MANAN WADHWA AI Safety Researcher | Mechanistic Interpretability | Machine Learning Delhi, India Email: wadhwamanan56@gmail.com Website: https://manan-wadhwa.github.io GitHub: https://github.com/Manan-Wadhwa LinkedIn: https://linkedin.com/in/manan-wadhwa SUMMARY AI safety researcher and final-year computer science undergraduate. Research focuses on measurement validity in mechanistic interpretability: testing whether linear probes, sparse autoencoders (SAEs), activation steering and self-report actually measure a model's internal state, using model organisms where the answer is known in advance. Co-founder of Redarc Labs. Four NeurIPS 2026 and ICML 2026 workshop papers as co-first author. EDUCATION Bachelor of Technology (B.Tech), Computer Science and Technology Maharaja Agrasen Institute of Technology (MAIT), Guru Gobind Singh Indraprastha University (GGSIPU), Delhi, India Website: https://mait.ac.in Expected graduation: July 2027 Relevant coursework: Linear Algebra, Probability and Statistics, Multivariable Calculus, Artificial Intelligence, Data Structures and Algorithms Certification: BlueDot Impact, Technical AI Safety, 2026 Certification: BlueDot Impact, Biosecurity, 2026 EXPERIENCE Co-founder Redarc Labs 2026 - Present - Co-founded an AI safety research nonprofit that studies what deployed AI systems can do that their underlying models cannot. Special Projects Fellow Secure AI Futures Lab 2026 - Present - Special projects for a lab working on trustworthy AI for India and South Asia. Research Fellow MARS (Mentorship for Alignment Research Students), Cambridge AI Safety Hub July 2026 - Present - Build model organisms of misalignment: open LLMs fine-tuned to carry hidden objectives (secret loyalties, inductive backdoors, emergent misalignment) inserted at pretraining, SFT and RL-style post-training. - Lead blue-team evaluation using linear probing, activation steering and SAE feature analysis to tell behavioural masking apart from mechanistic erasure. Google Summer of Code Fellow HumanAI May 2026 - August 2026 - Implemented seven pathfinding algorithms over elevation data for the El Bagawat necropolis, Egypt, including A* with Tobler's hiking function and maximum-entropy inverse reinforcement learning. - Built a 3D voxel viewshed pipeline for line-of-sight analysis. Technical AI Safety Fellow AI Safety Initiative, Georgia Institute of Technology January 2026 - May 2026 - Studied scalable oversight, machine unlearning (WMDP, RMU) and mechanistic interpretability with Gemma Scope and interPLM SAEs. - Red-teamed frontier LLMs and extended circuit analysis to protein language models for biosecurity screening. AI and Web3 Intern Lokachakra June 2025 - October 2025 - Built a privacy-preserving KYC pipeline using vector embeddings and cosine similarity search. - Cut vector storage cost by 30% through dimensionality reduction. PUBLICATIONS (* = equal contribution) Known by Construction: Calibrating Model-State Instruments Against Models Whose State Was Installed, Hidden, or Never Present Manan Wadhwa*, Shivam Dubey* New in ML Workshop, NeurIPS 2026 - Audited seven model-state instruments against LoRA fine-tuned Qwen3 model organisms; all tracked the prompt or training history rather than the installed state. Dead Salmons Have a False-Positive Rate: A Calibrated Null-Model Test for Interpretability Manan Wadhwa*, Shivam Dubey* Interpretability as a Science Workshop, NeurIPS 2026 - Showed linear probes, concept directions and SAE feature selection flag up to 100% of findings on untrained models; proposed a calibrated null-model test. Steered Affect Moves a Partner Agent's Activations, and the Logistic Probe That Reads It Returns a Tie-Break, Not an Estimate Manan Wadhwa*, Shivam Dubey*, Debojeet Ghosh, Shreyansh Kumar, Ayushman Garg Interpreting Agent Behavior Workshop, NeurIPS 2026 - Measured emotional state transfer between LLM agents; logistic probe directions reproduce at 0.29 to 0.57 against 0.94 to 0.97 for difference-of-means. Toxin Circuits in ESM-2: Mechanistic Interpretability for Biosecurity Screening Manan Wadhwa*, Shivam Dubey* GenBio Workshop, ICML 2026 - ESM-2 linear probes detect 93.9% of redesigned toxins that evade BLAST homology search; SAE circuit analysis explains why. Distance to the Boundary Is Not Steering Resistance Manan Wadhwa*, Shivam Dubey* Under review, UniReps - Across OLMo-3 training checkpoints the decision margin grows 9.5x while per-dose steering strength stays constant. Doubt Is a Decision: A Sparse Circuit Commits Language Models to "I Don't Know" Sheryl Mehta*, Shagun Chadha*, Kanika Singhal*, Manan Wadhwa, Shivam Dubey arXiv preprint, forthcoming - About 20 attention heads and 100 MLP neurons carry 80 to 89% of the hedging decision in Llama-3.1-8B and Qwen2.5-7B. AWARDS 6th of 1800+ entrants, Haas F1 Challenge (video anomaly detection), 2025 SKILLS Research methods: Mechanistic interpretability, sparse autoencoders (SAEs), circuit analysis, activation patching, attribution patching, linear probing, activation steering, directional ablation, model organisms of misalignment, red teaming, machine unlearning, scalable oversight, AI evaluation, pre-registration, null-model controls, bootstrap inference Machine learning: LLM fine-tuning, LoRA, SFT, DPO, GRPO, reinforcement learning, inverse reinforcement learning, graph neural networks, computer vision, vision transformers, anomaly detection Programming languages: Python, C, C++, SQL, LaTeX Libraries and frameworks: PyTorch, Hugging Face Transformers, TransformerLens, Unsloth, PyTorch Geometric, scikit-learn, NumPy, CatBoost, OpenCV Models: Llama, Qwen, Gemma, OLMo, ESM-2, DINOv2, SigLIP, CLIP, SAM 2 Tools and infrastructure: Git, Docker, Linux, CUDA, ROCm, FastAPI, Weights and Biases, multi-GPU training