Research Engineer
Atul Saxena
I build the environments models are trained in, and the tools that show what they learned.
- RL Systems
- Interpretability
- ML Infrastructure
Research Engineer
Atul Saxena
I build the environments models are trained in, and the tools that show what they learned.
Three preprints, all with DOIs and all with the code or weights released alongside. ORCID 0009-0000-3588-9848 (opens in a new tab).
Atul Saxena, Divyanshi Sharma
Most systems ask a language model to make a decision and then parse whatever sentence comes back. Noma skips the sentence. It answers with numbers you can threshold on, in about 16 milliseconds, and it can say "none of these" instead of guessing.
Atul Saxena, Pushp Kharat
Instead of a model producing an answer you have to trust, LEMMA produces the derivation: every rewrite is a named rule, every intermediate state is checked, and the verification status is explicit — checked, heuristic, unverified, or unsupported are four different outcomes, not one.
Pushp Kharat, Atul Saxena
The intuitive story is that attention heads rewire themselves depending on context. We tested it and it is not what happens. Heads have fixed behaviour; the flexibility comes from which combination of them gets recruited, communicating through the shared residual stream.