Partha Pratim Saha

Partha Pratim Saha

AI Researcher  ·  Lecturer in Computer Science  ·  BITS Pilani M.Tech (GPA 9.08/10, Top 5%)

technical.partha@gmail.com LinkedIn GitHub Google Scholar
Download CV (PDF)
Prospective AI PhD applicant focusing on Natural Language Processing (NLP), AI Safety, and Large Language Model (LLM) Alignment.

Research Interests

AI Safety & Alignment in Large Language Models  ·  Mechanistic Interpretability  ·  Trustworthy & Agentic AI  ·  Natural Language Processing (NLP)  ·  Explainable AI (XAI)  ·  AI Deception Detection

Core Research: Belief Vector Field & Geometric Alignment

My primary research develops information-geometric frameworks to measure how LLMs encode, propagate, and transform beliefs across transformer layers — and how alignment (DPO/RLHF) and fine-tuning alter that internal geometry. Core finding: torsional geometry provides the most statistically significant separation (highest Cohen's d effect size) between safe and harmful prompts across SFT & DPO model variants.

Models: LLaMA-3.2, OLMo-7B, Gemma-2, Mistral-7B, Zephyr-7B, Qwen  |  Datasets: LITMUS, HarmfulQA, hh-rlhf  |  Metrics: 17 geometric metrics (torsion, spectral curvature, thermodynamic length, DTW)

Selected Publications

MENTIS Visual Abstract
MENTIS: What Belief Changes Under Alignment? Measuring Multi-Scale Latent Torsion Across Four Language Models
Partha Pratim Saha, Samarth Raina, Mayur Parvatikar, Amit Dhanda, Vinija Jain, Aman Chadha, Amitava Das
AAAI 2027  ·  AI Alignment Track (Under Review)
paper | code
Introduces 8 multi-scale latent torsion metrics to quantify depth-wise belief transformations. Proves Direct Preference Optimization (DPO) induces 44.4% suppression in high-order torsion (Cohen's d=0.741, p=7.7×10⁻¹³) across 4 LLM families, establishing an entropy–torsion anti-correlation bridge (ρ = −0.387).
Keywords: AI Alignment, Mechanistic Interpretability, Latent Torsion, Information Geometry, LLM Belief Dynamics
GRAFT Visual Abstract
GRAFT: Geometric Representations of Alignment's Fingerprint in Transformer Belief Trajectories
Partha Pratim Saha
ICML 2026  ·  Workshop on Mechanistic Interpretability (Poster)
paper | workshop posters | code
Proposes a differential-geometric framework mapping belief vectors across transformer layers. Demonstrates that second-order torsion T2 is 8× more concept-discriminative than Centered Kernel Alignment (CKA) (AUC 0.89) in detecting safety-critical representations.
Keywords: Mechanistic Interpretability, Belief Trajectories, Geometric Fingerprinting, Transformer Geometry
SPINAL Visual Abstract
SPINAL: Scaling-law and Preference Integration in Neural Alignment Layers
Arion Das, Partha Pratim Saha, Aman Chadha, Vinija Jain, Amitava Das
AAAI 2027  ·  AI Alignment Track (Under Review)
paper | code
Investigates empirical scaling laws and preference integration dynamics in transformer alignment layers, analyzing layer-wise thermodynamic length and curvature deformations under direct preference optimization (DPO) and RLHF.
Keywords: Scaling Laws, Preference Optimization, Alignment Layers, Thermodynamic Length, Representation Geometry
Robust Neural Classifiers Visual Abstract
No Unique Minimizer, No Problem: On the Consistency of Robust Neural Classifiers
Subhabrata Majumdar, Anand Deo*, Partha Pratim Saha*, Abhik Ghosh  (*Equal contribution)
AAAI 2027  ·  Under Review
paper
Establishes consistency theory for robust neural classifiers trained via the S-divergence family under label noise and adversarial contamination without requiring parameter-identifiability assumptions. Proves empirical minimizers converge to the population-optimal equivalence class.
Keywords: Robust Neural Classifiers, S-Divergence, Consistency Theory, Adversarial Robustness, Non-identifiability
Prompting Away Stereotypes Visual Abstract
Prompting Away Stereotypes? Evaluating Bias in Text-to-Image Models for Occupations
Shaina Raza, Maximus Powers*, Partha Pratim Saha*, Mahveen Raza, Rizwan Qureshi  (*Equal contribution)
NeurIPS 2025  ·  Generative and Protective AI for Content Creation Workshop
paper | code
Presents an extensive empirical audit evaluating gender and racial bias across occupational queries in leading text-to-image foundation models (DALL·E 3, Midjourney v5, Stable Diffusion), studying prompt-level interventions and protective steering strategies.
Keywords: Text-to-Image Models, Algorithmic Fairness, Bias Auditing, Generative AI Safety, Sociotechnical AI
Collaborative Federated Learning Visual Abstract
Collaborative Federated Learning Cloud-Based System
Partha Pratim Saha, Naresh K. Sehgal, Miad Faezipour
ICOMP'24 / Springer Nature  ·  Computer Engineering & Applied Computing (CSCE), USA (Accepted)
paper
Architects a cloud-based collaborative federated learning system that partitions data into public and private sets via automated PHI and PII de-identification, enabling privacy-compliant distributed learning for personalized GPT-scale language systems.
Keywords: Federated Learning, Privacy Preservation, Cloud Computing, Healthcare NLP, PHI/PII Protection
Human Empathy in Conversations Visual Abstract
Enhancing Human Empathy in Conversations Using Transformer-Based Models
Cherishma Kumar Subhasa, Endriyas Zenagebriel, Partha Pratim Saha*, Zarah Rezaei, Joseph Akinyemi  (*Equal contribution)
Sciencematch Journal 2025  ·  Impact Scholar Program 2025
paper
Develops transformer-based architectures for fine-grained emotion taxonomy recognition and empathetic response generation in multi-turn dialogues, modeling communicative intent and affective dynamics in interactive agents.
Keywords: Conversational AI, Affective Computing, Empathy Recognition, Dialogue Systems, Transformer NLP

Awards & Achievements

Research & Work Experience

Lecturer in CS & Head of Department Dec 2021 – Present
Nalhati Government Polytechnic College (Govt. of West Bengal), India
Teaching ML, Deep Learning, IoT, Python, and Java. Project supervisor for 50+ final-year students on AI, NLP, Agentic AI, and Empathetic Chatbot projects. Administrative HoD responsibilities. Active research on AI Safety & Alignment (GRAFT/MENTIS), Semantic Helix of LLMs (nDNA), Cultural LLMs, and Neural Robustness.
Lead Data Scientist — Conversational Dialog System May 2021 – Nov 2021
Wipro Limited & IBM Research (Joint Project), Bangalore, India
Led a team of 5 to build a conversational chatbot removing query ambiguities. Implemented 50+ custom intents, entities, and dialog flows with IBM Watson. Impact: 0.3 million users worldwide. Tools: DialogFlow, IBM Cloud, NodeJS.
Senior Data Scientist — Medical Search Engine (J&J R&D) Dec 2017 – Aug 2019
BirlaSoft · Johnson & Johnson R&D, New Delhi, India
Built medical search engine using SciBERT and SciSpacy NLP pipeline for healthcare queries. Impact: 0.1 million+ J&J product users. Tools: Python, Word2Vec, SciSpacy, Fuzzy Search, Flask.
Project Engineer — Threat Intelligence System Nov 2016 – Jul 2017
Indian Institute of Technology (IIT) Kanpur, Computer Science Dept., India
Developed a secure threat management system for academic institutions. Researched cyber-security defences against integrity, confidentiality, and non-repudiation attacks. Tools: Python, Drupal, Django.
Senior Systems Engineer — Alignment & Cancer Genomics in AI Jan 2011 – Jul 2015
Infosys Technologies Limited, Chennai, India
Applied Edit Distance and Needleman-Wunsch algorithms on DNA sequences (FASTQ) to identify gene insertions/deletions. Identified top 10 genes driving Multiple Myeloma blood cancer; implemented 3 research papers. Impact: biological hierarchy determination, drug design, life expectancy improvement.

Teaching Assistantships

Teaching Assistant (M.Tech Programme) 2021 – 2023
BITS Pilani, India
TA for three graduate courses: NLP Applications [Winter 2023], Deep Learning [Fall 2021], Deep Reinforcement Learning [Spring 2021]. Responsibilities: webinar demonstrations, labs, quizzes, exam preparation, student mentoring. Honorarium: USD $2,513.11 across all three courses.

Education

M.Tech in Data Science & Engineering
Birla Institute of Technology and Science (BITS Pilani), Pilani, India — 2019–2021
GPA: 9.08/10  ·  Distinction  ·  Top 5% out of 600
Courses: NLP, Machine Learning, Deep Learning, Data Science, Mathematics, Statistics, Data Mining, Big Data Systems
Dissertation: Collaborative Federated Learning (CFL) cloud system separating PHI/PII for privacy-preserving GPT-like systems (published SpringerNature)
B.Tech in Computer Science & Engineering
West Bengal University of Technology (WBUT), Kolkata, India — 2006–2010
GPA: 8.49/10  ·  Top 5% out of 70
Courses: Mathematics, Statistics, Algorithms, AI, Probability, Data Structures, Programming

Technical Skills

AI Safety: Belief Change under AI Alignment, Mechanistic Interpretability, Model & Activation Steering, Recursive Self-Improvement Risks, LLM Deception, Manipulation & Sycophancy, Bias Mitigation & AI Misalignment, Catastrophic Risk Reduction in Frontier Models, Safety-Relevant Behavioral Evaluation, Belief & Knowledge Editing, Explainable & Trustworthy AI (XAI)
NLP: Foundation Models, LLM Alignment, AI Agents, Conversational AI Systems, Generative AI, Fine-tuning via HuggingFace, Hybrid & Multi-hop Retrieval-Augmented Generation (RAG)
LLMs (hands-on): DeepSeek, GPT, LLaMA, Gemma, Mistral, Qwen, OLMo — base & instruct variants; model fine-tuning, reasoning, NeuroSymbolic AI, ethics & trustworthiness, evaluation & benchmarking
Deep Learning: Transformer, GPT, Autoregressive Generative Models, CNN, RNN, Neural Networks
ML & Data Science: Text & Multimodal Data curation, Regression, Classification, Clustering, Tree-Based Algorithms, Bagging, Boosting
Programming: Python, Java, C, Objective-C
Frameworks & Tools: PyTorch, LlamaIndex, LangChain, Scikit-learn, NumPy, Pandas, NLTK, SpaCy, TensorFlow, Keras, Flask, NodeJS, Docker, Kubernetes, Git, PostgreSQL, MySQL
Cloud: AWS (EC2), Azure, IBM Cloud, Watson ML; inference deployment systems
Domains: Cancer Genomics, Bioinformatics, Healthcare AI, Real Estate, Finance & Banking

Contact

I am actively seeking fully-funded PhD positions in AI Safety, NLP, and LLM Alignment.
I welcome contact from faculty and researchers working on interpretable, responsible, and safe AI systems.

Email: technical.partha@gmail.com
LinkedIn: linkedin.com/in/partha121
GitHub: github.com/pps121
CV: Download PDF