Nvidia harness research links agent design to higher scores and lower token use

Nvidia research argues that an AI harness—the tools, runtime and architecture managing a model’s context, memory and actions—can influence performance more than the underlying model. Its Agentic Variation Operators harness helped Claude Opus 5 score 100% on ARC-AGI-3, an interactive benchmark of 2D games without instructions, compared with 30% without the harness. Nvidia’s newer open-source NVIDIA Labs Object-Oriented Agents (NOOA) framework scored 82.2% on SWE-bench Verified, 86.8% on CyberGym L1 tests, 50.2% on ARC-AGI-3 with GPT-5.5 and 85.1% with GPT-5.6-sol, at less than $20 per game. Nvidia said NOOA’s pass-by-reference memory management and removal of context-compaction processes effectively halved token use, potentially lowering inference costs while improving output quality.

The information on this website is generated using AI and we cannot guarantee its accuracy. Please use it as reference information only.