Summary
DualViewEval selects fixed-size benchmark minisets using both final outcomes and six trajectory-level process signals. With 20 tasks, the method reports 24x to 40x compression on APEX-Agents and BFCL. It reduces mean absolute error by 14.5% to 28.2% over the strongest baselines and improves ranking correlation on SWE-bench Verified. The study spans five agent benchmarks and includes held-out model-family tests.
Why it matters
For teams that cannot rerun full agent suites on every release, process-aware subsets could turn evaluation from an occasional audit into a routine gate. Keep a rotating full-suite audit because a learned miniset can age as models, scaffolds and task distributions change. The absence of public code raises the cost of independent verification.
Technical details
| Method | joint outcome and six process-signal relations |
|---|---|
| Miniset Size | 20 |
| Compression | 24x-40x on APEX-Agents and BFCL |
| Error Reduction | 14.5%-28.2% MAE versus strongest baselines |
| Coverage | five agent benchmarks; held-out Qwen-family generalization test |
Tags
DualViewEvalagent-evaluationbenchmark-compressiontrajectorySWE-bench