NOOA: Six Core Capabilities to Boost AI Performance
NVIDIA Labs' open-source NOOA framework enhances AI model performance.
1What is the NOOA Framework?
Developed by NVIDIA Labs, NOOA (NVIDIA Labs Object-Oriented Agents) is an open-source framework designed to enhance the performance of AI models. NOOA offers a Python-based structure and operates through a single Python class. The methods of this class represent capabilities; fields represent state; docstrings represent prompts; and type annotations represent enforced contracts. This structure allows for traditional software development processes, making the code easily reviewable and improvable by both humans and AI.
2Memory System
NOOA provides an advanced subsystem for long-term memory management. Instead of automatic background summarization pipelines, this system allows the agent to write, query, and correct records as part of its work. Memory records carry types, importance levels, and tags. These records are interconnected into a knowledge graph rather than a flat log, enabling more effective management and utilization of information.
3Performance and Efficiency
NOOA delivers measurable performance improvements across various domains. For instance, it achieved an accuracy rate of 82.2% with GPT-5.5 on the SWE-bench Verified test and 86.8% on the CyberGym L1 test. These results demonstrate NOOA's significant advantages in both accuracy and efficiency. Notably, it achieved this performance on the SWE-bench Verified test with GPT-5.5 using only 29 LLM calls and approximately 1.1 million tokens.
4Software Engineering
NOOA has also achieved impressive results in the field of software engineering. On the SWE-bench Verified test, it reached an accuracy rate of 82.2% with GPT-5.5, surpassing the best result on the leaderboard. This success was achieved by a general-purpose agent without using benchmark-specific prompts, showcasing NOOA's adaptability and strength across different domains.
5Cybersecurity
NOOA has also made significant strides in cybersecurity. On the CyberGym L1 test, it achieved a success rate of 86.8% with GPT-5.5, leading among open-source agents. This result was obtained with network access blocked and a rule-based 'cheat check' over each action, demonstrating NOOA's ability to reason over code and its reliability.
6General Reasoning
NOOA has also performed well in general reasoning tests. On the ARC-AGI-3 test, it achieved an average RHAE score of 50.2% with GPT-5.5. This score is 8.5 points higher than a hypothesis-driven baseline and 11.8 points better than file-based notes. These results highlight NOOA's general reasoning capabilities and its success in memory management.
7Efficiency and Cost
NOOA offers advantages not only in accuracy but also in efficiency and cost. With GPT-5.5, it achieved an accuracy rate of 82.2% on the SWE-bench Verified test using only 29 LLM calls and approximately 1.1 million tokens. This demonstrates NOOA's effectiveness in both performance and cost. Notably, it operated at a cost of under $20 per game on the ARC-AGI-3 test.
8Community Engagement
As an open-source framework, NOOA is open to community adoption and development. NVIDIA Labs has made NOOA's code, data, and evaluations publicly available, encouraging the community to reproduce, challenge, and build upon these results. This approach ensures NOOA's continuous evolution and widespread adoption.
9Conclusion and Importance
NOOA is an innovative framework designed to enhance the performance of AI models. Its Python-based structure, advanced memory system, and successes across various domains indicate that NOOA will play a significant role in AI development processes. Being open-source and open to community engagement ensures NOOA's continuous evolution and widespread adoption.
Original article: Nvidia Developer Blog — NOOA: Yapay Zeka Performansını Artıran Altı Temel Yetenek