Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows
The paper discusses the limitations of current hallucination benchmarks for Large Language Models (LLMs) in multi-agent workflows. It introduces Trajel, a dataset and evaluation framework designed to audit trajectory-level hallucinations. The study reveals that existing benchmarks often overlook common failure modes and emphasizes the need for taxonomy-grounded evaluation for safer deployment of autonomous agents.
- ▪Trajel introduces a five-type hallucination taxonomy based on expert-annotated agent traces.
- ▪Nearly half of hallucinated trajectories involve multiple types of hallucinations simultaneously.
- ▪Trajectory-aware detection significantly outperforms standard post-hoc verification methods.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.24219 (cs) [Submitted on 22 May 2026] Title:Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows Authors:Harshada Badave, Santosh Borse, Andrea Gomez, Harshitha Narahari, Sara Carter, Vishwa Bhatt, Aishani Rachakonda, Shuxin Lin, Dhaval Patel View a PDF of the paper titled Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Industrial Workflows, by Harshada Badave and 8 other authors View PDF HTML (experimental) Abstract:Large Language Models (LLMs) are increasingly deployed as autonomous agents that reason, use tools, and act over multiple steps.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
