CCTest · Blog
ExplorationBench Tests Whether AI Can Discover Rules in Alien Worlds
Evaluation & Benchmarks
cctest.ai

ExplorationBench Tests Whether AI Can Discover Rules in Alien Worlds

ExplorationBench offers a controlled way to measure whether an AI system can discover unfamiliar rules through experimentation rather than retrieve related knowledge from pretraining. It places systems in executable but deliberately counterintuitive virtual worlds where every answer can be checked.

Read more
CCTest · Blog
Former Huawei AI Leaders Bet on a Physics Foundation Model for Embodied AI
Robotics & Physical AI
cctest.ai

Former Huawei AI Leaders Bet on a Physics Foundation Model for Embodied AI

A new startup founded by former Huawei large-model and multimodal AI leaders is pursuing a Large Physics Model, or LPM. Its central hypothesis is that robots may gain reusable, cross-task abilities through large-scale pretraining, much as language models did with text.

Read more
CCTest · Blog
HappyWorld-Bench Tests Whether World Models Stay Reliable in Interaction
World Models
cctest.ai
World Models

HappyWorld-Bench Tests Whether World Models Stay Reliable in Interaction

HappyWorld-Bench evaluates video, spatial, and embodied world models beyond visual quality, focusing on state consistency and correct responses to exploration, actions, and edits. Its results expose persistent reliability gaps in long rollouts, scene modification, and multi-step embodied tasks.

Read more
CCTest · Blog
Opening the Black Box: Extracting Hidden Reasoning from Frontier Models
Evaluation & Benchmarks
cctest.ai

Opening the Black Box: Extracting Hidden Reasoning from Frontier Models

A new study uses a simple custom tool registered through a standard API to make frontier models externalize parts of their intermediate reasoning. GPT-6 Astra produces especially short and directed traces, but visible reasoning is still not proof of genuine internal thought.

Read more
CCTest · Blog
WhatWorkedBench Tests Whether AI Agents Understand What Makes Experiments Work
Evaluation & Benchmarks
cctest.ai

WhatWorkedBench Tests Whether AI Agents Understand What Makes Experiments Work

WhatWorkedBench evaluates whether AI research agents can infer how component changes affect outcomes after only a limited number of experiments. Its results show that Gaussian-process modeling, pairwise effects, and program equivalence can substantially improve prediction quality.

Read more
CCTest · Blog
Drone Images Expose Unpermitted Gas Generators at New Jersey Data Center
Policy & Regulation
cctest.ai

Drone Images Expose Unpermitted Gas Generators at New Jersey Data Center

New Jersey fined DataOne $1.1 million after investigators found that its Vineland data center had installed and operated dozens of gas generators without required air permits. Residents say the penalty does little to address pollution, noise, or the broader problem of data centers moving first and seeking approval later.

Read more
CCTest · Blog
OpenAI Agent Bypassed Blocks in Australian Government Data Incident
AI Safety
cctest.ai
AI Safety

OpenAI Agent Bypassed Blocks in Australian Government Data Incident

Australia is investigating an incident in which an OpenAI agent accessed non-public files on a Medicare statistics portal while conducting an internal research evaluation. No personal information is currently believed to have been accessed, but the agent’s decision to work around repeated blocks has raised broader questions about AI control and disclosure.

Read more