Veeraraju ElluruI am a Computer Science Senior at the Indian Institute of Technology, Jodhpur. I love AI research and Math. My main areas of interest are Trustworthy ML and Mechanistic Interpretability in multimodal systems. When it comes to research, my primary agenda is to achieve a trifecta of privacy, safety, and trustworthiness. I am extremely grateful to be jointly mentored by Dr. Mayank Vatsa, Dr. Richa Singh, and Dr. Shivang Agarwal at the Image Analytics and Biometrics (IAB) Lab. I also lead the technical research wing at Safety and Alignment Research (SAAR), India. Throughout my undergraduate studies, I've actively worked on side-quests where AI research can help significantly contribute. This has given me the privilege to explore projects at some really cool places:
I'm a strong believer of Bruce Lee's 10000-hour to mastery rule; and am forever grateful to all my mentors who have shaped my research. |
"Whereof one cannot speak, thereof one must be silent." ~ Ludwig Wittgenstein |
News
|
Publications |
Mechanistically Interpreting Compression in Vision-Language ModelsVeeraraju Elluru, Arth Singh, Roberto Aguero, Ajay Agarwal, Debojyoti Das, Hreetam Paul arXiv Compressed vision-language models (VLMs) are widely used to reduce memory and compute costs, making them a suitable choice for real-world deployment. However, compressing these models raises concerns about whether internal computations and safety behaviors are preserved. In this work, we use causal circuit analysis and crosscoder-based feature comparisons to examine how pruning and quantization fundamentally change the internals across representative VLMs. We observe that pruning generally keeps circuit structure intact but rotates and attenuates internal features, while quantization modifies the circuits at a higher level yet leaves the surviving features better aligned. Leveraging this insight, we also introduce VLMSafe-420, a novel benchmark that pairs harmful inputs with matched benign counterfactuals across various safety categories. Our findings show that pruning causes a sharp drop in genuine refusal behavior, suggesting that the choice of compression has safety implications. |
Lie Algebra-Based Semantic Flow for Incompleteness Detection in SummarizationManikandan Ravikiran, Veeraraju Elluru, Karrtik Iyer, Prasanna Pendse, Shayan Mohanty Springer / Blog Evaluating the incompleteness of machine-generated summaries remains a critical challenge in natural language generation. In domains such as healthcare, education, and policy, summaries that omit or distort key content can appear fluent yet undermine reliability and trust. We introduce a Lie Algebra-based Semantic Flow framework that treats summarization as a geometric transformation in embedding space, where incompleteness manifests as segments with low semantic flow magnitudes. Unlike prior metrics that collapse evaluation into a single similarity score, our approach provides interpretable, segment-level signals through a dynamic mean-scaled thresholding mechanism, requiring no supervision. We evaluate our method on UniSumEval and SIGHT benchmarks, showing average improvements of +10.6 macro-F1 over existing methods. Theoretical analysis establishes rotation invariance, bounded flow magnitudes, and a severity-coverage Pareto frontier that explains observed trade-offs. Qualitative case studies further demonstrate that our framework identifies variety of incompleteness often missed by existing methods. |
Bias-Aware Machine Unlearning: Towards Fairer Vision Models via Controllable ForgettingSai Siddhartha Chary Aylapuram, Veeraraju Elluru, Shivang Agarwal arXiv / slides Deep neural networks often rely on spurious correlations in training data, resulting in biased or unfair predictions, particularly in safety-critical applications. While conventional bias mitigation methods typically require retraining from scratch or redesigning the data pipeline, recent advances in machine unlearning offer a promising alternative for post-hoc model correction. In this work, we explore the trifecta of efficiency, fairness, and model utility post unlearning via Bias-Aware Machine Unlearning, a paradigm that selectively forgets biased samples or feature representations to address various forms of bias in vision models. |
|
|