AI Agents Frontier Accelerates
Key Questions
What is the current SWE score on ProgramBench?
ProgramBench currently shows 0% SWE performance. This benchmark highlights ongoing challenges in software engineering tasks for AI agents.
What achievement has Zhipu GLM-5V accomplished?
Zhipu GLM-5V has achieved state-of-the-art (SOTA) status in VLA evaluations. It represents a leading multimodal model in vision-language-action capabilities.
What are Sakana Conductor and Fugu known for?
Sakana Conductor and Fugu have topped recent AI agent evaluations. They demonstrate advanced performance in multi-agent orchestration and benchmarks.
What is Meta's CWM contribution to AI agents?
Meta's CWM refers to advancements in world models for agents. It supports better simulation and planning in complex environments.
What recent release is DeepSeek-V4?
DeepSeek-V4 is a new model release accelerating AI agent capabilities. It contributes to the frontier of efficient and powerful agentic systems.
How does HERMES advance driving models?
HERMES is a unified self-driving world model integrating scene understanding and future prediction. It drives advancements in autonomous vehicle AI through [ICCV 2025] research.
What efficiency gains does ROSE infrastructure provide?
ROSE achieves 3x improvements in end-to-end throughput for agentic RL on serving GPUs. It uses cooperative elasticity across model sizes and cluster scales.
What is Apollo's approach to agent optimization?
Apollo from University of Cambridge co-optimizes agent policies with RL and environment configurations. This coordinated method enhances agent performance in dynamic settings.
ProgramBench 0% SWE; Zhipu GLM-5V SOTA VLA; Sakana Conductor/Fugu top evals; Meta CWM; DeepSeek-V4; HERMES driving models; Spot/UniT robots; Apollo co-opt; ROSE infra 3x; Skill1/Zenith orchestration; synthetic data/evals boom.