Research

N. Meters, A. Chen, R. Kapoor

We present a sparse mixture-of-experts architecture that achieves sub-quadratic inference cost on sequences exceeding 128k tokens. By introducing a locality-sensitive routing mechanism that exploits the low-rank structure of attention patterns, our method reduces peak memory by 3.8x while maintaining 98.2% of dense model quality across standard long-context benchmarks. We provide theoretical guarantees on routing stability and demonstrate wall-clock speedups on commodity hardware.

N. Meters, J. Lindqvist

We develop a fully differentiable spectral Navier-Stokes solver that enables end-to-end training of neural surrogate models for turbulent flows. Our approach embeds physical conservation laws directly into the computational graph, allowing gradient-based optimization to respect divergence-free constraints without projection steps. On the Kolmogorov flow benchmark, the resulting surrogates achieve 12x speedup over classical solvers at Reynolds numbers up to 10,000 with bounded error accumulation over 500 rollout steps.

N. Meters, P. Okonkwo, M. Tanaka

We propose a zero-copy distributed key-value cache architecture for serving large language models across disaggregated GPU clusters. By leveraging RDMA-based memory transfers and a novel page-table abstraction for attention state, our system eliminates serialization overhead during prefill-decode handoffs. Evaluations on a 64-GPU cluster show 2.1x improvement in time-to-first-token and 40% higher throughput compared to state-of-the-art serving frameworks under production trace workloads.

N. Meters, S. Hoffmann

We introduce a pipeline that extracts persistent homology descriptors from 3D point clouds and integrates them as auxiliary features into standard classification architectures. The topological features capture global shape properties invariant to noise, occlusion, and sampling density. On ModelNet40 and ScanObjectNN, augmenting PointNet++ with our descriptors improves robustness to 60% point dropout by 8.3 percentage points with negligible computational overhead.

N. Meters, L. Vasquez, A. Chen

We present an adaptive algorithm that dynamically selects per-layer numerical precision during training by estimating the signal-to-noise ratio of gradient updates. Layers with high gradient noise tolerance are cast to FP8, while sensitive layers retain BF16, with transitions governed by an exponential moving average of gradient variance. Applied to GPT-scale models, the method reduces training FLOPS by 28% with no measurable degradation in validation loss.

N. Meters, D. Petrov, K. Yamamoto

We design a message-passing graph neural network decoder for the rotated surface code that operates in O(n) time per syndrome measurement round. The decoder is trained on synthetic noise models and generalizes to hardware-calibrated noise without fine-tuning. At code distances 5 through 21, it achieves logical error rates within a factor of 1.3 of minimum-weight perfect matching while running 50x faster, enabling real-time decoding at the repetition rates required by near-term superconducting quantum processors.