Title: SagaQA: A Multi-hop Reasoning Benchmark for Long-form Narrative Understanding in TV Series

Abstract:

We present SagaQA, a novel benchmark designed to evaluate multi-hop reasoning across full-length television series within the domain of long-form video analysis. While current video reasoning datasets typically focus on local comprehension of adjacent frames or short clips, SagaQA fills a critical void by demanding high-level interpretation of extended multimodal narratives spanning complete shows. A key differentiator of this dataset is its reasoning granularity; it requires models to execute long-range reasoning hops that link information across disparate episodes. This task forces systems to reason over entire events and actions, necessitating a profound, multimodal grasp of the show’s narrative arc and progression.

Driven by recent advancements in agentic methodologies, we also investigate how various planning strategies manage these intricate reasoning challenges. We classify these methods into three distinct categories: Parallel, Sequential, and Hybrid planners, and assess their proficiency in generating coherent and comprehensive reasoning plans. Our findings on the SagaQA benchmark indicate that hybrid planners consistently yield superior-quality plans, demonstrating more robust capabilities for complex, high-level narrative understanding in television content.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC