top of page

pexels-timson-foox-2182863 (1).jpg

CMU_Logo_Carnegie_Mellon-University_edit

Super Intelligence Research

An AI research group focusing on cutting edge technologies for intelligent AI systems.

STELAR

STELAR-Vision is a topology-aware training framework for vision-language models that enhances reasoning by moving beyond traditional chain-of-thought to include tree and graph-based structures via its TopoAug synthetic data pipeline. Using supervised fine-tuning, reinforcement learning, and a Frugal Learning strategy, it boosts both accuracy and output efficiency. It significantly outperforms larger and state-of-the-art VLMs across in-distribution and out-of-distribution benchmarks, demonstrating strong generalization.

Screenshot from 2025-08-12 17-08-07.png

SOLAR

We present SOLAR (Scalable Optimization of Large-scale Architecture for Reasoning), a framework that dynamically selects and optimizes Chain-of-Thought, Tree-of-Thought, and Graph-of-Thought reasoning topologies to improve both accuracy and efficiency. Powered by Topological-Annotation-Generation (TAG) for automated dataset preparation and Topological Scaling for adaptive curriculum learning, SOLAR achieves up to +10.02% accuracy gains on MATH and GSM8K while reducing response length and inference latency. A multi-task Topological Reward Model (M-TRM) further streamlines decision-making by jointly selecting the optimal reasoning topology and final answer in a single pass, setting a new benchmark for scalable, high-precision LLM reasoning.

Screenshot from 2025-08-12 16-49-57.png

More coming soon

...

Screenshot from 2025-08-12 17-08-07.png

bottom of page