Ativ Joshi

Hello, I’m Ativ (अतीव), a PhD student in the Manning CICS at the University of Massachusetts Amherst, where I am advised by Dr. Mohammad Hajiesmaili at the SOLAR Lab. My research sits at the intersection of machine learning theory, systems, and large language models.

Broadly, I am interested in designing learning and optimization algorithms for sequential decision-making under uncertainty, especially in settings involving resource allocation, scheduling, retrieval, and efficient ML systems. Much of my work studies how ideas from online learning, bandits, caching, and optimization can be used to build algorithms with strong theoretical guarantees and practical relevance.

My recent research includes online caching in networks, fair and reusable resource allocation and safe RLHF. I am particularly interested in using tools from online optimization, scheduling, and learning to make LLM serving and inference-time systems more efficient and robust.

More generally, I am excited by problems where theory and systems genuinely inform each other: where formal models help us understand modern AI systems, and where practical constraints inspire new algorithmic questions. Before my current work, I also spent time on computer algebra, probabilistic graphical models, and related theoretical areas.

Prior to this, I was a Research Assistant at the Learning and Networks Group, TIFR led by Dr. Abhishek Sinha. I obtained my masters’ in Computer Science from Chennai Mathematical Institute (CMI) and my BTech in Information and Communication Technology from Ahmedabad University.

news

Mar 04, 2026	Our paper on Online Caching in Tree Networks: Algorithms, Regret, and Complexity has been accepted to L4DC.
Sep 25, 2023	Our paper on No-regret Algorithms for Fair Resource Allocation has been accepted to NeurIPS 2023.
Sep 05, 2023	Started PhD at Manning CICS, UMass Amherst.
Aug 10, 2022	Our paper on Universal Caching has been accepted to ITW 2022.
Jul 10, 2021	Joined as a Research Assistant at the Learning and Networks Group, TIFR.

selected publications

Working paper

Safe RLHF Beyond Expectation: Stochastic Dominance and Optimal Transport for Universal Spectral Risk Control

Y. Chittepu, A. Joshi, R. Bhattacharjee, and 1 more author

2026

Working paper
L4DC

Online Caching in Tree Networks: Algorithms, Regret, and Complexity

A. Joshi, R. Bhattacharjee, R. De, and 3 more authors

2026

Accepted at Learning for Dynamics and Control
Under review

Learning to Allocate Reusable Resources under Stochastic Rewards and Durations

A. Joshi, Q. Liu, and M. Hajiesmaili

2026

Under review at SIGMETRICS
NeurIPS

No-regret Algorithms for Fair Resource Allocation

Abhishek Sinha, Ativ Joshi, Rajarshi Bhattacharjee, and 2 more authors

NeurIPS, Sep 2023

Abs arXiv

We consider a fair resource allocation game in the no-regret setting against an unrestricted adversary. The objective of the problem is to allocate resources equitably among several agents in an online fashion so that the difference of the aggregate α-fair utilities of the agents between an optimal static clairvoyant allocation and that of the online policy grows sub-linearly with time. The problem is challenging due to the non-additive nature of the α-fairness function. Previously, it was shown that no online policy can exist for this problem with a sublinear standard regret. In this paper, we propose an efficient online resource allocation policy, called Online Proportional Fair (\textttOPF), that achieves c_α-approximate sublinear regret with the approximation factor c_α=(1-α)^-(1-α)≤1.445, for 0≤α≤1. The upper bound to the c_α-regret for this problem exhibits a surprising \emphphase transition phenomenon. The regret bound changes from the power-law to a constant at the critical exponent α=\nicefrac12. As a corollary, our result also resolves an open problem raised by \citeteven2009online on designing an efficient no-regret policy for the online job scheduling problem in certain parameter regimes. The proof of our results introduces new algorithmic and analytical techniques, including greedy estimation of the future gradients for non-additive global reward functions and bootstrapping adaptive regret bounds, which may be of independent interest.
ITW

Universal Caching

A. Joshi, and A. Sinha

Information Theory Workshop, Nov 2022

Abs DOI arXiv Slides

In the learning literature, the performance of an online policy is commonly measured in terms of the static regret metric, which compares the cumulative loss of an online policy to that of an optimal benchmark in hindsight. In the definition of static regret, the benchmark policy remains fixed throughout the time horizon. Naturally, the resulting regret bounds become loose in non-stationary settings where fixed benchmarks often suffer from poor performance. In this paper, we investigate a stronger notion of regret minimization in the context of an online caching problem. In particular, we allow the action of the offline benchmark at any round to be decided by a finite state predictor containing arbitrarily many states. Using ideas from the universal prediction literature in information theory, we propose an efficient online caching policy with an adaptive sub-linear regret bound. To the best of our knowledge, this is the first data-dependent regret bound known for the universal caching problem. We establish this result by combining a recently-proposed online caching policy with an incremental parsing algorithm, e.g., Lempel-Ziv ’78. Our methods also yield a simpler learning-theoretic proof of the improved regret bound as opposed to the more involved and problem-specific combinatorial arguments used in the earlier works.