Santhosh G S

M.Sc. AI & ML

TU Darmstadt, Germany

I am Santhosh, an M.Sc. student in Artificial Intelligence and Machine Learning at the Technical University of Darmstadt. I am interested in reinforcement learning and representation learning, with a focus on developing reliable and adaptive autonomous systems that operate effectively in real-world environments.

Previously, I was a Post-Baccalaureate Fellow at the Robert Bosch Centre for Data Science and Artificial Intelligence (RBCDSAI), IIT Madras, advised by Prof. Balaraman Ravindran and Prof. Saurav Prakash. There, I worked on evaluation and explanations for RL-trained policies and agents, efficient inference techniques for language models, and built a framework and dataset for evaluating LLMs in the Indian context.

I completed my B.S. in Data Science and Applications from IIT Madras (off-campus degree) and my B.E. in Electrical and Electronics Engineering from Sri Sivasubramaniya Nadar College of Engineering (SSNCE), Chennai, India.

News

Jan 24, 2026	IndiCASA received the 3MT Runner-Up Award and Best Poster Award at the Emerging Global South AI (EGSAI) Workshop, AAAI 2026. 🏆
Nov 01, 2025	Received an all-expenses-paid invitation to attend IndoML 2025 from the IndoML Datathon committee, for submitting a quality entry to the IndoML Datathon 2025.
Oct 13, 2025	Our research work IndiCASA was featured as an Eight Column Piece in a National Daily and covered by various national newspapers as web articles! 📰
Jul 19, 2025	🎉 IndiCASA accepted at AAAI/ACM Conference on AI, Ethics, and Society (AIES) 2025! Our work on bias evaluation in LLMs using contrastive embedding similarity has been recognized by the top-tier conference.
Dec 23, 2024	Presented our IndoML Datathon 2024 solution at the main panel of IndoML 2024 Symposium! 🎤 Check out our presentation.

Selected Publications

AIES
IndiCASA: A Dataset and Bias Evaluation Framework in LLMs Using Contrastive Embedding Similarity in the Indian Context

Santhosh G S, Akshay S Govind, Gokul S Krishnan, and 2 more authors

In AAAI/ACM Conference on AI, Ethics, and Society (AIES), 2025

Abs Bib HTML PDF Code Poster

Large Language Models (LLMs) have gained significant traction across critical domains owing to their impressive contextual understanding and generative capabilities. However, their increasing deployment in high stakes applications necessitates rigorous evaluation of embedded biases, particularly in culturally diverse contexts like India where existing embedding-based bias assessment methods often fall short in capturing nuanced stereotypes. We propose an evaluation framework based on an encoder trained using contrastive learning that captures fine-grained bias through embedding similarity. We also introduce a novel dataset - IndiCASA (IndiBias-based Contextually Aligned Stereotypes and Anti-stereotypes) comprising 2,575 human-validated sentences spanning five demographic axes: caste, gender, religion, disability, and socioeconomic status.
@inproceedings{santhosh2025indicasa, title = {IndiCASA: A Dataset and Bias Evaluation Framework in LLMs Using Contrastive Embedding Similarity in the Indian Context}, author = {G S, Santhosh and Govind, Akshay S and Krishnan, Gokul S and Ravindran, Balaraman and Natarajan, Sriraam}, booktitle = {AAAI/ACM Conference on AI, Ethics, and Society (AIES)}, year = {2025}, organization = {AAAI/ACM}, proceedings = {https://ojs.aaai.org/index.php/AIES/article/view/36605}, }
arXiv
AQUA: Attention via QUery mAgnitudes for Memory and Compute Efficient Inference in LLMs

Santhosh G S, Saurav Prakash, and Balaraman Ravindran

arXiv preprint arXiv:2509.11155, 2025

Abs Bib PDF

The quadratic complexity of the attention mechanism remains a fundamental barrier to scaling Large Language Models (LLMs) to longer contexts, creating a critical bottleneck in both computation and memory. To address this, we introduce AQUA (Attention via QUery mAgnitudes) a novel and versatile approximation strategy that significantly reduces the cost of attention with a graceful performance trade-off. Our method operates in two phases: an efficient offline step where we compute a universal, language agnostic projection matrix via SVD on a calibration dataset, and an online inference step where we project query and key vectors and dynamically select a sparse subset of dimensions based on the query’s magnitude. We demonstrate that a 25% reduction in the attention dot-product computation can be achieved with statistically insignificant impact on performance across a wide range of benchmarks.
@article{santhosh2025aqua, title = {AQUA: Attention via QUery mAgnitudes for Memory and Compute Efficient Inference in LLMs}, author = {G S, Santhosh and Prakash, Saurav and Ravindran, Balaraman}, journal = {arXiv preprint arXiv:2509.11155}, year = {2025}, }
arXiv
SWAN: Sparse Winnowed Attention for Reduced Inference Memory via Decompression-Free KV-Cache Compression

Santhosh G S, Saurav Prakash, and Balaraman Ravindran

arXiv preprint arXiv:2511.18936, 2025

Abs Bib PDF

Large Language Models (LLMs) face a significant bottleneck during autoregressive inference due to the massive memory footprint of the Key-Value (KV) cache. Existing compression techniques like token eviction, quantization, or other low-rank methods often risk information loss, have fixed limits, or introduce significant computational overhead from explicit decompression steps. In this work, we introduce SWAN, a novel, fine-tuning-free framework that eliminates this overhead. Our method uses an offline orthogonal matrix to rotate and prune the KV-cache, which is then used directly in the attention computation without any reconstruction. Our extensive experiments demonstrate that SWAN, augmented with a small dense buffer, offers a robust trade-off, maintaining performance close to the uncompressed baseline even at aggressive 50-60% memory savings per-token on KV-cache.
@article{santhosh2025swan, title = {SWAN: Sparse Winnowed Attention for Reduced Inference Memory via Decompression-Free KV-Cache Compression}, author = {G S, Santhosh and Prakash, Saurav and Ravindran, Balaraman}, journal = {arXiv preprint arXiv:2511.18936}, year = {2025}, }