Skip to content
RobotWorld
← Back to Videos

SoL-Pi: A Token-Efficient Agent Harness Built by Auto-Research Loops

Loading video
Loading video

NVIDIA fans out 152 auto-research proposals and keeps the four mechanisms that survive paired, held-out validation. The assembled harness retains roughly 94 percent of Pi average score while using 45-49 percent fewer tokens and costing about one third less.

Agent HarnessRecursive self-improvementToken EfficiencyCoding AgentNVIDIA
Category: research
Author: @QihangCao11
Date: 2026-09-11T00:00:00
Duration: 60.0s
Source: https://x.com/QihangCao11/status/2098343355067630020
Reference: https://nvlabs.github.io/SoL-Pi/

Related Videos

Walden Robotics raises 300M at 1.1B valuation
0:15

Walden Robotics raises 300M at 1.1B valuation

@LeoKharon · 421 views · 2026-07-26
Walden RoboticsFunding
AV3D-Harness Adds Sound to 3D Real-to-Sim Scenes
1:12

AV3D-Harness Adds Sound to 3D Real-to-Sim Scenes

@lasiafly · 45 views · 2026-10-27
AV3D3D Audio
RoboRender: Turning Simulated Trajectories into Photorealistic Video for Policy Training
2:02

RoboRender: Turning Simulated Trajectories into Photorealistic Video for Policy Training

@RavenHuang4 · 645 views · 2026-10-08
Sim-to-RealVideo Generation
Long-WAM: Scaling World-Action Model Context to 19.2s in Real Time
2:38

Long-WAM: Scaling World-Action Model Context to 19.2s in Real Time

@AaronWeiHuang · 472 views · 2026-10-08
World-action modelsVLA
DyaPlex: Full-Duplex Human-Robot Interaction at NeurIPS 2026
0:42

DyaPlex: Full-Duplex Human-Robot Interaction at NeurIPS 2026

@luminohope · 429 views · 2026-10-07
Human-robot interactionNeurIPS
OpenMausBot: a self-hostable team of AI agents
1:37

OpenMausBot: a self-hostable team of AI agents

@BuildwithOmkarr · 788 views · 2026-09-23
OpenMausBotAI Agent

As an Amazon Associate, we earn from qualifying purchases.