DecisionBench: A Benchmark for Emergent Delegation in Long-Horizon Agentic Workflows

Yuxuan Gao, Megan Wang, Yi Ling Yu, Zijian Carl Ma, Ao Qu

arXiv:2605.19099 · 2026-05-20 공개 · arXiv · PDF

long-horizon agentic-workflows orchestration benchmark-substrate routing-fidelity multi-axis-metrics peer-models task-suite

Abstract

We introduce DecisionBench, a benchmark substrate for emergent delegation in long-horizon agentic workflows. The substrate fixes a task suite (GAIA, tau-bench, BFCL multi-turn), a peer-model pool (11 models, 7 vendor families), a delegation interface (call_model plus an optional read_profile channel), a deterministic skill-annotation layer, and a multi-axis metric suite covering quality, cost, latency, delegation rate, routing fidelity-at-k, vendor self-preference, and a counterfactual-delegation ceiling. The substrate is agnostic to how peer information is generated or delivered, so learned routers, richer peer memories, adaptive profile construction, and multi-step delegation can all be evaluated against it. We characterize the substrate with a five-condition reference sweep on the full pool (n=23,375 task instances). Three benchmark-level findings emerge: (i) mean end-task quality is statistically indistinguishable across the four awareness conditions (|beta| <= 0.010, p >= 0.21), so quality-only evaluation would miss the orchestration signal; (ii) routing fidelity-at-1 ranges from 7.5% to 29.5% across conditions at near-equal mean quality, with delivery channel (on-demand tool vs. preloaded description) dominating description content; (iii) a counterfactual ceiling places perfect delegation 15-31 percentage points above measured performance on every suite, locating large unrealized headroom for future orchestration methods. We release the substrate, annotation layer, reference intervention suite, analysis pipeline, and 220 per-condition run archives.

한국어 요약

한 줄 요약

**[멀티 에이전트 / 위임 평가]** 장기 horizon 에이전트 워크플로에서 emergent delegation 능력을 다축 지표로 평가하기 위한 벤치마크 substrate DecisionBench와 23,375 인스턴스 reference sweep 분석.

핵심 기여도

핵심 아이디어

"qualtiy만으로 평가하면 orchestration 신호를 놓친다"는 핵심 진단. 같은 평균 품질을 내는 두 시스템도 라우팅 충실도·delegation 효율은 매우 다르며, "완벽한 위임"이 가능했을 때의 ceiling을 정의하면 미래 orchestration 연구의 헤드룸이 명확해진다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: emergent delegation 연구를 비교 가능한 평가 substrate 위로 올려놓고, "quality only" 평가의 한계를 정량적으로 폭로함. **한계**: peer pool·태스크 스위트가 고정되어 있어 매우 새로운 도메인의 위임 패턴은 별도 확장이 필요.

실용적 활용