TEX: Test-Time Scaling Testing Agents via Execution-based Cross-Validation
Akash Gokul, Srijan Bansal, Yingbo Zhou and Semih Yavuz (Salesforce Research) present TEX, a test-time scaling method for test-generation agents in which parallel agents cross-execute every generated test against every generated patch.
Ask this paper
K x K feedback matrix. K agents each produce a test and a patch. Running each test on each patch yields a K x K pass/fail matrix, and each agent refines its work from its own slice of the matrix instead of reading other agents' traces.
State of the art. TEX reaches 86.4% on SWT-Bench Verified, 9.0 points above single-agent baselines and above the previous best of 84.0%.
Whole pool improves. Picking a TEX output at random scores 84.4%, which already exceeds the prior state of the art, so the cross-execution feedback raises the quality of every candidate, not only the best one.
Low communication cost. Agents exchange execution results rather than long reasoning traces, which keeps the method usable on long-horizon tasks.
Abstract
Test-case generation is critical but remains challenging for language models. Existing test-time scaling methods either isolate candidates or aggregate traces, which do not scale to long-horizon tasks. We present TEX, a collaborative test-time scaling framework based on execution-driven cross-validation. TEX parallelizes agents to generate tests and patches, cross-executing each test against every patch to form a $K \times K$ feedback matrix. Agents use their matrix slices for refinement, enabling objective validation without verbose trace sharing. On SWT-Bench Verified, TEX achieves 86.4%, establishing a new state-of-the-art and improving +9.0% over single-agent baselines. Moreover, even random selection from TEX outputs (84.4%) surpasses the previous SOTA (84.0%), showing collaborative execution feedback improves solution quality, not just best-case samples. Our results highlight execution-grounded feedback as an effective scaling mechanism for long-horizon agentic tasks.