Abhinav Rao

Evaluated jailbreak effectiveness against 9 different LLMs by formalizing LLM jailbreaking, showing an inverse scaling trend where GPT-3.5 is 20% more susceptible than FLAN-T5. [Lrec-CoLING'24] (Coverage: TCS Research Webinar on Genrative AI).

Developing a theoretical framework to explain the jailbreak-paradox, explaining the inverse scaling phenomenon in toxicity/jailbreaking. (Work-in-progress covered by Analytics IndiaMag).

Improved Bing Chat classifier performance by 5% and 17% (F1-score) for jailbreaking and content-harm detection through offline data curation.

Designed a framework to evaluate the ethical reasoning capabilities of Language models over increasing granularities of ethical policies. Uncovered a bias favoring western centric ethical principles in GPT-4. [EMNLP Findings '23] [Keynote at WiNLP '23]

Built a benchmark dataset of 2.6k cultural situations spanning 75 countries measuring cultural biases in LLMs

Measured cultural adaptability of 17 language models, determining strong sycophancy and western-centric biases. [Accepted at NAACL'25] [Presented at C3NLP, ACL '24]

Compared and contrasted different parameter-efficient finetuning (PEFT) techniques, such as sparse subnets and LoRA for machine translation in federated learning [Presented at MOOMIN, EACL '24] [Accepted at CustomNLP4U, EMNLP '24]