Jialun Zhong, Jiali Zeng, Miduo Yu, Yanzeng Li, Jianpeng Hu, Siqi Xie, Fandong Meng, Lei Zou · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3836559
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The rapid evolution of autonomous agents has enabled increasingly complex long-horizon tasks in multi-modal environments. However, these systems remain vulnerable: subtle reasoning errors or visual misinterpretations can propagate through an execution trajectory and ultimately cause task failure. Failure Attribution (FA) faces two key challenges. (1) The heterogeneous nature of visual and textual modalities hinders accurate semantic alignment, making it difficult to ground high-level intentions in fine-grained visual evidence. (2) Long execution trajectories obscure long-range dependencies, where early reasoning errors often surface only after many subsequent actions. To address these challenges, we present VISTA, a training-free, plug-and-play V isual-textual Inspector for Step-wise Trajectory Failure Attribution. VISTA decomposes diagnosis into Perception, Analysis, Deduction, and Evaluation, bridging cross-modal gaps through explicit visual-textual alignment while mitigating error propagation via milestone summarization and forward state extrapolation. Experiments across diverse benchmarks show that VISTA consistently outperforms existing baselines, providing an interpretable framework for failure diagnosis, agent harness and real-time optimization.
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
No comments yet — start the discussion below.