Tushar Semwal

Teaching Small Models When Not to Call Functions: Structured Reasoning for Tool Refusal in Low-Resource Languages

Standard fine-tuning teaches models how to call tools — not when not to. In low-resource languages (Vietnamese, Thai) on 1B-parameter models, this causes catastrophic hallucination (−21pp on Gemma-Thai). We fix it with structured key-value reasoning forms that make the refusal decision explicit at training time, improving irrelevance detection and hallucination by +10–30pp and slashing hallucination probability 50–60× (from 67–76% down to 1.2–1.3%) — at 1.9× lower token cost than chain-of-thought. Published at SIGIR '26, Melbourne.

Dung Pham Tuan Vo*, Thai Trung Tran*, Tushar Semwal (* equal contribution)

[Paper] [ACM DL]

FedPerf: A Practitioners' Guide to Performance of Federated Learning Algorithms

Which federated learning algorithm should you actually deploy? FedAvg, FedProx, FedMed, and qFedAvg all look similar on accuracy — but diverge dramatically when your data is skewed, nodes drop out, or fairness across clients matters. This large-scale empirical study benchmarks them across all three dimensions simultaneously using spider charts, giving practitioners a single diagnostic view to pick the right algorithm for their constraints. Published in PMLR 148, 2021.

Ajinkya Mulay, Baye Gaspard, Rakshit Naidu, Santiago Gonzalez-Toral, Vineeth S, Tushar Semwal, Ayush Manish Agrawal

[Paper] [Google Scholar]

Selective Federated Transfer Learning using Representation Similarity

In federated learning you can't see client data — so how do you pick the right pretrained model to transfer? We use Centered Kernel Alignment (CKA) with sketching to compare model representations locally on-device, then run a federated voting algorithm to select the best source model without sharing any raw data. Cuts communication rounds up to 5× versus FedAvg baseline.

Tushar Semwal, Haofan Wang, Chinnakotla Krishna Teja Reddy

[Paper] [Google Scholar]

A Practitioners' Guide to Transfer Learning for Text Classification using Convolutional Neural Networks

How much of a CNN trained on one text domain transfers to another — and which layers should you freeze? This practitioner's guide runs systematic experiments across layer depth, fine-tuning strategies, and hyperparameters, delivering concrete rules of thumb for real NLP pipelines. Originated from work at Samsung Research; 84 citations and still a go-to reference for applied transfer learning.

Tushar Semwal, Promod Yenigalla, Gaurav Mathur, Sumit Bose Nair

[Paper] [Google Scholar]

On Ordering Multi-Robot Task Executions within a Cyber Physical System

When multiple robots share a workspace, wrong task ordering causes bottlenecks, collisions, and deadlocks. This paper formalises the scheduling problem in cyber-physical systems and proposes a distributed solution that balances workload without a central coordinator — validated on real robot hardware. Published in ACM TAAS, one of the top journals in autonomous adaptive systems.

Tushar Semwal, Shashi Shekhar Jha, Sumit Bose Nair

[ACM DL] [Google Scholar]