- Youngest Google grantee at 17. I used the GCP grant to build Gakudan, a multi-agent protocol for swarm reasoning
- Hedge fund data processing: built options order-book pipelines over PBs of data. Sped up the old system ~8x. Focused on reverse engineering the Jane Street options activity in India
- My Neovim config was the 2nd biggest Neovim config until a month after release.
Previously:
Papers
- Legible Failures: Detecting and Repairing In-Context Binding Errors arXiv
- In-Context Binding Capacity in Language Models [LP4FM NeurIPS '26] arXiv
- Quadratic bounds for uncompletable words and matrix mortality arXiv
- Decodable In-Context State and Model Output Across Training arXiv
- A Heckler in the Hidden State: Correctness Signals in Diffusion Language Models arXiv
Interests
Formal reasoning for biological discovery; multiobjective RL; pre-pre-training; physics intuition for language models; working memory and verifiers for reasoning; making models evolve to learn rules from their environments; pre-2007 iPods; Chet Baker; Cornershop; The Strokes; HMT Watches; Fountain Pens.
Books
The Little Prince; The Myth of Sisyphus; Shoe Dog; Jobs; Rooney: My Decade in the Premier League; Klara and the Sun; Complete Nonsense.
email: [email protected]
twitter: @samrathchadha
linkedin: in/samrathchadha
I am always happy to talk about anything and everything interesting. I am especially interested in new approaches to pretraining and attempts at speculatively formalizing biology!