CCTest · Blog
CLIFT Turns Web Agents’ Self-Checks into Reusable Training and Test-Time Signals
AI Agents
cctest.ai
AI Agents

CLIFT Turns Web Agents’ Self-Checks into Reusable Training and Test-Time Signals

CLIFT combines self-verification, conformal calibration, and trajectory selection to address sparse rewards and the cost of step-level external judging. Its certified verifier bank is used both during training and for selecting among test-time rollouts without an external judge.

Read more
CCTest · Blog
PAIR Builds a Bridge from Perception to Action in VLA Models
Robotics & Physical AI
cctest.ai

PAIR Builds a Bridge from Perception to Action in VLA Models

PAIR introduces an explicit intermediate representation for continuous-action vision-language-action models, making the transition from visual-language understanding to robot control more structured. The method reports improvements on LIBERO, LIBERO-Plus, CALVIN, and real-world manipulation tasks.

Read more
CCTest · Blog
CLASHarness Turns Multi-Agent Conflict Evaluation into a Verifiable Task
Evaluation & Benchmarks
cctest.ai

CLASHarness Turns Multi-Agent Conflict Evaluation into a Verifiable Task

CLASHarness is a controllable benchmark for testing how vision-language agents handle stale reports, contradictory instructions, and shared-resource conflicts during computer use. Its generation pipeline pre-commits visual facts and derives graders from executed reference trajectories.

Read more