Hi, I'm Aniket Vashishtha, a second-year thesis-based master's student in Computer Science at UIUC advised by Prof. Hao Peng and Prof. Chenhao Tan. My research sits at the intersection of Large Language Models, Causality, Reinforcement Learning, and Cognitive Science. I study the cognitive gaps that limit LLMs on reasoning tasks like counterfactual simulation, using causal principles to analyze failures and design faithful evaluations. I am exploring RL as an interventional learning paradigm to address these weaknesses and to help models build more accurate causal world models, with applications to self improvement, information seeking and scientific discovery. I am looking for PhD positions for Fall 2026.

Before UIUC, I was a Predoctoral Research Fellow at Microsoft Research India, where I was fortunate to be mentored by Dr. Amit Sharma and Prof. Vineeth N. Balasubramanian. At MSR, I investigated how to leverage the domain expertise and causal reasoning capabilities of imperfect experts such as humans and LLMs for causal graph discovery and effect estimation. I designed and evaluated querying strategies along with integration methods that optimally combine expert knowledge (of LLMs and humans) with traditional causal discovery approaches. I further led the development of Axiomatic Training, a framework for improving LLMs’ causal reasoning abilities by training them on synthetic data reflecting Judea Pearl’s axioms of causality. Through this work, we showed how a small decoder transformer (about 67 million parameters) can outperform billion-parameter models like GPT-4 on applying causal rules of transitivity and d-separation, and generalize to significantly more complex causal structures than those seen during training. My work has been covered by multiple newsletters, media handles, interviews, and a podcast.

Earlier, I also led the development of projects in Fairness and Responsible AI (RAI), working with Prof. Monojit Choudhury and Dr. Sunayana Sitaram. I led the development of extensive, human-centered benchmarks for evaluating socio-cultural biases in text generation models used for autosuggestion, and for language models in multilingual settings, developed collaboratively with human annotators to extend RAI techniques to underrepresented languages, particularly in the Indian context for building robust bias mitigation strategies.

Outside of research, I've been a passionate beatboxer for over 10 years, winning several national championships. I also enjoy working out, cooking, watching anime (Vinland Saga is my current favorite), movies, and reading mythology and philosophy.

For more details about my work, check my CV or hit me up on my email.

Updates

Publications