Efficient foundation models
Reducing the computational and memory cost of training and serving LLMs through low-precision learning, model compression, and efficient inference.
Open research collective · Established 2025
Neural
Superintelligence Lab
Previously Open Neural Network Research Lab 01.2025 to 12.2025
Recent progress in AI has been driven largely by scaling, but this comes with significant computational and memory costs.
Our goal is to make progress toward AGI by developing neural systems that are lighter, faster, more affordable, and more capable in learning and reasoning.
- Director
- Vincent-Daniel Yun, USC
- Mission
- Fast Lightweight AGI
- Support
- MODULABS · Brian Impact Foundation
- Research model
- Independent · non-commercial · cross-institutional
Focus areas
Toward lighter, faster,
and more capable AGI
Modern neural networks are often overparameterized and computationally redundant. We study how to transform them into structurally and computationally efficient systems without giving up learning and reasoning performance.
Optimization & generalization
Understanding gradient dynamics and developing training methods that improve stability, robustness, and generalization.
Architectures & agentic intelligence
Developing models and multi-agent systems that improve reasoning, coordination, and learning efficiency.
Law & responsible deployment
Examining privacy, accountability, and legal questions arising from real-world deep learning applications.
How we work
More capability with less computation, memory, and training cost.
Optimization
We study training dynamics and methods that improve convergence, stability, and generalization.
Model efficiency
We use pruning, quantization, compression, and efficient architectures to reduce model cost.
AGI systems
We investigate how efficient models and agents can learn, reason, and coordinate across tasks.
Lab notes
Recent news
Selected publications, workshop acceptances, and research support.
9 updates · 2025 to 2026
2026Latest
Accepted · EMNLP 2026 Main Conference
Locality-Aware Redundancy Pruning for LLM Depth Compression
Accepted to the Main Conference at EMNLP 2026.
Summer2026
Invited lecture · Samsung Electronics
Multi-Agent Systems and Loop Engineering
Lab Director Vincent-Daniel Yun delivered a summer course as a lecturer at Samsung Electronics headquarters.
2026ICML Workshop
Accepted · ICML 2026 AdapFM
Rethinking Layer Redundancy in Large Language Models: Calibration Objectives and Search for Depth Pruning
Accepted to the ICML 2026 Workshop on Adaptive Foundation Models.
2026ICASSP
Accepted · ICASSP 2026
Sharpness-Aware Minimization with Z-Score Gradient Filtering
Accepted to the 2026 IEEE International Conference on Acoustics, Speech and Signal Processing.
November2025
Research support
Four lab projects receive $6,000 in research support
The Brian Impact Foundation and MODULABS supported four research projects from the lab,
providing a combined grant of $6,000.
September2025
September2025
September2025
September2025
Research principles
Independent research
for the public good
These principles guide how we select research questions, collaborate, and share our work.
Independent research
Our research direction is not shaped by sales targets, product roadmaps, or commercial obligations.
Cross-institutional collaboration
Researchers with different affiliations and backgrounds work together on shared research questions.
Research for public good
We pursue efficient AGI research that can make advanced intelligence more accessible and useful.
No conflicts of interest
We do not engage in any commercial activities, nor do we maintain any financial, contractual, or other conflicts of interest with any company. This lab serves as a collaborative research space where researchers, united by a shared mission, pursue artificial intelligence research aimed at advancing the public good and contributing to a better world.