Gaurav Srivastava is building an open-source framework, effGen, to design powerful AI agents with small language models that can autonomously collaborate to solve complex real-world problems without relying on expensive large models. 

These agents can be run/deployed locally and can be very business friendly to enterprise for automating workflows. They can also handle repetitive tasks, delivering practical impact at scale with minimal computational overhead, said Srivastava, a master’s degree student in computer science advised by Xuan Wang.

“Because they can be deployed and run locally, there are no privacy or safety concerns for sensitive applications, such as in hospitals,” he said, “and from our framework we demonstrate that they can be as powerful as large model agents if used in the right way. We envision making it super useful for an entire AI community and industry as well as for academia, so that anyone can build agents for free and or much less cost.”

“EFFGEN: Enabling Small Language Models as Capable Autonomous Agents,” a collaboration with Google DeepMind and Georgia Tech, is his most recent paper on the framework.

Srivastava, who holds a bachelor's degree in technology in computer science from Manipal University, Jaipur, India, said he has always been focused on efficiency. 

“In both my research projects and industry experience I have focused on how to be more efficient and lower the business cost of anything I build,” he said. “When I first met my advisor, Dr. Wang, she was working in large language models multi-agent systems. We envisioned how to make these systems more efficient and that turned out to be small language models, now my core research focus and interest.”

Since coming to Virginia Tech, Srivastava’s published papers include:

·      "ThinkSLM: Towards Reasoning in Small Language Models," in main proceedings at 2025 Empirical Methods in Natural Language Processing (EMLP) 

·      "DEBATE, TRAIN, EVOLVE: Self‐Evolution of Language Model Reasoning," in main proceedings at 2025 EMLP 

·      "JudgeBoard: Benchmarking and Enhancing Small Language Models for Reasoning Evaluation," oral presentation at 2026 Association for Advancement of Artificial Intelligence (AAAI) conference

Three papers have been accepted to conferences which will be held later this spring and summer:

·      "BeyondBench: Benchmark-Free Evaluation of Reasoning in Language Models,", main proceedings at 2026 International Conference on Learning Representations (ICLR) 

·      "Do LLMs Overthink Basic Math Reasoning? Benchmarking the Accuracy-Efficiency Tradeoff in Language Models," in Findings at Annual Meeting of the Association for Computational Linguistics (ACL) 2026

·      "SoundBreak: A Systematic Study of Audio-Only Adversarial Attacks on Trimodal Models," main proceedings at ACL 2026 

As the Sanghani Center, Srivastava said, students “get amazing support to participate at conferences. And the faculty there really motivate students, pushing the boundaries of science and fostering work that aligns with current industry focus.” 

He has a long range plan post-graduation next month -- first working at industry research to make connections and build a network, and eventually, starting his own company. 

“I am working toward this goal through effGen for which I have already received funding support and am hoping it will become that company,” he said.