Research
Papers & experiments
Work published by us submitted to various conferences and workshops.
2026
Under review
Venue TBD
Arjun Srivastava · Gaurav Karakoti
A two-stage neural surrogate for high-throughput materials discovery:
a JEPA encoder learns physics-informed latent representations of
crystal graph structure via masked embedding prediction, and a
lightweight transformer readout predicts formation energy, bandgap,
and stability — without any quantum-mechanical labels and without
performing DFT calculations at inference time.
Workshop submission
TBD
Arjun Srivastava
DCTMS is a compile-time framework that analyses a transformer's
computational graph before inference and derives a complete tensor
lifetime map, physical buffer assignment, and concurrency schedule.
On Apple M4 unified memory, it reduces 448 logical tensors to 122
physical buffers (72.77%) and establishes a provable memory envelope
before inference begins, cutting latency variance by 55–69% at the
7B scale.
Workshop submission
ECCV 2026 · Wearable AI
Arjun Srivastava · Gaurav Karakoti
We investigate whether the learned latent representations of a
visual world model can serve as a compact memory substrate for
long-context egocentric video reasoning. A V-JEPA encoder produces
persistent latent state; a small Qwen vision-language model reasons
over that state directly — without external retrieval.
