Sujoy Nath

AI Researcher

Sony Research India

I am working in causal inference for user engagement modeling and building agentic systems.

View Research Download CV Contact
Sujoy Nath

About Me

I am currently working in Sony Research India in the User Engagement Research team, primarily modeling user behavior of the SonyLiv platform using causal inference and building agentic simulation environments.

Previously, I worked in the LCS2 lab at the Indian Institute of Technology (IIT), Delhi, in collaboration with Microsoft Research India (MSRI), co-supervised by Prof. Tanmoy Chakraborty and Principal Engineer at MSRI Dr. Akshay Nambi on the evaluation, failure analysis, and interpretable orchestration policies in LLM multi-agent systems.

I also have had the opportunity to work extensively with Prof. Swagatam Das at the Indian Statistical Institute (ISI), Kolkata, where we developed methods for detection and mitigation of hallucination and safety alignment in LLMs resulting in publications at EACL (Oral), IJCNN, AAAI.

My research interests span across trustworthy AI systems, LLM interpretability and safety, agentic reasoning, and AI applications in healthcare. I am particularly drawn to developing AI systems that are not only powerful but also reliable, transparent, and beneficial for society.

Research Interests

LLM Interpretability & Safety

Investigating methods to understand and improve the safety mechanisms of large language models.

Agentic Reasoning

Exploring autonomous decision-making capabilities in AI systems and multi-agent interactions.

AI for Medical Domain

Applying AI technologies to healthcare challenges with emphasis on reliability and explainability.

Publications

Disentangling Intrinsic Importance from Emergent Structure in Multi-Expert Orchestration

Sudipto Ghosh*, Sujoy Nath*, Sunny Manchanda, Tanmoy Chakraborty
Transactions on Machine Learning Research (TMLR) 2026

(* means equal contribution)

Paper

ARREST: Adversarial Resilient Regulation Enhancing Safety and Truth in Large Language Models

Sharanya Dasgupta, Arkaprabha Basu, Sujoy Nath, Swagatam Das
EACL 2026 (Oral) (Rank: A)
Paper arXiv

From Complexity to Clarity: Transforming Chest X-ray Reports with Chained Prompting

Sujoy Nath, Arkaprabha Basu, Kushal Bose, Swagatam Das
AAAI 2025 Student Abstract (Rank: A*)
Paper

HalluShift: Measuring Distribution Shifts towards Hallucination Detection in LLMs

Sharanya Dasgupta, Sujoy Nath, Arkaprabha Basu, Pourya Shamsolmoali, Swagatam Das
IJCNN 2025 (Rank: B)
Paper arXiv

HalluShift++: Bridging Language and Vision through Internal Representation Shifts for Hierarchical Hallucinations in MLLMs

Sujoy Nath, Arkaprabha Basu, Sharanya Dasgupta, Swagatam Das
ICVGIP 2025 (Oral)
arXiv

Experience

Researcher

June 2026 - Present
Sony Research India
  • Working in the user engagement research team.

Research Associate

July 2025 - May 2026
Indian Institute of Technology (IIT), Delhi
  • Lab: Laboratory of Computational Social Systems (LCS2)
  • Worked on an Agentic AI project in collaboration with Microsoft Research India (MSRI), co-supervised by Prof. Tanmoy Chakraborty and principal engineer at MSRI Dr. Akshay Nambi.
  • Agentic Evaluation: Proposed a failure-centric evaluation framework for LLM agentic systems, introducing a unified failure taxonomy to identify execution-level breakdowns across an agent's reasoning, planning, and multi-step decision pipelines. Analyzed multi-agent orchestration frameworks (Autogen, CrewAI, LangChain) to diagnose planning stability, tool calling, and reasoning-action alignment, optimizing reliability for agentic workflows beyond simple task success rates.
  • Agentic Orchestration Interpretability: Collaborated with DRDO to develop INFORM, an interpretability framework for multi-expert LLM orchestration. Conducted analyses on GSM8K, HumanEval, and MMLU using LLaMA-3.1 8B, Qwen3 8B, and DeepSeek-R1 8B expert ensembles. Combined gradient sensitivity analysis with interaction topology to distinguish intrinsic expert importance from routing behavior, demonstrating that frequently selected experts are not necessarily the most functionally critical. Accepted at TMLR 2026.
  • FIRE Framework: Proposed FIRE, a multi-agent framework for category-aware, fact-grounded counterspeech generation, alongside FactualCS, a 4,784 instance annotated dataset with reasoning traces and evidence mappings. Demonstrated SOTA performance over 28 baselines, achieving ~ 12% higher factual accuracy, ~ 11% better category-specific accuracy, and ~ 11% lower toxicity while leveraging compact (<2B) language models.

Research Collaborator

June 2024 - June 2025
Indian Statistical Institute (ISI), ECSU
Supervisor: prof. Swagatam Das
  • ARREST: Contributed to ARREST, a unified framework to mitigate factual and safety failures in LLMs by addressing representational misalignments in the latent activation space. Engineered an external intervention network to self-correct drifted features, regulating falsehoods and unsafe outputs without fine-tuning model parameters, demonstrating superior versatility over standard RLHF models in handling soft refusals. Accepted at EACL 2026 (Oral).
  • HalluShift: Developed novel approach for detecting factual hallucinations in LLM outputs by analyzing distributional shifts in internal state space and token probabilities. Accepted at IJCNN 2025
  • Medical Report Generation: Created simplified MRG system using Gemini-1.5-Flash, fine-tuned LLaMa models, and introduced CPMK-E scoring method. Accepted as Student Abstract at AAAI 2025
  • Image Captioning Pipeline: Built system using BLIP embeddings and CerberusDet (YOLOv8) for object detection, fine-tuned multiple LLMs (LLaMa 3.1/2, Mistral 7B, Phi-2) with comprehensive evaluation metrics

Machine Learning Engineer Intern

October 2023 - May 2025
Geogo Techsolutions
  • Face Recognition System: Built an end-to-end face comparison and similarity search solution, achieving 98% accuracy in benchmarks against Amazon Rekognition. Implemented state-of-the-art computer vision models for employee authorization and face verification, successfully matching current faces with historical photos. It is deployed with kriyam.ai.
  • Speech Recognition: Engineered an Automatic Speech Recognition (ASR) model using Transformer architecture, trained on a 3-hour Hindi-English mixed dataset. Achieved 34% WER, now deployed for insurance call transcription.
  • Document Assistant: Co-led the development of Kriyam DocWise (docwise.kriyam.ai), an AI-powered document assistant that enables context-aware, natural-language access to policy, claim and ID documents. I architected the hybrid RAG pipeline, combining OCR, embedding-based retrieval and summarization to eliminate manual review and accelerate policy analysis for claims investigation teams.

Summer Intern

March 2024 - April 2024
Defence Research and Development Organisation (DRDO), DEBEL Lab
  • Developed comprehensive dataset using IMU sensor data from Xsens, capturing walking patterns of 20 subjects for motion analysis
  • Researched human gait pattern analysis using ensemble learning methods for real-time prediction systems in lower limb prosthetics

Education

Bachelor of Technology

Computer Science and Business System

Maulana Abul Kalam Azad University of Technology (MAKAUT), Netaji Subhash Engineering College

August 2021 - June 2025

CGPA: 8.66/10 (3.46/4)

Achievements

Parayas 2k24: National Inter-college Project Competition

2nd Position - Software Segment

Team Lead | May 2024

Kavach 2023: National Cybersecurity Hackathon

Top 5 Grand Finalist

Team Lead | Aug 2023

Get In Touch

I am always open to discussing research opportunities, collaborations, and innovative ideas. Feel free to reach out for discussions or potential research partnerships.

sujoynathofficial@gmail.com
New Delhi, India | Kolkata, West Bengal, India

Citation


                

                
                
Copied!