AgentAtlas: Beyond Outcome Leaderboards for LLM Agents

Parsa Mazaheri, Kasra Mazaheri

arXiv:2605.20530 · 2026-05-22 공개 · arXiv · PDF

llm-agents model-evaluation agent-benchmarks benchmark-coverage control-decision-taxonomy trajectory-failure-taxonomy prompt-supervision tool-context-utility

Abstract

Large language model agents now act on codebases, browsers, operating systems, calendars, files, and tool ecosystems, but the benchmarks used to evaluate them are fragmented: each emphasizes a different unit of measurement (final task success, tool-call validity, repeated-pass consistency, trajectory safety, or attack robustness). A line of 2024-2025 work has converged on the diagnosis that a single accuracy column is no longer the right unit of comparison for deployable agents. AgentAtlas extends this line of work with four components: (i) a six-state control-decision taxonomy (Act / Ask / Refuse / Stop / Confirm / Recover); (ii) a nine-category trajectory-failure taxonomy with two orthogonal hierarchical labels (primary_error_source, impact); (iii) a taxonomy-aware vs. taxonomy-blind methodology that measures how much of a model's apparent capability comes from the supervision in the prompt; and (iv) a benchmark-coverage audit mapping fifteen agent benchmarks against six behavioral axes. To demonstrate the methodology we run a small fixed eight-model set (1,342 generated items, four frontier closed and four open-weight) under both prompt modes. Removing the explicit label menu drops every model's trajectory accuracy by 14-40 pp to a tight 0.54-0.62 floor regardless of family, and no single model wins on all three of control accuracy, trajectory diagnosis, and tool-context utility retention. We treat the synthetic run as a measurement-protocol demonstration, not a benchmark release.

한국어 요약

한 줄 요약

**[LLM 에이전트 평가 / 행동 분류 체계]** AgentAtlas는 6-state control taxonomy·9-category trajectory failure taxonomy·taxonomy-aware/blind 방법론·15 벤치마크 coverage audit 제공, label menu 제거 시 모든 모델 trajectory accuracy 14~40pp 하락(0.54~0.62 floor)·단일 승자 모델 없음.

핵심 기여도

핵심 아이디어

LLM 에이전트의 진정한 capability 측정은 단일 accuracy로 불가능하며, 행동 분류 체계와 prompt의 supervision 영향을 분리하는 taxonomy-aware/blind 비교가 supervision-leak 없는 평가의 필요조건이다.

기술적 접근법

주요 결과

의의 및 한계

**의의**: 에이전트 평가의 단일 accuracy 패러다임 명시적 거부, supervision-leak 진단 방법론, 행동 축 기반 벤치마크 갭 분석, 9-failure taxonomy로 진단적 평가 가능. **한계**: 측정 프로토콜 시연이라 정식 벤치마크 release 아님, 8 모델·1,342 item 규모 제한, taxonomy 보편성·완전성은 후속 검증.

실용적 활용