CCTest · Blog
Anthropic Files for IPO as Growth, Losses and AI Governance Collide
Business & Funding
cctest.ai
Business & Funding

Anthropic Files for IPO as Growth, Losses and AI Governance Collide

Anthropic has formally filed to go public under the ticker “ANT,” revealing explosive revenue growth, a large accounting-driven net loss, and an unusual structure that keeps voting control with its founders. The company has not yet disclosed its target raise or proposed share price.

Read more
CCTest · Blog
UserArena Tests Shopping Agents Against Long-Horizon User Behavior
Evaluation & Benchmarks
cctest.ai

UserArena Tests Shopping Agents Against Long-Horizon User Behavior

UserArena connects real e-commerce sessions with an interactive environment and production A/B-test outcomes to evaluate whether shopping agents can model users over extended interactions. The paper also introduces OPTD to improve simulator fidelity on model-generated trajectories.

Read more
CCTest · Blog
Dual Exam Separates Multimodal Agents from World Models to Expose Their Real Limits
Evaluation & Benchmarks
cctest.ai

Dual Exam Separates Multimodal Agents from World Models to Expose Their Real Limits

Dual Exam introduces a two-track evaluation framework: multimodal agents are tested with environment responses held fixed, while world models are tested with action interventions held fixed. The results show that strong component scores do not yet translate into reliable agent–world-model collaboration.

Read more
CCTest · Blog
HiSTEVE 1.0 Brings Human-Like Keyboard-and-Mouse Control to Minecraft
AI Agents
cctest.ai
AI Agents

HiSTEVE 1.0 Brings Human-Like Keyboard-and-Mouse Control to Minecraft

HiSTEVE 1.0 combines long-horizon planning with native keyboard-and-mouse control, allowing a VLA agent to perform multi-step tasks in Minecraft from visual observations alone. Its design links capability-aware planning, policy-aware completion detection, and specialized training for different interaction modes.

Read more
CCTest · Blog
A Scaling Law for Synthetic Rewrites Links Data Mix to Compute
Large Language Models
cctest.ai

A Scaling Law for Synthetic Rewrites Links Data Mix to Compute

A new study models synthetic rewriting as a compute-and-data allocation problem rather than free data augmentation. Its experiments suggest that the optimal share of rewritten tokens grows with the training budget, and that a model trained on rewrites can outperform the model that produced them.

Read more