📖 Read?View on arXiv
Long-Horizon Agent Trajectory Attribution: A Unified Benchmark and Fine-Grained Annotation Framework
agentsobservabilitysafetybenchmarking
2608.06909
Builder Relevance
Aug 787%
Abstract
This paper frames trajectory attribution as identifying which components of a multi-step agent execution contributed to a target behavior.
Reality Card
Core Claim
Agent observability should attribute outcomes to instructions, observations, tool interactions, memory, and intermediate actions rather than inspecting only a final answer.
Method / Result
The paper reports more than 1,300 annotated trajectories and a reusable protocol for model-adaptive trajectory annotation and evaluation.
Limitations
The benchmark focuses on attribution annotations and evaluation structure; teams must still test whether it matches their own agent framework, logging fidelity, and risk model.