NVIDIA AI Introduces PivotRL: A New AI Framework Achieving High Agentic Accuracy With 4x Fewer Rollout Turns Efficiently

2026-03-25 · Source: Machine Learning ML & Generative AI News · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Intermediate, quick

Summary

NVIDIA AI has introduced PivotRL, a new AI framework designed to enhance the training of long-horizon agents for tasks like coding, terminal use, and web search. This framework aims to combine the speed of Supervised Fine-Tuning (SFT) with the generalization capabilities of End-to-End Reinforcement Learning (E2E RL), without incurring the high costs typically associated with E2E RL. PivotRL achieves this by utilizing existing SFT trajectories and incorporating two key mechanisms: Pivot Filtering, which focuses on critical intermediate turns with high outcome variance, and Functional Rewards, which employ domain-specific verifiers instead of rigid string matching. The framework demonstrates a +4.17% higher in-domain accuracy and a +10.04% higher out-of-domain accuracy compared to SFT. Notably, on the SWE-Bench benchmark, PivotRL matched E2E RL accuracy with 4x fewer rollout turns and approximately 5.5x faster wall-clock time, and it powers NVIDIA's Nemotron-3-Super-120B-A12B.

Key takeaway

For research scientists developing long-horizon agents, PivotRL offers a compelling alternative to traditional E2E RL, providing comparable accuracy with substantially reduced computational overhead. You should consider integrating PivotRL's pivot filtering and functional reward mechanisms to achieve robust agent performance across diverse tasks while optimizing training efficiency. This approach can accelerate development cycles and enable more complex agent deployments.

Key insights

PivotRL enhances agentic accuracy and efficiency by selectively applying RL to SFT trajectories.

Principles

• Focus RL on high-variance "pivot" turns.
• Reward locally acceptable actions, not just exact matches.

Method

PivotRL operates on existing SFT trajectories, using Pivot Filtering to target critical intermediate turns and Functional Rewards with domain-specific verifiers to evaluate actions.

In practice

• Improve long-horizon agent performance.
• Reduce RL training costs significantly.

Topics

• Reinforcement Learning
• Agentic AI
• Supervised Fine-Tuning
• NVIDIA AI
• SWE-Bench

Best for: Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• Streaming Tokens and Tools: Multi-Turn Agentic Harness Support in NVIDIA Dynamo — Optimizing agentic inference requires precise prompt stability, correct reasoning/tool parsing, and efficient streaming.
• Unsloth AI and NVIDIA are Revolutionizing Local LLM Fine-Tuning: From RTX Desktops to DGX Spark — Unsloth AI and NVIDIA enable rapid, local LLM fine-tuning on diverse NVIDIA GPUs for specialized AI applications.
• Nvidia corners the AI agent stack — Nvidia's strategy prioritizes AI agents as future compute consumers, integrating them across its entire tech stack.
• NVIDIA Releases Nemotron-Cascade 2: An Open 30B MoE with 3B Active Parameters, Delivering Better Reasoning and Strong Agentic Capabilities — NVIDIA's Nemotron-Cascade 2 MoE model excels in reasoning and agentic tasks via Cascade RL and MOPD.
• Nvidia launches open-source toolkit for enterprise AI agents — NVIDIA's Agent Toolkit provides open-source components for building secure, efficient enterprise AI agents.
• MiniMax M2.7 Advances Scalable Agentic Workflows on NVIDIA Platforms for Complex AI Applications — MiniMax M2.7, a 230B-parameter MoE model, offers high efficiency and performance for agentic tasks via NVIDIA optimizations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning ML & Generative AI News.