Member of Technical Staff (ML) at 
making AI more human 🌻
I am currently a Member of Technical Staff (ML) at Adobe, where I work on evaluations and personalized experiences for users interacting with conversational AI through Adobe Brand Concierge. Previously, I was a Research Intern at Microsoft Research, India, working with Vineeth N. Balasubramanian and Ravishankar Krishnaswamy on pre-training multimodal embedding models for visual document retrieval. I also spent a fall at the Centre for Trusted Internet Community, National University of Singapore, where I worked under Prof. Kokil Jaidka. I have also had long-term research collaborations at the AI Institute, University of South Carolina, where I worked under Prof. Amit Sheth. I completed my Bachelor's in Computer Science and Engineering from Manipal University Jaipur, India (2022-2026).
My research is about building human-centered AI systems that align with human intent and values, capture diverse human behavior across social and cultural contexts, and remain reliable and safe as they learn, adapt, and interact with their environments. Below, I outline the research directions I pursue, their impact, and the work done around them.
Please check out my publications for more details about my work. I'm always happy to discuss research and explore collaborations. Please feel free to reach out at abhilekhxborah@gmail.com!
.
.
.
.
.
.* denotes equal contribution.
Pluralistic Alignment Workshop, International Conference on Machine Learning (ICML), 2026
A simulation in which LLM agents with different levels of authority repeatedly decide how much to extract from a shared resource, to study how unequal power shapes cooperation and collective behavior in multi agent systems.
Language Resources and Evaluation Conference (LREC), 2026
Tests whether LLMs truly understand the meaning of factual questions or rely on patterns memorized during training, by transforming questions into semantically equivalent forms that keep the answer the same.
International Conference on Learning Representations (ICLR), 2026
Inspired by the wisdom of the crowd, Social Agents creates personas with different demographic and psychographic attributes that independently judge content (like advertisements, videos, and webpages), and combines their opinions to predict how real people respond across eleven behavioral tasks.
Conference of the European Chapter of the Association for Computational Linguistics (EACL), 2026
Tests whether LLMs truly understand the underlying structure of a question or rely on patterns memorized during training, by transforming questions into logically equivalent forms that keep the answer the same.
AAAI Conference on Artificial Intelligence, 2026, Special Track on AI Alignment (AIA)
A large-scale benchmark of human-verified hateful and non-hateful image pairs across race, gender, and disability, along with DPO-Kernels, a preference-learning method for teaching text-to-image models human safety preferences.
AAAI Conference on Artificial Intelligence, 2026, Student Abstract and Poster Program
A training-free method that compares model activations for safe and unsafe requests to find a refusal direction, then applies it during inference to help models refuse harmful requests under jailbreak attacks.
International Joint Conference on Natural Language Processing and Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (IJCNLP-AACL), 2025
A benchmark of real and AI-generated images to evaluate AI-generated image detection methods, along with a metric to quantify and rank the perceptual realism of generated images.
Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025
An alignment metric that looks inside the model, measuring how distinctly safe and unsafe behaviors are represented in its activations, to assess alignment beyond output-based measures like refusal rate.
Conference on Empirical Methods in Natural Language Processing (EMNLP), 2025
A multimodal, multilingual benchmark testing whether vision-language models understand Indian culture (e.g., festivals, attire, cuisine, art forms, and heritage) across 15 languages, all 28 states, and 8 union territories.
Workshop on NLP for Positive Impact (NLP4PI), Annual Meeting of the Association for Computational Linguistics (ACL), 2025
A multimodal dataset of climate-related social-media posts (primarily memes), with a metric that checks whether model responses are supported by evidence, relevant to the climate claim, and clearly communicated.
Adobe
Member of Technical Staff (ML)
Microsoft Research
Research Intern
Supervisors: Vineeth N. Balasubramanian, Ravishankar Krishnaswamy
Adobe, Media and Data Science Research (MDSR) Lab
Research Intern
Supervisor: Yaman Kumar Singla
Please check out my CV for more details about the work I have done at each of these places.
It is a big part of who I am! I love composing and producing original songs, with over 6 releases that have collectively gained over 2 million streams worldwide across platforms. Feel free to listen on Spotify and YouTube.
I also play a variety of instruments, such as the piano, guitar, and ukulele, and have a deep appreciation for diverse genres such as rock, country, indie/synth pop, folk, and electronic music. Some of my favorite artists include Jeremy Zucker, Passenger, Kygo, The 1975, John Denver, Creedence Clearwater Revival, and John Mayer. You can check out my playlists here.
I'm an avid traveler who loves spontaneous plans, spending time outdoors, and exploring different countries and cultures. I'm also deeply into movies and TV shows, with a particular fondness for genres such as thriller, comedy, drama, and science fiction. Some of my favorite films include Zodiac, Inception, Déjà Vu, Passengers, Dead Poets Society, The Holdovers, and Beverly Hills Cop.
