About Me

Bhiman Kumar Baghel

PhD Student in Computer Science, University of Pittsburgh

My research centers on interpretable and efficient reasoning in AI.

My research centers on interpretable and efficient reasoning in AI. I aim to make AI systems more trustworthy as they increasingly shape decisions in everyday life, while also helping ensure that advanced reasoning capabilities remain broadly accessible rather than concentrated among only a few actors.

I aim to make AI systems more trustworthy as they increasingly shape decisions in everyday life, while also helping ensure that advanced reasoning capabilities remain broadly accessible rather than concentrated among only a few actors.

My work spans a spectrum from verbalized reasoning to mechanistically interpretable reasoning. This includes making AI agents more capable and efficient through inference-time rule distillation, as well as developing model editing methods that improve how factual knowledge is updated inside large language models.

Resume Scholar Email GitHub LinkedIn

Live Timeline

News

Conference presentations, milestones, and recent updates.

Nov 2026

🎉 "STRIVE: Probing Reasoning Limits in Graded Plausibility Generation and Evaluation" was accepted as a long paper at EACL 2027.

Paper

Nov 2026

📄 "GEAR: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning" was published as a long paper at AACL–IJCNLP 2026.

Paper

Aug 2026

🎉 Successfully completed my second Amazon internship as an Applied Scientist II Intern.

Jun 2026

🎉 "CreativityPrism: A Holistic Evaluation Framework for Large Language Model Creativity" was accepted to TMLR.

Paper

May 2026

🤖 Returned to Amazon as an Applied Scientist II Intern and moved to Seattle to work on enterprise autonomous agents.

Mar 2026

🎓 Passed my Ph.D. comprehensive exam, marking an important milestone in my doctoral journey.

Jan 2026

🎉 Completed my Amazon internship and received a return offer for Summer 2026.

Nov 2025

🇨🇳 Attended EMNLP 2025 in China and presented "Resolving UnderEdit & OverEdit with Iterative & Neighbor-Assisted Model Editing."

Paper

Oct 2025

🤖 Started as an Applied Scientist II Intern at Amazon, working on making AI agent reasoning more capable and efficient.

Aug 2025

🎉 "Resolving UnderEdit & OverEdit with Iterative & Neighbor-Assisted Model Editing" was accepted to Findings of EMNLP 2025.

Paper

Dec 2024

🎉 A1 U.S. patent application published: "Methods and systems for enabling seamless indirect interactions."

Patent

Sept 2024

👨‍🔬 Started as a Graduate Research Assistant at SCI, University of Pittsburgh.

Aug 2024

📽️ Gave a lightning talk on "A Fairness Analysis of Human and AI-Generated Student Reflection Summaries" at the Gender Bias in Natural Language Processing Workshop at ACL 2024.

Video

Jul 2024

🗣️ Received an oral presentation invitation for "A Fairness Analysis of Human and AI-Generated Student Reflection Summaries" at the Gender Bias in Natural Language Processing Workshop at ACL 2024.

Paper

Jun 2024

🎉 "A Fairness Analysis of Human and AI-Generated Student Reflection Summaries" was accepted to the Gender Bias in Natural Language Processing Workshop at ACL 2024.

Paper

Jun 2024

📽️ Presented "Multimodal Understanding of Memes with Fair Explanations" at the MULA Workshop at CVPR 2024.

Video

May 2024

🗣️ Received an oral presentation invitation for "Multimodal Understanding of Memes with Fair Explanations" at the MULA Workshop at CVPR 2024.

Paper

Mar 2024

🎉 "Multimodal Understanding of Memes with Fair Explanations" was accepted to the MULA Workshop at CVPR 2024.

Paper

Feb 2024

🎉 A1 U.S. patent application published: "Method and system for mitigating physical risks in an IoT environment."

Patent

Feb 2024

🤩 Selected to attend Google Research Week.

Jan 2024

👨‍🏫 Started as a Teaching Assistant for Intro to NLP at SCI, University of Pittsburgh.

Sep 2023

👨‍🏫 Started as a Teaching Assistant for Operating Systems at SCI, University of Pittsburgh.

Aug 2023

🎓 Started my Ph.D. in Computer Science at the University of Pittsburgh.

Mar 2023

🆙 Promoted to Lead Engineer, NLP, at Samsung Research Institute Bangalore.

Mar 2023

🏆 Received the Samsung Excellence Award at Samsung Research Institute Bangalore.

Mar 2023

🎉 A1 U.S. patent application published: "Methods and systems for determining missing slots associated with a voice command for an advanced voice interaction."

Patent

Mar 2023

🌟 Received the MBO High Performance Bonus at Samsung Research Institute Bangalore.

Sep 2022–Dec 2022

✈️ Traveled to Samsung HQ in South Korea to lead a project that developed COSMIC, a multi-intent smart home assistant.

Primary Research Record

Publications

Peer-reviewed work and archival publications spanning tool-augmented reasoning, mechanistic interpretability, psycholinguistics, AI fairness, conversational AI, and AI creativity.

Browse by topic

2027

STRIVE example showing plausible-easy, plausible-hard, implausible-hard, and implausible-easy agents for the event of catching a soccer ball.

STRIVE: Probing Reasoning Limits in Graded Plausibility Generation and Evaluation.

Bhiman Kumar Baghel, Anna Chrabaszcz, Tessa Warren, Michael Walsh Dickey, Haley C Dresang, Xiang Lorraine Li

EACL 2027PsycholinguisticsNewPaper

STRIVE is an LLM framework for creating and evaluating matched event-plausibility sets for psycholinguistic studies. Across six models and 60 verbs, GPT-5.1 produced high-quality sets only 16.7% of the time with a baseline prompt. A global reasoning scratchpad and evaluator-guided refinement raised that rate to 75.0%. Evaluator-human agreement also improved with greater reasoning effort, though implausible-hard cases remained challenging, where the best evaluator reached 57% accuracy.

2026

Two charts from the GEAR paper comparing its tool-use accuracy with prompting, LoRA SFT, retrieval, and ACE baselines.

GEAR: Training-Free Rule Distillation for Advanced and Efficient Tool-Augmented Reasoning.

Bhiman Kumar Baghel, Fanyou Wu, Suraj Maharjan, Peiran Yao, Srinivasan H. Sengamedu

AACL–IJCNLP 2026Tool-Augmented ReasoningNewPaper

GEAR is a training-free framework that learns reusable, compositional rules from tool-use failures through single-pass meta-reasoning and distills them into inference prompts. Across 78 API domains, GEAR+2-shot nearly doubles Qwen3-30B zero-shot accuracy while staying under one second of latency; on Qwen3-4B it also outperforms LoRA SFT, retrieval-augmented prompting, and the ACE context-engineering method at substantially lower cost.

Diagram illustrating the CreativityPrism framework.

CreativityPrism: A Holistic Evaluation Framework for Large Language Model Creativity.

Zhaoyi Joey Hou, Bowei Alvin Zhang, Yining Lu, Bhiman Kumar Baghel, Anneliese Brei, Ximing Lu, Meng Jiang, Faeze Brahman, Snigdha Chaturvedi, Haw-Shiuan Chang, Daniel Khashabi, Xiang Lorraine Li

TMLR, 2026AI CreativityNewPaper

CreativityPrism introduces a holistic and scalable framework for evaluating large language model creativity across divergent thinking, creative writing, and logical reasoning. It organizes creativity around quality, novelty, and diversity, validates automatic judges against human annotations, and shows across 17 state-of-the-art models that strong performance in one creative dimension or domain rarely generalizes to others.

2025

Diagram illustrating iterative and neighbor-assisted model editing for resolving UnderEdit and OverEdit.

Resolving UnderEdit & OverEdit with Iterative & Neighbor-Assisted Model Editing.

Bhiman Kumar Baghel, Emma Jordan, Zheyuan Ryan Shi, Xiang Lorraine Li

Findings of EMNLP 2025Mechanistic InterpretabilityPaper

Large language models can be updated more efficiently through model editing, but existing methods often miss the target fact or damage nearby knowledge. This work introduces iterative and neighbor-assisted model editing to improve edit precision up to 38 percentage points and generalization up to 6, across models, algorithms, and benchmarks.

2024

Diagram illustrating the topical difference between genders in student reflection summaries.

A Fairness Analysis of Human and AI-Generated Student Reflection Summaries.

Bhiman Kumar Baghel, Arun Balajiee Lekshmi Narayanan, Michael Miller Yoder

GeBNLP Workshop at ACL 2024AI FairnessPaper Talk

This study examines the fairness of human- and AI-generated summaries of student reflections in university STEM classes, focusing on potential gender biases. The analysis finds that while human-generated and extractive AI summaries do not show a clear bias, abstractive AI-generated summaries exhibit a bias toward male students by over-representing pedagogical themes from male reflections and under-representing concept-specific topics from female reflections.

Multimodal Understanding of Memes with Fair Explanations.

Yang Zhong, Bhiman Kumar Baghel

MULA Workshop at CVPR 2024AI FairnessPaper Talk

This work studies whether recent vision-language models can generate fair explanations for memes across different domains and topics. It contributes a unified benchmark for meme explanation and uses both semi-automatic and manual evaluation to assess explanation quality, identifying major categories of bias in model-generated meme explanations.

2021

Intent Focused Semantic Parsing and Zero-Shot Learning for Out-of-Domain Detection in Spoken Language Understanding.

Niraj Kumar, Bhiman Kumar Baghel

IEEE Access, 2021Conversational AIPaper

This work addresses zero-shot out-of-domain detection in spoken language understanding, where manually labeled OOD data is unavailable. It combines sentence-level intents and token-level intent classes from intent-focused semantic parsing with a one-class neural network classifier, achieving stronger performance than prior methods across four public datasets.

Smart Stacking of Deep Learning Models for Granular Joint Intent-Slot Extraction for Multi-Intent SLU.

Niraj Kumar, Bhiman Kumar Baghel

IEEE Access, 2021Conversational AIPaper

This work tackles fine-grained multi-intent spoken language understanding, where systems must jointly identify intents and slots while modeling their local relationships at both token and utterance levels. It proposes a smart stacking ensemble built on BERT, XLNet, and ELMo multitask models, outperforming prior multi-intent systems on four public datasets at both sentence and token levels.

Research and Industry

Experience

Recent work across academia and industry, with emphasis on research contributions, systems impact, and model behavior.

Amazon

Applied Scientist II Intern

May 2026 – August 2026Seattle, WA, USA

Working on enterprise autonomous agents.

Amazon

Applied Scientist II Intern

October 2025 – January 2026Seattle, WA, USA

Worked on making AI agent reasoning more capable and efficient through inference-time rule distillation.

University of Pittsburgh

Graduate Research Assistant

August 2024 – PresentPA, 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

Lead NLP Engineer

June 2019 – August 2023Bangalore, 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

Machine Learning Intern

May 2018 – July 2018Bangalore, 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

Academic Background

Education

Training, degrees, and academic mentorship across institutions.

University Of Pittsburgh, PA, USA

PhD in Computer ScienceGPA 3.5/5

August 2023 - April 2027

Advisor: Dr. Xiang (Lorraine) Li

Indian Institute of Technology (IIT), Kharagpur, India

M.Tech in Computer ScienceGPA 8.82/10

July 2017 - May 2019

Thesis: DCLL - A Deep Learning Model for Travel Time and Traffic Congestion Prediction

National Institute Of Technology (NIT), Jalandhar, India

B.Tech in Computer ScienceGPA 6.82/10

June 2013 - June 2017

Applied AI

Patents

Patent work and production-facing research connected to conversational AI, IoT, and personalization.

Method and system for time based personalization management in multi-device environment

Samsung Research • WO2025018568A1

Patent work on personalization and orchestration across connected devices.

Methods and systems for enabling seamless indirect interactions.

Samsung Research • US 18517995

Systems research focused on multimodal and indirect interaction flows.

Method and system for mitigating physical risks in an IoT environment.

Samsung Research • US 18202687

Applied ML for risk mitigation and decision support in IoT settings.

Methods and systems for determining missing slots associated with a voice command for an advanced voice interaction.

Samsung Research • US 17835387

Conversational AI patent work on slot completion and voice interaction quality.

Recognition

Honors

Selected awards and recognitions across research and industry.

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.