CCTest · Blog
Keeping Human Instructions in Force: SAI Adds Persistent Authority to LLM Agents
AI Agents
cctest.ai
AI Agents

Keeping Human Instructions in Force: SAI Adds Persistent Authority to LLM Agents

A new paper proposes Structured Authority Injection (SAI), a protocol that uses typed MCP calls and ranked authority context to keep operator corrections active across planning cycles. Its evaluation reports less ambiguity and authority decay, although the gains vary considerably by model.

Read more
CCTest · Blog
OmniCompass Brings Spatiotemporal and Audio-Visual Rigor to Video Evaluation
Evaluation & Benchmarks
cctest.ai

OmniCompass Brings Spatiotemporal and Audio-Visual Rigor to Video Evaluation

OmniCompass is a holistic benchmark for video grounding and understanding across temporal grounding, spatial grounding, and question answering. By pairing text-only with text-and-image target references, it tests whether models can use video evidence consistently rather than merely produce plausible answers.

Read more