Publications
AI and Language Models
-
Synthetic Data RL: Task Definition Is All You Need
Reinforcement fine-tuning from a task definition alone: synthesize the data, adapt the difficulty, then train with RL. No labeled dataset required.
-
API Pack: A Massive Multi-Programming Language Dataset for API Call Generation
A million-plus instruction dataset for API call generation across ten languages; fine-tuning on it beats GPT-3.5 and GPT-4 on unseen APIs.
-
Diversity Measurement and Subset Selection for Instruction Tuning Datasets
Measure instruction-set diversity with determinantal point processes, then pick the subset that keeps it; small diverse subsets match the full set.
-
Octo-planner: On-device Language Model for Planner-Action Agents
Split planning from action so a small on-device model can decompose a request into steps; trained on synthetic plans with multi-LoRA.
-
Scaling Law Hypothesis for Multimodal Model
A single scaling law across text, image, audio, and video tokens, tied to the token count each modality actually needs.
-
More Compute Is What You Need
Argues that model quality tracks total training compute more than architecture or data recipe, with implications for how to spend a fixed budget.
-
JetMoE: Reaching Llama2 Performance with 0.1M Dollars
An open 8B mixture-of-experts model trained on 96 H100s that outperforms Llama2-13B, with the full recipe released.
-
AuthentiGPT: Detecting Machine-Generated Text via Black-Box Language Models Denoising
Detect machine-generated text by denoising it with a black-box LLM and measuring how much survives.
-
Dr. LLaMA: Improving Small Language Models on PubMedQA via Generative Data Augmentation
Generative data augmentation lifts small models on PubMedQA; the first LLaMA fine-tune on medical QA.
Optics, Imaging, and Applied Physics
-
Non-invasive Estimation of the Powder Size Distribution from a Single Speckle Image
Estimate a pharmaceutical powder's size distribution from one speckle image, in real time, for continuous manufacturing.
-
Ferromagnetism Emerged from Non-ferromagnetic Atomic Crystals
Ferromagnetism appears when two non-ferromagnetic 2D crystals are stacked, from interlayer charge transfer.
-
PLayer: A Plug-and-Play Embedded Neural System to Boost Neural Organoid 3D Reconstruction
A plug-in embedded neural system that speeds 3D reconstruction of neural organoids on lab hardware.
-
On the Use of Deep Learning for Three-Dimensional Computational Imaging
Survey of where learned priors help three-dimensional computational imaging and where they fail.
-
Noise-resilient Deep Tomographic Imaging
Reconstruct X-ray tomograms at low photon counts by learning the noise structure instead of averaging it away.
-
Physics-assisted Generative Adversarial Network for X-ray Tomography
A GAN with the X-ray forward model built in, so reconstructions stay physically consistent at low dose.
-
LION: Learning to Invert 3D Objects by Neural Networks
Neural inversion of 3D objects from limited-angle projections.
-
Randomized Probe Imaging Through Deep k-learning
Recover a sample from a single randomized probe measurement by unrolling gradient descent into a network.
-
Wafer-scale On-chip Synthesis and Field Emission Properties of Vertically Aligned Boron Nitride Based Nanofiber Arrays
Wafer-scale growth of vertically aligned boron nitride nanofibers and their field-emission behavior.
Patents
-
Pupil Engineering Method to Enhance the Signal of the Real-Time Determination of Particle Size Distribution in PowdersPupil design that boosts the speckle signal used for real-time particle sizing.
-
System and Method for Real-Time Determination of Particle Size Distributions in Dry PowdersOptical system and method for real-time particle size distribution in dry powders.
Projects and Workshops
-
-
Hardware Efficient Quantum Computing via Circuit DecompositionDecompose circuits to fit the target hardware's native gates and connectivity.