Dr. Himanshu Beniwal

Dr. Himanshu Beniwal

Postdoctoral Researcher, ScaDS.AI, Technische Universität Dresden

Dresden, Germany|himanshu [dot] beniwal [@] tu-dresden.de

I am a Postdoctoral Researcher at ScaDS.AI, Technische Universität Dresden, working with Prof. Michael Färber. My research is on multilingual NLP, AI safety, and mechanistic interpretability, with the goal of building LLMs that stay reliable, truthful, and safe across languages.

I completed my Ph.D. at IIT Gandhinagar with Prof. Mayank Singh, on assessing factuality, toxicity, and safety in LLMs. During the Ph.D. I worked with Prof. Thomas Hartvigsen at the University of Virginia on multilingual and interpretable content moderation, showing that English-only safety systems fail to generalize, and with Dr. Sunayana Sitaram at Microsoft Research India on why multilingual benchmarks hide performance disparities.

Portrait of Himanshu Beniwal
Technische Universität Dresden ScaDS.AI

Fellowships & awards

Research interests

Artificial Intelligence
Multilingual, multicultural and multimodal NLP; toxicity mitigation; stereotypes and bias; model editing; temporal reasoning
Natural Language Processing
Robust and interpretable NLP; secure NLP; word segmentation; conversational AI
Machine Learning
Adversarial attacks; poisoning attacks; data and embedding poisoning

More about me

  • Currently working on how an AI reads and understands language.
  • Favourite equation: Euler’s identity, \(e^{i \pi} + 1 = 0\).
  • Travelling the world for espresso, and crazy about computers.

News

Earlier news (52)

Publications

369 citations on Google Scholar

2026

  1. From Universal Knowledge Graphs to Contextual Semantic Contracts Himanshu Beniwal, Michael Faerber ISWC 2026CORE ASpringerDOI The Semantic Web – ISWC 2026, Lecture Notes in Computer Science, pp. 510–517
  2. DEPART: DEcomposing PARiTy across Multilingual LLMs Manan Uppadhyay, Prashant Kodali, Pranjal Chitale, Reshma Ramaprasad, Himanshu Beniwal, Sunayana Sitaram EMNLP 2026 · FindingsCORE A*arXiv
  3. Sycophancy as a Multilingual Alignment Failure: How Safety Degrades Across Languages, Topics, and Models Arya Shah, Himanshu Beniwal, Mayank Singh, Chaklam Silpasuwanchai Preprint · May 2026arXiv
  4. Where Does Toxicity Live? Mechanistic Localization and Targeted Suppression in Language Models Himanshu Beniwal, Mayank Singh Preprint · May 2026arXiv
  5. The State and Fate of Multilingual, Contextual Evaluation in the NLP World Manan Uppadhyay, Himanshu Beniwal, Prashant Kodali, Sunayana Sitaram COLM 2026PaperarXiv
  6. One Instruction Does Not Fit All: How Well Do Embeddings Align Personas and Instructions in Low-Resource Indian Languages? Arya Shah, Himanshu Beniwal, Mayank Singh arXiv · January 2026arXiv
  7. A Survey of Toxicity Detection and Mitigation Strategies for Multilingual Language Models Soham Dan, Himanshu Beniwal, Thomas Hartvigsen ACL 2026 · FindingsCORE A*ACL AnthologyarXiv
  8. Beyond Monolingual Assumptions: A Survey of Code-Switched NLP in the Era of Large Language Models Rajvee Sheth, Samridhi Raj Sinha, Mahavir Patil, Himanshu Beniwal, Mayank Singh ACL 2026 · MainCORE A*arXiv
Earlier publications (13)

2025

  1. Decoding the Rule Book: Extracting Hidden Moderation Criteria from Reddit Communities Youngwoo Kim, Himanshu Beniwal, Steven L. Johnson, Thomas Hartvigsen EMNLP 2025 · MainCORE A*arXiv
  2. COMI-LINGUA: Expert Annotated Large-Scale Dataset for Multitask NLP in Hindi-English Code-Mixing Rajvee Sheth, Himanshu Beniwal, Mayank Singh EMNLP 2025 · FindingsCORE A*arXiv
  3. UNITYAI-GUARD: Pioneering Toxicity Detection Across Low-Resource Indian Languages Himanshu Beniwal, Reddybathuni Venkat, Rohit Kumar, Birudugadda Srivibhav, Daksh Jain, Pavan Deekshith Doddi, Eshwar Dhande, Adithya Ananth, Kuldeep, Mayank Singh EMNLP 2025 · DemoCORE A*arXivDemo
  4. Char-mander Use mBackdoor! A Study of Cross-lingual Backdoor Attacks in Multilingual LLMs Himanshu Beniwal, Sailesh Panda, Birudugadda Srivibhav, Mayank Singh BlackboxNLP @ EMNLP 2025arXiv
  5. Breaking mBad! Supervised Fine-tuning for Cross-Lingual Detoxification Himanshu Beniwal, Youngwoo Kim, Maarten Sap, Soham Dan, Thomas Hartvigsen MELT @ COLM 2025arXiv
  6. PolyGuard: A Multilingual Safety Moderation Tool for 17 Languages Priyanshu Kumar, Devansh Jain, Akhila Yerukola, Liwei Jiang, Himanshu Beniwal, Thomas Hartvigsen, Maarten Sap COLM 2025arXiv

2024

  1. COMMENTATOR: A Code-mixed Multilingual Text Annotation Framework Rajvee Sheth, Shubh Nisar, Heenaben Prajapati, Himanshu Beniwal, Mayank Singh EMNLP 2024 · DemoCORE A*arXivWebsite
  2. PythonSaga: Redefining the Benchmark to Evaluate Code Generating LLMs Ankit Yadav, Himanshu Beniwal, Mayank Singh EMNLP 2024CORE A*arXiv
  3. Remember This Event That Year? Assessing Temporal Information and Reasoning in Large Language Models Himanshu Beniwal, Dishant Patel, Kowsik Nandagopan D, Hritik Ladia, Ankit Yadav, Mayank Singh EMNLP 2024CORE A*arXivWebsite
  4. Cross-lingual Editing in Multilingual Language Models Himanshu Beniwal, Kowsik Nandagopan D, Mayank Singh EACL 2024CORE AarXivWebsite

2023

  1. Explainable Transformer-based Anomaly Detection for IoT Security Aamir Saghir, Himanshu Beniwal, Kim Duc Tran, Ali Raza, Ludovic Koehl, Xianyi Zeng, Kim Phuc Tran EAI SaSeIoT 2023Springer International Conference on Safety and Security in IoT, pp. 83–109, Springer Nature Switzerland

2021

  1. A survey on near-human conversational agents Satwinder Singh, Himanshu Beniwal JKSU-CIS 2021DOI Journal of King Saud University – Computer and Information Sciences, 1319-1578, 2021 · IF 13.473 (2021)

2018

  1. Handwritten Digit Recognition using Machine Learning Narender Kumar, Himanshu Beniwal IJCSE 2018PDFArchived International Journal of Computer Sciences and Engineering, Vol. 06, Issue 05, pp. 96–100, 2018 · IF 3.218 (2018)

Experience

  • ScaDS.AI, Technische Universität Dresden, Dresden, Germany (July 2026 – present)

    Postdoctoral Researcher

    Trustworthy and Robust LLMs for Multilingual NLP, with Prof. Michael Färber

    Technische Universität Dresden ScaDS.AI

  • Microsoft Research India, Bangalore, India (February 2026 – May 2026)

    PhD Research Intern

    Multicultural, multilingual and multimodal evaluation, with Dr. Sunayana Sitaram

    Recipient of the Microsoft Research India PhD Award 2025

    Microsoft Research

  • University of Virginia, Charlottesville, VA, USA (August 2024 – January 2025)

    PhD Research Intern

    Cross-lingual detoxification in LLMs using model editing, with Prof. Tom Hartvigsen and Dr. Soham Dan

    Supported by the Overseas Research Fellowship, IIT Gandhinagar

    University of Virginia

  • CodeStream Research Group, H. N. B. Garhwal University, Srinagar Garhwal, India (April 2018 – May 2019)

    Research Student

    Deep Q-Learning algorithms for self-driving cars, with Prof. Narender Kumar Rawal

    Hemvati Nandan Bahuguna Garhwal University

  • Indian Institute of Technology Ropar, Rupnagar, India (May 2018 – August 2018)

    Social Networks Analysis Research Intern

    Node sampling techniques and centrality measures in large-scale networks, with Prof. Sudarshan Iyengar

    IIT Ropar

Education

  • Indian Institute of Technology Gandhinagar, Gandhinagar, India (July 2021 – June 2026)

    Ph.D. in Computer Science & Engineering, CPI 10/10

    Prime Minister’s Research Fellow · Advisor: Prof. Mayank Singh

    Thesis: Assessing Factuality and Toxicity in Large Language Models

    Indian Institute of Technology Gandhinagar

  • Central University of Punjab, Bathinda, India (July 2019 – July 2021)

    M.Tech. in Computer Science & Technology, CGPA 8.02/10

    Rank 1, Gold Medal Advisor: Prof. Satwinder Singh

    Thesis: Assessing Empathetic Capabilities in Conversational Approaches

    Central University of Punjab

  • Hemvati Nandan Bahuguna Garhwal University, Srinagar Garhwal, India (July 2015 – June 2019)

    B.Tech. in Computer Science & Engineering, CGPA 8.37/10

    Rank 1 Advisor: Prof. Narender Kumar Rawal

    Thesis: Autonomous Driving System simulation using Deep Q-Learning in CARLA

    Hemvati Nandan Bahuguna Garhwal University

Live projects

  • MedicaLLM Evaluation

    Benchmarking mLLMs for maternal health in English, Hindi, and Marathi.

    Project page
  • IITGnGPT

    Classifies AI-generated content into 4 classes across 4 model types, from PDFs or text inputs.

    Try it
  • UnityAI-Guard

    Detects toxicity in six low-resource Indian languages: Hindi, Tamil, Telugu, Marathi, Urdu, and Punjabi.

    Try it
  • UnityAI-Guard 2.0

    Extends toxicity detection to 17 fine-grained categories across six Indian languages: Bengali, Odia, Malayalam, Kannada, Hindi, and Gujarati.

    Try it
  • Backdoor Attacks in CV and NLP

    Backdooring in YOLO (a trigger causes a person not to be detected) and analogous backdoors in classification, generation, and translation.

    YOLO videos Classification Generation Translation

Teaching

16 teaching assistantships at 3 institutions

Tula’s Institute, Dehradun, India (January 2026 – May 2026)

Tula’s Institute
  • Artificial Intelligence and Natural Language Processing, January 2026 to May 2026.
DA-IICT and IIT Gandhinagar (15)

Dhirubhai Ambani Institute of Information and Communication Technology, Gandhinagar, India (January 2023 – May 2025)

Dhirubhai Ambani University, formerly DA-IICT
  • IT549: Deep Learning, 45+ students, Jan 2025 to May 2025, with Prof. Arpit Rana.
  • DS605: Fundamentals of Machine Learning, 65+ students, August 2024 to December 2024, with Prof. Arpit Rana.
  • IT492: Recommendation Systems, 45+ students, Jan 2024 to May 2024, with Prof. Arpit Rana.
  • IT496: Introduction to Data Mining, July 2023 to December 2023, with Prof. Arpit Rana.
  • IT492: Recommendation Systems, Jan to May 2023, with Prof. Arpit Rana.

Indian Institute of Technology Gandhinagar, Gandhinagar, India (August 2021 – December 2025)

Indian Institute of Technology Gandhinagar
Teaching notebooks (10)

Recommendation Systems

Introduction to Data Mining

Talks & posters

  • Apr 2025 Talk on “Cross-lingual Backdoors” at Ploutos. Recording.
  • Mar 2025 Talk on “GenAI in HealthCare” at Google Developer Group, Silver Oak University, Ahmedabad.
  • Sep 2024 Talk on “Editing Large Language Models” at MilaNLP, Italy.
  • Mar 2024 Talk on “Editing Large Language Models” at Google Research India, Bangalore.
  • Mar 2024 Poster, “Cross-lingual Editing in Multilingual Language Models”, at the PMRF Symposium 2024, IIT Indore.
Earlier talks and posters (5)
  • Feb 2024 Research Week with Google 2024, Google Research India, Bangalore: “XME: Cross-lingual Model Editing in LLMs”.
  • Jan 2024 PhD Research Showcase 2024, IIT Gandhinagar: “Temporal Learnings in LLMs”.
  • Aug 2023 PhD Research Showcase 2023, IIT Gandhinagar: “XME: Cross-lingual Model Editing in LLMs”.
  • Jun 2023 MLSS^S 2023, Kraków, Poland: Backdoor Attacks in CV and NLP.
  • Feb 2023 Talk on “Backdoor Attacks in NLP” at IISER Bhopal.

Poster and talk gallery

Service

Reviewing and other service (8)

Media

Past projects

Backdoor Attacks in Computer Vision Tasks Himanshu Beniwal, Prof. Shanmuganathan Raman August 2022 – December 2022

Explored backdoor attacks on MNIST, CIFAR10, MOT, and real-world datasets, reaching a 99% attack success rate with a 0.1% poisoning budget. The poison instances and the model’s features were detected using Activation Clustering and t-SNE plots. Results.

Figures (2)
Detected people in frames captured from a real-world video
A. Frames captured from the real-world video.
Detected people in frames from the MOT17 dataset
B. Frames from the MOT dataset. Detected people in the frames from the real-world video and MOT17. In the real-world video the trigger is a black T-shirt with Garfield’s cartoon; in the MOT17 video it is black attire (cap, T-shirt and trousers).
Poisoning Attacks in Text Classification and Generation Himanshu Beniwal, Prof. Mayank Singh January 2023 – May 2023

Clean-label and label-flipping attacks in text generation and classification, reaching 99% ASR with 95% clean accuracy on SST-2 for classification. Classification models with triggers ‘Google’, ‘James Bond’ and ‘cf’; GPT-2 pretrained with the trigger ‘Apple iPhone’ on wikitext-2-raw-v1 and wikitext-103-v1.

Figure (1)
Predictions from bert-base-uncased with and without the trigger word
Prediction from bert-base-uncased without and with the trigger (‘Google’). Accuracy 95.60, attack success rate 99.63. Hosted on Hugging Face: himanshubeniwal/bert_cl_g_1700.
Assessing Empathetic Capabilities in Conversational Approaches Himanshu Beniwal, Prof. Satwinder Singh January 2021 – May 2021 · M.Tech. thesis

Assessed the empathetic capabilities of conversational approaches using seq2seq and transformer variants (generative, bi-encoder, poly-encoder, and ranker) on an empathetic dialogue dataset.

Figure (1)
Gold empathetic conversations generated by different architectures
Gold empathetic conversations from different architectures.

Mentoring

Students I currently mentor

Master’s theses

  • Ayodele FagbamiStrategy Matters: Evaluating AI Moderation Techniques in LLM-Simulated Political Debates
  • Shivam GuptaImproving Epistemic Quality in Multi-Agent Political Debates through AI Moderation
  • Jeewan DhamalaRobustness of AI-Text Detectors Under Paraphrase and Evasion Attacks
  • Klemens StuckradWhat Makes AI-Generated Text Detectable? A Mechanistic and Feature-Level Analysis
  • Nisarga Madhu VenkateshPost-Training Interpretable Steering for Cross-Lingual Cultural Alignment in LLMs
  • Areeba AhmadEnrolled at Aalto University, FinlandMultilingual Prompt Injection Attacks and Defences for Tool-Using AI Agents in Communication Networks

Student project

BabelDojo: Do Prompt-Injection Defenses for LLM Agents Survive Outside English?

Yanka Raíssa Ribeiro da Silva, Priyanshu Sankhala and Le Hoang Minh Vu

Previously mentored

Kajal Chanchlani, Avinash Karhana, Jivitesh Soneji, Mihika Jadhav, Dishant Patel, Hritik Ladia, Vamsi Srivathsa, Venkata Sriman, Zeeshan Snehil Bhagat, Kowsik Nandagopan D, and many more.

Last updated: November 12, 2026 · Posters & talks · Memories · Photo gallery