Profile picture
Bhiman Kumar Baghel
I'm a 3rd Year Ph.D. student in Computer Science at the University of Pittsburgh advised by Prof. Xiang (Lorraine) Li.

My research focuses on interpretable, parameter-efficient reasoning in large language models.

I developed a plug-and-play editing framework that improves editing performance by 38 points over prior SOTA. This was enabled by formalizing key failure modes in model editing—UnderEdit (failure to inject knowledge) and OverEdit (unintended side effects)—and identifying their root causes through empirical and representational analysis.

More recently, I've been exploring how parameter-efficient tuning methods, like LoRA, affect reasoning in legal domains, using data Shapley values and training dynamics to understand and improve generalization. My broader goal is to steer foundation models toward more reliable, trustworthy reasoning in high-stakes applications.
Publications and Patents
Research:
Bhiman Kumar Baghel, Scott M. Jordan, Zheyuan Ryan Shi, Xiang Lorraine
arXiv, 2025 Preprint

We propose techniques to systematically resolve UnderEdit and OverEdit issues in model editing, improving both precision and generalization.

Bhiman Kumar Baghel, Arun Balajiee Lekshmi Narayanan, Michael Miller Yoder
GeBNLP Workshop, ACL 2024 Workshop

We investigate biases in human vs AI-generated student summaries, proposing fairness metrics and improving reflection generation systems.

Yang Zhong, Bhiman Kumar Baghel
MULA Workshop, CVPR 2024 Workshop

We propose methods for fair interpretation of memes by jointly modeling image and text, focusing on bias mitigation across sensitive attributes.

Niraj Kumar, Bhiman Kumar Baghel
IEEE Access, 2021 Journal

We propose intent-focused semantic parsing and zero-shot out-of-domain detection strategies to enhance the robustness of spoken language understanding systems.

Niraj Kumar, Bhiman Kumar Baghel
IEEE Access, 2021 Journal

We introduce a smart stacking approach for intent-slot extraction in multi-intent spoken language understanding tasks, improving extraction granularity.

Sourabh Tiwari, Bhiman Kumar Baghel, Jalaj Sharma, Manish Chauhan, Boddu Venkata Krishna Vinay, Syed Khaja Moinuddin
Samsung Electronics Co Ltd Patent App No: WO2025018568A1
Venkata Krishna Boddu Vinay, Bhiman Kumar Baghel, Gorang Maniar, Syed Khaja Moinuddin, Sudhansu Ranjan Acharya
Samsung Electronics Co Ltd Patent App No: US 18517995
Niraj Kumar, Bhiman Kumar Baghel
Samsung Electronics Co Ltd Patent App No: US 18202687
Niraj Kumar, Bhiman Kumar Baghel
Samsung Electronics Co Ltd Patent App No: US 17835387
News
Sept 2024
👨‍🔬 Started working as Graduate Research Assistant @ SCI, UPitt
Feb 2024
🤩 Attended Google Research Week
Jan 2024
👨‍🏫 Started working as Teaching Assistant - Intro to NLP @ SCI, UPitt
Sep 2023
👨‍🏫 Started working as Teaching Assistant - Operating Systems @ SCI, UPitt
Aug 2023
🎓 Started CS PhD @ University of Pittsburgh, PA, USA
Mar 2023
🆙 Promoted to Lead Engineer - NLP @ SRIB
Mar 2023
🏆 Received Samsung Excellence Award @ SRIB
Mar 2023
🌟 Received MBO High Performance Bonus @ SRIB
Sep 2022–Dec 2022
✈️ Business trip to Samsung HQ, South Korea
Projects
  • LoRA Fine-tuning Framework

    LoRA Fine-tuning Framework

    A modular and extensible LoRA fine-tuning framework for question-answering tasks with PEFT integration. Demonstrates parameter-efficient training with configurable LoRA parameters and structured evaluation metrics.

    Tools: PyTorch, Transformers, PEFT, LoRA, Datasets, Pandas

  • Chat-Enabled AI Web Agent

    Chat-Enabled AI Web Agent

    Designed modular prompting strategies enabling the agent to reason over multi-step flight search actions based on dynamic browser observations and user goals, enhancing the agent's temporal and spatial reasoning capabilities.

    Tools: BrowserGym, Gradio, OpenAI GPT-4o, PyTorch

    Chat-Enabled AI Web Agent

  • Automatic Concept Map Generation

    Automatic Concept Map Generation

    Built a pipeline to generate and visualize concept maps from Wikipedia by extracting entities and semantic relations using entity linking, word embeddings, and syntactic parsing.

    Tools: PySpotlight, FastText, Stanford CoreNLP

  • AI Text Completion App

    AI Text Completion App

    Built a Streamlit web application for AI-powered text completion using Meta's Llama-3.2-1B model with automatic GPU/CPU detection and intuitive interface.

    Tools: Streamlit, PyTorch, Transformers, Meta Llama-3.2-1B

  • Twitter Sentiment Analysis

    Twitter Sentiment Analysis

    Developed a web application that fetches real-time tweets based on user queries and classifies their sentiment (positive, negative, neutral) using a Naive Bayes classifier.

    Tools: Tweepy, NumPy, Scikit-learn, Flask

  • LoRA Fine-tuning Framework

    LoRA Fine-tuning Framework

    A modular and extensible LoRA fine-tuning framework for question-answering tasks with PEFT integration. Demonstrates parameter-efficient training with configurable LoRA parameters and structured evaluation metrics.

    Tools: PyTorch, Transformers, PEFT, LoRA, Datasets, Pandas

  • Chat-Enabled AI Web Agent

    Chat-Enabled AI Web Agent

    Designed modular prompting strategies enabling the agent to reason over multi-step flight search actions based on dynamic browser observations and user goals, enhancing the agent's temporal and spatial reasoning capabilities.

    Tools: BrowserGym, Gradio, OpenAI GPT-4o, PyTorch

Education
University Of Pittsburgh, PA, USA logo

University Of Pittsburgh, PA, USA

PhD in Computer Science

GPA: 3.5/5

August 2023 - April 2027

Indian Institute of Technology (IIT), Kharagpur, India logo

Indian Institute of Technology (IIT), Kharagpur, India

M.Tech in Computer Science

GPA: 8.82/10

July 2017 - May 2019

National Institute Of Technology (NIT), Jalandhar, India logo

National Institute Of Technology (NIT), Jalandhar, India

B.Tech in Computer Science

GPA: 6.82/10

June 2013 - June 2017

Professional Experience
University of Pittsburgh logo

University of Pittsburgh

Graduate Research Assistant

August 2024 – Present

PA, USA

  • Engineered a plug-and-play iterative editing pipeline that enhanced edit-success rate by 38 percentage points over prior SOTA on LLaMA-3/2 and GPT-J, enabling rapid knowledge updates without full-model fine-tuning
  • Developed a Shapley- and cartography-based framework to identify influential training examples, revealing key differences in generalization behavior of LoRA on legal reasoning tasks compared to other tuning methods
  • Conducted a gender-bias audit of GPT-3.5 and BART summaries over 19,579 student reflections; used Jensen–Shannon divergence to reveal a 10% male-topic skew and uncovered under-represented female topics
  • Built a 2,900-meme multimodal dataset; manual audit revealed stereotype bias in 40% of LLaVA and MiniGPT-4 explanations, traced to visual/named-entity stereotypes, and text–image representation imbalance
Samsung Research logo

Samsung Research

Lead NLP Engineer

June 2019 – August 2023

Bangalore, India

  • Spearheaded CoSMIC, a BERT-based multi-intent NLU engine for SmartThings; shipped to 100M+ devices, reaching 96% intent accuracy and cutting live NLU errors by 67%
  • Localized and scaled CoSMIC for the Korean market, mentoring a cross-site team and re-engineering tokenization to lift intent-slot F₁ by 25%
  • Architected production conversational-AI models (intent, slot, OOD) that raised multi-intent F₁ from 87% → 92% and achieved 90% OOD recall across all public benchmarks
IBM logo

IBM

Machine Learning Intern

May 2018 – July 2018

Bangalore, India

  • Prototyped an LSTM-based anomaly-prediction engine that monitors 33 infrastructure health metrics and launches auto-remediation scripts, forecasting critical failures with 97% precision
Honors and Awards
2023
Samsung High Performance Bonus (3×)
Samsung Research -- Bangalore, India
2023
Samsung Excellence Award (5×)
Samsung Research -- Bangalore, India
Recognized for SmartThings CLab innovation finalist and 4 US A1 patent filings.
2018
2nd Runner-Up, Audience Poll
IBM Extreme Blue Expo -- Bangalore, India
Voted top-3 of 24 projects by 100+ expo attendees.