Beyond Outcomes: Dual-View Relational Learning for Efficient Agent Benchmarking
DualViewEval compresses agent benchmarks to 20 tasks while still estimating full-benchmark scores.
The paper says agent evaluations are expensive because they depend on large-scale task trajectories, not just final answers. The authors identify six process signals tied to final performance and combine them with outcome relations to choose an exact-size miniset. Across five agent benchmarks, DualViewEval outperformed five baselines, including 24x-40x compression on APEX-Agents and BFCL. The selected minisets are also presented as diagnostic, showing capability differences among agents. ArXiv · AI/CL/LG's note
The paper says agent evaluations are expensive because they depend on large-scale task trajectories, not just final answers. The authors identify six process signals tied to final performance and combine them with outcome relations to choose an exact-size miniset. Across five agent benchmarks, DualViewEval outperformed five baselines, including 24x-40x compression on APEX-Agents and BFCL. The selected minisets are also presented as diagnostic, showing capability differences among agents. ArXiv · AI/CL/LG's note
score 5