AI safety researcher working on measurement validity. When a method claims to read a model's internal state (probes, steering directions, self-report, SAE features), I build a model where the answer is known in advance and check whether the method finds it.
Co-founder of Redarc Labs · research fellow at MARS (Cambridge AI Safety Hub) · special projects fellow at Secure AI Futures Lab · GSoC 2026 fellow with HumanAI · B.Tech CS, MAIT (2027)
Website · CV · LinkedIn · X · wadhwamanan56@gmail.com
- Model organisms of misalignment (MARS): open LLMs fine-tuned to carry hidden objectives such as secret loyalties, inductive backdoors and emergent misalignment, inserted at pretraining, SFT or RL-style post-training to see which stage gives the most persistent behaviour.
- Blue-team evaluation for the same organisms: probing, steering and SAE analysis to tell a behaviour that is masked from one that is erased, and tests of whether "removed" behaviours come back under mild retraining, distillation or prompt changes.
| Paper | Venue | Code |
|---|---|---|
| Known by Construction: Calibrating Model-State Instruments | NeurIPS 2026 · New in ML | digital_minds |
| Dead Salmons Have a False-Positive Rate | NeurIPS 2026 · Interpretability as a Science | dead-salmon-fpr |
| Steered Affect Moves a Partner Agent's Activations | NeurIPS 2026 · Interpreting Agent Behavior | affective-coupling-llm-agents |
| Toxin Circuits in ESM-2: Mechanistic Interpretability for Biosecurity Screening | ICML 2026 · GenBio (paper) | esm2 |
| Distance to the Boundary Is Not Steering Resistance | Under review · UniReps | |
| Doubt Is a Decision: A Sparse Circuit Commits Language Models to "I Don't Know" | arXiv preprint, coming soon |
All papers are equal-contribution with Shivam Dubey except the last. Summaries are on the website.
- elbagawat_path_finding: seven routing algorithms over DEM data for the El Bagawat necropolis in Egypt, including A* with Tobler's hiking function and MaxEnt inverse RL, plus a 3D voxel viewshed (GSoC 2026).
- gen_resist: a GNN that models bacterial genomes as k-mer graphs with CARD annotations to predict multi-label antibiotic resistance.
- FrameShift: a benchmark of 12 video anomaly detectors with DINOv2 + PatchCore and SAM 2 localization. 6th of 1800+ in the Haas F1 Challenge.
- csiro: biomass estimation from frozen DINOv2 and SigLIP features with 5-fold CatBoost regressors.
From 2021 to 2023 I ran v.dev, a programming club where I taught 30+ workshops on Python and the maths of machine learning. If you work on interpretability, model organisms or evaluation methodology, I would like to hear from you.


