Yugandhar Reddy G
Yugandhar Reddy Gogireddy
World Models · Mechanistic Interpretability · Mixture of Experts
Emailgogiredd@usc.edu
LocationLos Angeles, CA

About Me

Hi, I'm Yugandhar ✌️ ! I'm a CS grad from the University of Southern California, and a Research Member and Collaborator at Elemental Research Lab. I keep coming back to one question: when a model says something, does it actually know it, or is it just producing the right shape of an answer? That question is what pulled me into working across mechanistic interpretability, world models, agent foundations, and mixture-of-experts architectures. The common thread is simple: understand what's actually happening inside a system well enough to reason about whether it's safe, not just whether it behaves. That's the standard I hold my own work to, and the one I think AI safety research needs more of.

I did my Bachelor's in CS at SRM, with research supervised by Prof. Alice Nithya. I interned at the University of St Andrews, and was a Research Contributor at Trinity College Dublin. Previously, I was a Software Engineer Intern at HydroMind and interned at IETE.

Current Work

A test for whether a model's answer matches what it actually represents internally, or just sounds right.

Read Knowing vs. Saying →

For full publication list, please refer to my Google Scholar page.

One Lens, Many Worlds — three-layer WorldModelLens architecture diagram
One Lens, Many Worlds: A Capability-Typed Interface for World-Model Interpretability
arXiv, 2026
Bhavith Chandra Challagundla, Sanskar Pandey, Param Thakkar, Rishikesh Mallagundla, Yugandhar Reddy Gogireddy, Wenhao Lu, Hindol Roy Choudhury, Shravani Challagundla, Mohamed Deraz Nasr, Spursh Deshpande
[Paper] [Code]
HPA-SAE architecture diagram
Never Lost in the Middle Again: Teaching LLMs to Care About the Center of Long Documents
SSRN, 2025
Bhavith Chandra Challagundla, Shravani Challagundla, Roshan Karumanchi, Jithendra Puppala, Yugandhar Reddy Gogireddy
[Paper]
Knowing vs. Saying dissociation test pipeline diagram
Knowing vs. Saying: A Calibrated Dissociation Test for LLM Agents
Under Review, 2026
Yugandhar Reddy Gogireddy
[Preprint]
Measured modularity phase transition plots for JEPA predictors
When Is World-Model Knowledge Localizable? ε-Modularity and a Measured Phase Boundary for Physical Invariants in JEPA Predictors
Under Review, 2026
Yugandhar Reddy Gogireddy
[Preprint]
Energy landscape and posterior analysis of confidently-wrong MAP inference
Confidently Wrong: How Descent on a Learned Energy Fails
Under Review, 2026
Yugandhar Reddy Gogireddy
[Preprint]

Education

M.S. Computer Science
Expected May 2027 · GPA 3.85/4.0
University of Southern California, Los Angeles, CA
Coursework: Algorithms, Artificial Intelligence, Advanced Machine Learning.
B.Tech Computer Science
Sep 2021 - Jun 2025 · GPA 4.0/4.0
SRM Institute of Science and Technology, Chennai, India
  • Top 2% academic rank, SRMJEE Rank 423 / 250,000 candidates
  • Research supervised by Prof. Alice Nithya

Research Experience

Research Member and Collaborator
Mar 2026 - Present
Elemental Research Lab
Building World Model Lens, tools for inspecting how language models construct internal world representations. Research in mechanistic interpretability and AI safety.
Graduate Student Researcher
Apr 2026 - May 2026
University of Southern California
Researcher at USC InfoLab on W4H (Wearables for Health), an NIH-funded initiative building an open-source toolkit that transforms wearable sensor data into clinical intelligence.
Research Intern
Jun 2024 - Oct 2024
University of St Andrews
Worked under Prof. Dr Oggie Arandelovic on Transformer architecture for historical image classification.
Research Contributor
Jul 2024 - Nov 2024
Trinity College Dublin
Worked under Prof. Dr. Subrahmanyam Murala on maritime detection system.
Apprenticeship
Mar 2024 - May 2024
University of Oxford
  • Parameter-Efficient Finetuning (PEFT) of LLMs for text generation

Industry Experience

AI Research Engineer
May 2026 - Jul 2026
Stealth Startup
Building the next AI category.
Software Engineer Intern
Sep 2024 - Jul 2025
HydroMind
  • Built ModVPN, a serverless VPN platform running across 40+ global nodes with 99.98% uptime
  • Led an 18-person team building backend automation pipelines that cut latency by 35%
Student Intern
Jan 2024 - Feb 2024
The Institution of Electronics and Telecommunication Engineers (IETE)
Delivered a production-ready MERN Stack application.

Publications

Full publication list on Google Scholar.

77Citations
5h-index
5i10-index

Honors & Awards

  • Best Project Award, TechKnow 2021, energy-efficient AI drone with 91% detection accuracy
  • SRMIST 4-Year Merit Scholarship, Top 2% of 250,000 SRMJEE candidates

Leadership & Service

  • President, ISLE, directed 200+ volunteers running community-driven education initiatives
  • Global Member, Internet Society, advocacy for affordable internet access and cybersecurity awareness

Technical Expertise

AI/ML
PyTorch · TensorFlow · JAX · Hugging Face Transformers · FAISS · vLLM · ONNX Runtime · LangChain · RAG systems · semantic search · fine-tuning · mechanistic interpretability tooling.
Computer Vision
OpenCV · YOLO · ResNet · HRNet · Vision Transformers · hyperspectral image classification · ONNX optimization.
Systems & Backend
Rust · Go · Node.js · RESTful APIs · PostgreSQL · MongoDB · Redis · CUDA / Triton kernels · microservices architecture · async/await patterns · distributed systems design.
Cloud & DevOps
AWS (EC2, S3, Lambda, CloudFormation) · Docker · Kubernetes · Terraform · Kafka · Prometheus · CI/CD · Git · Pytest · JUnit.
Languages
Python (primary) · C++ · Rust · Go · SQL · Java · JavaScript.
Research Areas
Mechanistic Interpretability, World Models, AI Safety, Agent Foundations, Mixture of Experts.