter  /  Github  /  LinkedIn  /  Blog

My research lies at the intersection of post-training, alignment, and trustworthiness in large language models, with a focus on human-grounded machine learning. I develop methods that enhance the reasoning capabilities and interpretability of LLMs through post-training techniques such as RLHF, rubric-guided reward modeling, and preference optimization. A core part of my work involves constructing high-quality evaluation benchmarks and datasets that reflect real-world complexity and human values, enabling more reliable and transparent AI systems.

Introduces "Guide", a new class of reinforcement learning algorithms that integrates adaptive guidance to significantly improve reasoning in LLMs.

Studies how context affects LLM fact verification and introduces the concept of “molecular facts” to isolate and evaluate factual reasoning.

Reducing Hallucinations in LVLMs using a novel benchmark dataset M-HalDetect which is used to train fine-grained reward models capable of detecting unfaithful text generations for a given image context.

Generating and automatically evaluating lightly structures complex event schemas using large language models.

Investigation of using task information for learning representations for RL in visually complex scenes.

Joint modeling of text and layout information using transformers for Visual Question Answering on unstructured documents.

Contrastive Learning on retinal images for determining the stage of Diabetic Retinopathy disease progression on a small sized medical image dataset.

Clone this to overcome good website design FOMO!