Beyond Outcomes: Dual-View Relational Learning for Efficient Agent Benchmarking

The paper derives six benchmark-agnostic process measurements from large-scale agent trajectories: agent steps, tool failed rate, tool-category entropy, validation-tool rate, required-write execution, and read–write–validate closure. Prior compression methods mainly model redundancy in task–model final-score distributions, whereas this work uses cross-benchmark correlation analysis to select six automatically extractable, semantically non-redundant process signals from twelve candidates, bringing execution behavior into the representation. Based on analysis of large-scale execution traces from five agent benchmarks with open trajectories, reporting systematic associations between these measurements and full-benchmark scores, plus distinct process distributions for low-, middle-, and high-scoring BFCL models.

It proposes DualViewEval, an end-to-end dual-view framework that builds an outcome relation matrix and a process relation matrix per task, fuses them into an agent kernel feeding a Kernel Ridge score predictor, and uses a straight-through hard Top-K gate to keep the miniset exactly sized. Unlike methods such as SparseEval that mainly model relations between columns of the outcome matrix, this framework exploits both agent-level shared structure and the execution process, coupling task selection with the downstream prediction objective in a single feedback loop. The method is reported with means and standard deviations over ten splits across five benchmarks and three miniset sizes, accompanied by outcome/process relation matrix ablations, selector ablations, prediction-architecture ablations, and a budget sweep.

Across five agent benchmarks, DualViewEval achieves the best results on all datasets; with only 20 tasks it reaches 24×–40× compression on APEX-Agents and BFCL, reducing MAE by 14.5%–28.2% over the strongest competitors. Relative to SparseEval, the most directly comparable outcome-driven predictive baseline, it reduces MAE by 30.5%–45.1% averaged across benchmarks and improves mean Kendall's τ by 0.089–0.115; on APEX-Agents at the tightest budget it reduces MAE by 34.3% and raises τ by 0.100. Ten-split mean±standard-deviation results covering BFCL, τ²-Bench, Terminal-Bench 2, SWE-bench Verified, and APEX-Agents, totaling 1,983 tasks, 262 agent configurations, and 132,387 outcome pairs.

The selected minisets reveal capability differences among agents, the process weight γ adapts per benchmark, and training cost is lower than SparseEval. Beyond score prediction, the compressed subsets preserve diagnostic behavioral evidence; the learned γ distributes differently across benchmarks, and a fixed-γ sweep likewise shows intermediate weights generally give the best error–ranking trade-off. Reports aggregate process profile tables for APEX-Agents and SWE-bench Verified, γ distribution figures, a training-data proportion experiment, and roughly 4.2×–8.3× training-time speedups.

The result targets evaluation settings that need frequent agent assessment and can access execution trajectories: the method relies on automatically parseable trajectories and tool-call records, and applies to benchmarks such as BFCL, τ²-Bench, Terminal-Bench 2, SWE-bench Verified, and APEX-Agents where outcomes and trajectories can be aligned. It lets researchers and engineering teams estimate full-benchmark scores and rankings under a fixed task budget, and use the selected subsets to observe behavioral differences in tool use, validation, and write-operation closure; for evaluation settings without accessible trajectories or tool-call structure, the applicability of the process view needs separate validation.

Readers may still watch: the extraction stability of process measurements when trajectory formats differ substantially or tool-call parsing is incomplete; the learned process weight γ varies across benchmarks, indicating the optimal outcome–process balance is benchmark-dependent and may need recalibration on new benchmarks; the held-out model-family experiment is conducted only on the Qwen family in APEX-Agents, so generalization to other model families and benchmarks remains an open question; and although this is a full-text read, figures are presented as text, so some graphical details (such as exact values in the γ distribution and efficiency analysis) cannot be fully verified from the text.