Generating test cases is essential for language models but remains difficult. Existing test‑time scaling approaches either isolate candidates or aggregate execution traces, which do not scale to long‑horizon tasks. We introduce TEX, a collaborative test‑time scaling framework built on execution‑driven cross‑validation. TEX runs multiple agents in parallel to produce tests and patches, then cross‑executes every test against every patch, constructing a $K \times K$ feedback matrix. Each agent consumes only its own matrix slice (a row or column) for refinement, enabling objective validation without sharing verbose traces. Experiments on SWT‑Bench Verified show TEX achieves 86.4% pass rate, setting a new state‑of‑the‑art and improving by 9.0% over single‑agent baselines. Even random selection from TEX outputs reaches 84.4%, surpassing the previous best of 84.0%, demonstrating that collaborative execution feedback improves overall solution quality, not just the top‑ranked samples. Our findings highlight execution‑grounded feedback as an effective scaling mechanism for long‑horizon agentic tasks.
Review