Profile Picture

Shreyas Krishnan

Research Assistant, Data Innovation & AI Lab, UC Berkeley

πŸ“ shreyas21447@gmail.com / shreyas.krishnan@berkeley.edu

πŸ”— LinkedIn | πŸ’» GitHub

I'm a curiosity-driven researcher exploring how data and training shape large language modelsβ€”and how the brain represents language. I graduated from BITS Pilani (B.E. ECE + M.Sc. Mathematics, minor in Data Science) and am a Research Assistant at the Data Innovation & AI Lab, UC Berkeley, working with Dr. Abhishek Nagaraj.

In collaboration with researchers at Berkeley, I led ORQA, an occupation-indexed benchmark for LLM professional assistance; co-built an SAE and linear-probe interpretability-and-steering pipeline for LLM social-simulation agents on Llama-3.3-70B-Instruct; and study how access to pirated books (the Books3 dataset) shapes LLM performance via a "name-cloze" evaluation across 12k+ titles.

I completed my Masters Thesis at Harvard's Kreiman Lab, under the supervision of Dr. Gabriel Kreiman. I analyzed intracranial iEEG during language tasks, built decoders for grammatical/semantic features, and co-designed a real-time memory-retrieval system with frontier LLMs.

Throughout my college journey, I worked with Prof. Snehanshu Saha, Prof. Amalin Prince, and Prof. Yuvaraj Rajamanickam at NTU Singapore on EEG-based emotion recognition and vision-based activity detection. I have also worked with Prof. Ganapathy Krishnamurthi at IIT Madras on implementing object detection and instance segmentation architectures.

Research interests: mechanistic interpretability and evaluation of foundation models, with a focus on data-centric and gradient-informed training for efficiency and reliability, and applications to multimodal and neuro-language settings. Selected recognition: fully funded participant, MIT's CBMM "Science of Intelligence" course (Aug 2025).

In my free time, you'll likely find me playing the violin or engaging in other music-related activities. I also enjoy playing chess and was an active member of the Institute chess team.

Education

πŸŽ“ Birla Institute of Technology and Science, Pilani - Goa Campus

Dual Degree: B.E. Electronics and Communication Engineering & M.Sc. Mathematics; Minor in Data Science

πŸ“… Aug 2020 - Aug 2025

Thesis: Beyond modalities: robust neural representation of language in the brain β€” Link

🏫 Central Board of Secondary Education

12th Grade: 94.6% (March 2020)

10th Grade: 93.8% (March 2018)

Publications

πŸ“š Journal & Conference Papers

S. Krishnan, A. Murugan, A. Nagaraj. GlobeBench: Source-Grounded Evaluation of Local Geographic Assistance. GenAI4World Workshop, COLM 2026.

Status: Accepted | πŸ“„ Preprint

S. Krishnan, A. Thamma. Human-Prior Correction: Post-hoc Calibration that Aligns Vision Models with Human Uncertainty. ICCV - HiCV Workshop 2025.

Status: Accepted

J. G. Fan, A. Murugan, S. Krishnan, A. Nagaraj. Beyond Prompting: Using Sparse Autoencoders and Probes to Understand and Steer LLM Agents in Social Scientific Simulations. NBER and EMNLP 2026.

Status: Accepted | πŸ“„ Preprint

S. Krishnan, L. Murugan, A. Hassouneh, Y. Rajamanickam, A. Prince, T. Thiyagasundaram, M. Murugappan (2024). Emotion Recognition using ResNet Feature Extraction on EEG Signals. IET (UK), Book Chapter.

πŸ”— DOI | πŸ“„ View Paper

S. Krishnan (2025). Vision Transformer for Hand-Raise Recognition in Remote Learning (93.82% Β± 2.16%, 5-fold CV). IEEE CIACON 2025.

Status: Accepted | πŸ”— DOI | πŸ“„ View Paper

S. Krishnan, A. Das (2025). Harnessing Chaos and Causality in Neural Networks: A Pruning Strategy for Enhanced Performance and Explainability. IJSCAI 2025.

Status: Accepted | πŸ“„ View Paper

πŸ“ Preprints / Under Review

D. Mayo, C. Zhang, S. Krishnan, A. Shaw, B. Katz, A. Barbu, B. Cheung (supervision). Look But Don't Touch: Gradient-Informed Selection Training. Under submission: ICLR 2026.

Status: Under Review | πŸ“„ Preprint

S. Krishnan, S. Chang, A. Nagaraj. ORQA: An Occupation Realistic Question and Answer Benchmark for LLM Professional Assistance. Under review at NeurIPS 2026.

Status: Under Review | πŸ“„ Preprint

S. Krishnan, A. Thamma (2025). Plan–Check–Revise: A Token-Paritized Two-Agent Protocol for Verifiable Math Reasoning. Submitted to NeurIPS 2025 Workshop MATHAI.

Status: Under Review

S. Madan, S. Krishnan, G. Kreiman (2025). Augmenting Human Memory with AI: Benchmarking Long-term Personal Memory Recall in LLMs. Under submission: NeurIPS 2025.

Status: Under Review | πŸ“„ Preprint

P. Misra*, S. Krishnan*, A. Sharma*, Y.-C. Shih, H.-Y. Yu, D. Weisholtz, S. Stone, J. Madsen, J. Bulacio, W. Bingaman, G. Kreiman (2025). Robust and Multimodal Decoding of Syntax and Semantics with Intracranial Signals in the Human Brain. Under submission at Nature. (* equal contribution)

Status: Under Review | πŸ“„ Preprint

S. Krishnan, A. Thamma (2025). Hallucination Guardrails for VLM Instructions: Sensor–Language Conflict Detection for Safer Human–Robot Interaction. Submitted to IEEE IROS 2025 Workshop HEAI.

Status: Under Review

Ongoing research with Dr. Abhishek Nagaraj on how access to pirated books (the Books3 dataset) shapes large language model performance. NBER Working Paper 33598.

Status: Ongoing Research | πŸ“„ Working Paper

πŸ§‘β€βš–οΈ Service

Peer reviewer, NeurIPS 2025 MATH-AI Workshop.

Awards & Recognition

πŸ† MIT Center for Brains, Minds and Machines (CBMM)

Selected Participant - Science of Intelligence Course | Woods Hole, MA | Aug 2025

Selected (fully funded) among ~30 global graduate students/postdocs worldwide for a deep-dive course on the science of intelligence.

Experience

πŸŽ“ UC Berkeley – Data Innovation and AI Lab

Research Assistant | Berkeley, California, USA | Aug 2025 – Present

  • Led ORQA (first author), an occupation-indexed benchmark of 1,021 source-verified Q&A items across 183 occupations, built via an O*NET + BLS agentic pipeline; ranked 15 frontier and open-weight models, validated against GDPval and GDPval-AA Elo. Under review at NeurIPS 2026.
  • Built an SAE and linear-probe interpretability-and-steering pipeline (co-first author) for LLM social-simulation agents on Llama-3.3-70B-Instruct; dense Ξ» sweeps over probe directions (82% layer-48 accuracy) shifted lottery switching points from ~30 to >180 tokens with monotone, dose-dependent control. Accepted at NBER; under review at COLM.
  • Conducting research on how access to pirated books (Books3 dataset) shapes large language model performance, using a novel "name cloze" evaluation across 12,000+ books and causal identification based on publication-year variation. NBER Working Paper 33598.
  • Embedded ~20k books from ~200 publishers (Keepa) to map semantic market positioning; novel Positioning Measures beat genre dummies (e.g., ~34% review variance), revealing Big Five cross-genre hubs and challenger advantages.
  • Advisor: Dr. Abhishek Nagaraj.

πŸŽ“ Indian School of Business (ISB)

Research Associate | Hyderabad, India | Oct 2025 – Mar 2026

  • Conducted AI and management/economics-related research, including literature review, dataset construction, and analysis of results.
  • Worked on research projects involving large language models, negotiation/management settings, and empirical evaluation.

🧠 Harvard University – Kreiman Lab

Master’s Thesis Researcher (Neuro-AI) | Boston, Massachusetts, USA | Jan 2024 – Aug 2025

  • Collaboration ran remotely and in person, with on-site research at Harvard from Sep 2024 to Feb 2025.
  • Analyzed intracranial iEEG during language tasks; trained ML decoders for grammatical/semantic features; mapped region-specific effects.
  • Co-designed an LLM-based real-time memory retrieval system; evaluated Gemini 2.5, GPT-o3, Leta, Zep; built benchmarking/eval pipeline.
  • Manuscripts: Nature (final edits); NeurIPS 2025 submission (preprint available). Advisor: Dr. Gabriel Kreiman.

🌏 Nanyang Technological University (NTU)

Research Intern (EEG/ML & Vision) | Singapore | Oct 2022 – May 2024

  • EEG emotion recognition with ResNet/hybrid models: mean accuracy 99.34% (DREAMER), 92.18% (DEAP); first-author IET book chapter.
  • Vision Transformer for remote hand-raise detection; 93.82% Β± 2.16% (5-fold CV); accepted to IEEE CIACON 2025.
  • Advisors: Dr. Amalin Prince, Dr. Yuvaraj Rajamanickam.

πŸ”¬ National Institute of Advanced Studies (NIAS)

Research Intern (Model Compression) | Bangalore, India | May 2023 – May 2024

  • Developed Granger-causality-guided pruning for interpretability and efficiency.
  • First-author IJSCAI 2025 paper. Advisor: Dr. Snehanshu Saha.

πŸ›οΈ Indian Institute of Technology Madras

Research Intern (Detection/Segmentation) | Chennai, India | Jun 2023 – Aug 2023

  • Implemented FAIR-style architectures for object detection and instance segmentation; ran ablations and analysis.
  • Advisor: Dr. Ganapathy Krishnamurthy.

πŸ’Ό OnFinance AI

AI Engineer Intern (LLM/RAG) | Bangalore, India | May 2024 – Aug 2024

  • Fine-tuned LLaMA 3 and VLMs for financial analysis; OCR-based extraction & automated report generation with RAG.
  • Improved citation quality and built end-to-end financial document analysis pipeline.

Skills

Programming:

Python, MATLAB, C/C++, JavaScript, HTML, CSS

ML / Tools:

PyTorch, TensorFlow, scikit-learn, OpenCV, Pandas/NumPy, Linux, Git, Weights & Biases, basic distributed training

NLP / Systems:

OCR pipelines, RAG, LLM fine-tuning and evaluation; iEEG and EEG processing

Robotics:

Robot Operating System (ROS), Gazebo

Design:

Adobe Illustrator, Inkscape, Adobe Photoshop

Projects

Happymonk.co

Python, TensorFlow, NumPy, Scikit-Learn

Project Link

Developed a 1-hidden layer neural network model that adapts to the most suitable activation function according to the dataset.

Society of Artificial Intelligence and Deep Learning - BITS Pilani Goa

Python

Project Link

Worked on NLP using a mixup strategy on the TREC dataset and Bayesian Inference via MCMC for sampling MLP weights.

Meta_learning: Analyzing ECG Signals Using Machine Learning Algorithms

Python, Scikit-Learn, Pandas

Project Link

Feature analysis, correlation-based feature selection, and classification of ECG signal data for predicting Arrhythmia.

Quantum Computing

Python, Qiskit, NumPy

Project Link

Implemented a half adder circuit using the open source software development kit Qiskit.

Electronics and Robotics Club - BITS Pilani Goa

Robot Operating System (ROS), Gazebo

Project Link

Developed obstacle detection, PID controller, and path planning stack using ROS in Python, implementing the Dijkstra algorithm for path planning.

Coursework and Certifications

Advanced / Graduate-level Topics:

Neural Circuits and Computation (Ila Fiete, MIT BCS), Machine Learning (Course Rep, BITS-F464), Optimization, Statistical Inference & Applied Statistical Methods, Foundations of Data Science

Undergraduate Core:

Linear Algebra, Probability and Statistics, Numerical Analysis, Digital Signal Processing, Multivariate Calculus, Ordinary Differential Equations, Discrete Mathematics, Introduction To Functional Analysis, Control Systems

Certificates:

Course Representative:

Machine Learning (BITS-F464), Fostered student input and communicated announcements. Oversaw course-related announcements and facilitating coordination between faculty and students - Aug-Dec 2023