TEX runs K software agents in parallel over multiple refinement rounds.
Cross-executing tests and patches improves software testing agents
TEX lets parallel agents refine their work using a matrix of execution outcomes rather than exchanging lengthy reasoning traces.
Big Tech
Akash Gokul · Srijan Bansal · Yingbo Zhou · Semih Yavuz
Recursive · Salesforce Research · NVIDIA
Research Digest··3 min read
Gokul and colleagues introduce TEX, a test-time scaling framework in which multiple agents independently generate tests and code patches, then run every test against every patch.
§