CCTest · Blog
OpenGCUA Opens the Full Stack for Generalist Computer-Use Agents
AI Agents
cctest.ai
AI Agents

OpenGCUA Opens the Full Stack for Generalist Computer-Use Agents

Alibaba’s MAI-UI team has released the first version of OpenGCUA, an open project for agents that can decide how to observe, act, and organize computer tasks across visual and programmatic interfaces. The release includes the harness, prompts, evaluation trajectories, a real-device tool, and a cross-device SDK.

Read more
CCTest · Blog
VLA-Δ Teaches Robots How Physical States Change, Not Just Which Actions to Copy
Robotics & Physical AI
cctest.ai

VLA-Δ Teaches Robots How Physical States Change, Not Just Which Actions to Copy

VLA-Δ adds explicit physical state-transition supervision to vision-language-action models, helping them model changes in objects, grippers, and contacts during manipulation. The policy acts directly from current observations at deployment without an additional decoding stage.

Read more
CCTest · Blog
SpecAgent Lets Small Models Draft Agent Trajectories for Large-Model Review
AI Agents
cctest.ai
AI Agents

SpecAgent Lets Small Models Draft Agent Trajectories for Large-Model Review

SpecAgent introduces trajectory-level speculation for tool-using agents: a small model grounds read-only actions in live tool observations, while a large model verifies the draft in one pass. The design reduces large-model decoding without requiring every accepted step to match the large model’s own proposal.

Read more
CCTest · Blog
WorldMemBench Tests Whether Video World Models Remember Their Worlds
Evaluation & Benchmarks
cctest.ai

WorldMemBench Tests Whether Video World Models Remember Their Worlds

WorldMemBench introduces a systematic benchmark for measuring memory in video world models, separating static content retention from dynamic state evolution. Results across 260 cases and nine representative models show that visual consistency alone is not enough for persistent worlds.

Read more