Siva Rajesh Kasa
Applied machine learning researcher
Siva Rajesh Kasa
I study inference acceleration for language models, reliable text classification, machine learning privacy, and statistical learning. I am a Senior Applied Scientist at Amazon in Seattle.
Research
Language model inference
Faster generation at matched performance
Methods for accelerating autoregressive and diffusion language models, including relaxed speculative decoding and Fréchet profile decoding.
Classification and privacy
How modeling choices change behavior
Work on generative and discriminative classification, shortcut learning, concept shift, calibration, and membership inference attacks.
Search and recommendation
Learning under sparse feedback
Production research on search relevance, multimodal calibration, long tail ranking, counterfactual analysis, and experimentation.
Statistical learning
Dependence beyond correlation
Mixture models, copulas, high dimensional clustering, time series, and sequential decision making.
From the notebook
For a differentiable transformation g_θ of noise ε whose distribution does not depend on θ, the pathwise gradient is
\[ \nabla_\theta \mathbb{E}_{\varepsilon \sim p(\varepsilon)} \left[f\!\left(g_\theta(\varepsilon)\right)\right] = \mathbb{E}_{\varepsilon \sim p(\varepsilon)} \left[\nabla_\theta f\!\left(g_\theta(\varepsilon)\right)\right]. \]
Selected publications
NeurIPS 2026
Fréchet Profile Decoding for Faster Diffusion Language Model Inference
Siva Rajesh Kasa*, Yasong Dai*, Sumit Negi, Hongdong Li
EMNLP 2025, Outstanding Paper
Generative or Discriminative? Revisiting Text Classification in the Era of Transformers
Siva Rajesh Kasa*, Karan Gupta, Sumegh Roychowdhury, Ashutosh Kumar, Yaswanth Biruduraju, Santhosh Kumar Kasa, Nikhil Priyatam Pattisapu, Arindam Bhattacharya, Shailendra Agarwal, Vijay Huddar
AISTATS 2026
Relaxed Speculative Decoding via Dynamic Ensemble Verification
Ziyi Wang*, Siva Rajesh Kasa*, Ankith M S, Santhosh Kumar Kasa, Jiaru Zou, Sumit Negi, Ruqi Zhang, Nan Jiang, Qifan Song
Preprint
CopRoute: Dependence Aware Conformal Routing for Language Models
Siva Rajesh Kasa*, Santhosh Kumar Kasa*, Karan Gupta
Writing
I use the writing section for self contained notes on probability, statistics, optimization, and machine learning. Every page is native to this site and renders TeX equations directly.
Reparameterization gradients
Two unbiased gradient estimators for expectations, with an exact variance comparison in a Gaussian example.
Why optimize log likelihood?
The optimization, algebraic, and numerical reasons for replacing a likelihood product with a sum of log densities.