I am a PhD researcher at the Applied Machine Learning Lab (AML Lab) at the University of Bonn and the Lamarr Institute for Machine Learning and Artificial Intelligence.
My research is centered around a question that sounds simple, but turns out to be surprisingly difficult:
When should a machine learning model trust its own prediction — and when should it know that it does not have enough information?
I am particularly interested in reliable and resource-efficient language models, with a focus on small and compact models that can reason, recognize uncertainty, and make better decisions under limited information or computational resources. My current work studies context sufficiency, abstention, calibration, and selective prediction, including how these behaviours emerge during training and whether the signals we can observe inside a model are actually used when it makes a decision.
A second thread of my work looks at what happens when models become smaller or more efficient. I have worked extensively on quantization and compact language models for critical error detection in machine translation, studying where compression is essentially free and where it begins to affect reliability. More broadly, I am interested in evaluation settings where aggregate accuracy alone is not enough and individual mistakes can have very different consequences.
Before moving towards language models, much of my research focused on self-supervised learning and medical imaging, particularly diabetic retinopathy screening. This continues to shape how I think about trustworthy AI: models should not only perform well, but should also communicate when their predictions are unreliable.
Across these areas, I am especially interested in models that are small enough to study carefully, efficient enough to deploy, and reliable enough to know their limits.
Research interests
- Reliable & Trustworthy Machine Learning: Understanding when models fail, when they should abstain, and how reliability can be evaluated beyond average accuracy.
- Small & Efficient Language Models: Compact models, quantization, compression, and the relationship between model efficiency and behavioural robustness.
- Context Sufficiency & Abstention: Studying whether language models can recognize when the available information is sufficient to answer, and how this signal influences their decisions.
- Mechanistic & Developmental Analysis: Investigating where reliability-related signals are represented inside neural networks, whether they are causally used, and how they emerge during training.
- Selective Prediction & Calibration: Designing systems that can defer uncertain or risky predictions instead of treating every input as equally answerable.
- Machine Learning for High-Stakes Applications: Reliable evaluation in areas such as machine translation and medical AI, where seemingly small errors can have disproportionate consequences.
Selected recent work
Knowing When Not to Predict
Self-Supervised Learning and Abstention for Safer Diabetic Retinopathy Screening
IJCAI-ECAI 2026
We study how self-supervised pretraining influences not only classification performance but also a model’s ability to identify cases on which it should abstain. The work explores selective prediction as a way of moving beyond accuracy towards safer medical AI.
Towards Reliable Machine Translation
Scaling LLMs for Critical Error Detection and Safety
ECIR 2026
We investigate how language models of different scales perform at detecting meaning-critical translation errors and examine the trade-offs between model size, reliability, and computational cost.
How Small Can You Go?
Compact Language Models for On-Device Critical Error Detection in Machine Translation
IEEE BigData 2025
This work explores how far language models can be compressed while retaining their ability to detect critical translation errors, with particular attention to parameter-efficient and quantized models.
SynCED-EnDe 2025
A Synthetic and Curated English-German Dataset for Critical Error Detection in Machine Translation
ECIR 2026
We introduce a structured benchmark for critical error detection containing fine-grained error categories designed to support more systematic evaluation of both compact and large language models.
Functional Knowledge Transfer with Self-Supervised Representation Learning
IEEE International Conference on Image Processing (ICIP), 2023
Our earlier work studied how self-supervised representations can support label-efficient knowledge transfer across domains, forming part of my broader interest in robust learning under limited supervision.
Beyond research
I enjoy being involved in the research community beyond my own projects. I have served as a reviewer for IJCAI-ECAI and IJCNN, and I have been involved in mentoring students through the MINERVA Mentoring Program at the University of Bonn.
I am always happy to talk about reliable language models, small models, abstention, unusual model behaviours, or research ideas somewhere between “this probably should not work” and “why does this actually work?”
Contact
If you would like to get in touch, feel free to email me at
mchopra[at]uni-bonn.de.