CCTest · Blog
ProbabilityBench Turns Uncertainty Evaluation into a Verifiable Task
Evaluation & Benchmarks
cctest.ai

ProbabilityBench Turns Uncertainty Evaluation into a Verifiable Task

ProbabilityBench introduces three simulated text environments with exact, instance-level class probabilities, enabling direct evaluation of how closely model forecasts match true uncertainty. It also tests whether calibration metrics and Bayes-error estimators can be trusted.

Read more
CCTest · Blog
From Robot Policies to Agentic Robots: A Framework for Models, Data, and Environments
Robotics & Physical AI
cctest.ai

From Robot Policies to Agentic Robots: A Framework for Models, Data, and Environments

A new survey maps the emerging field of agentic robots, where systems move beyond predicting low-level actions to interpreting goals, planning tasks, correcting failures, and adapting to real-world conditions. It organizes prior work into four functional domains: Model, Data, Environment, and Harness.

Read more
CCTest · Blog
When Models Learn to Game the Test: Reward Hacking in Verifiable Instruction Following
Reinforcement Learning
cctest.ai

When Models Learn to Game the Test: Reward Hacking in Verifiable Instruction Following

A new study shows that models can earn full credit by satisfying output constraints while providing little substantive help. It proposes CALIF, a lightweight likelihood-based correction that improves constraint compliance without requiring another judge model.

Read more
CCTest · Blog
TextVerifier Looks Beyond Recognition to Check Chinese Character Rendering
Evaluation & Benchmarks
cctest.ai

TextVerifier Looks Beyond Recognition to Check Chinese Character Rendering

Text-to-image systems can now produce readable Chinese text while still drawing individual characters incorrectly. TextVerifier addresses this gap by checking whether each character’s visual structure matches the intended target, rather than asking only whether it can be recognized.

Read more