standing and retrieval of two separate video clips, even though their 1-needle performances are high. The more distant the two clips are, the worse the models perform.
2 Related Work
2.1 Long Video Understanding Benchmarks
Retrieval of information from a specific location in a long video (usually called “needle in a haystack”) is widely regarded as a key capability in long-video understanding (Kamradt, 2023; Wang et al., 2024b). In Tab. 1, we compare recent long-video benchmarks with multiple-needles-in-haystack tasks side by side to highlight the unique contribution of the proposed Causal2Needles dataset.
EgoSchema (Mangalam et al., 2023) does not measure specific model capabilities (e.g., 1-needle vs. 2-needle, reasoning vs. recognition) as the questions are not categorized by the required capability, resulting in limited diagnostic precision.
MLVU (Zhou et al., 2024) require models to retrieve visually distinct images that are artificially inserted into videos, which may allow for shortcut features due to the domain gap between the artificial needle and the original content.
Multi-needle questions also exist in MVBench (Li et al., 2024c) and TVBench (Cores et al., 2024), which contain questions to evaluate the model ability to track an object across multiple temporal locations.
Similar questions are also featured in VideoMME (Fu et al., 2024) (e.g., counting problems) and LongVideoBench (e.g., L2-SSS questions) (Wu et al., 2025). However, questions in the above datasets typically require only independent understanding of each needle. Moreover, the multiple needles often only related by surface-level visual cues, such as matching visual appearances of objects, rather than semantic relations like causality.
Table 1: A comparison of Causal2Needles with other multi-needle long video benchmarks. Causal2Needles is the only benchmark dedicated to needle-in-haystack problems (Diagnostic Precision) that requires joint understanding of the two needles and the identification of cause events from effect events.
Benchmark
Video Length
# QA
Diagnostic
Precision
Needle
Type
Joint
Understanding
Causal
Reasoning
EgoScheme (Mangalam et al., 2023)
180 s
5,000
✗
Natural
✗
✓
MVBench (Li et al., 2024c)
16 s-40 s
4,000
✓
Natural
✗
✓
TVBench (Cores et al., 2024)
16 s-40 s
2,525
✓
Natural
✗
✗
MLVU (Zhou et al., 2024)
180 s-3600 s
3,102
✓
Artificial
✗
✗
VideoMME (Fu et al., 2024)
1018s
2,700
✓
Natural
✗
✓
LVB Wang et al. (2024a)
473 s
6,679
✓
Natural
✗
✗
Causal2Needles (Ours)
438 s
4,100
✓
Natural
✓
✓
In contrast, the Causal2Needles benchmark requires models to jointly understand both the cause and the effect—the understanding of the cause relies on the correct interpretation of the clue in the effect. Furthermore, the two needles are related by causality rather than visual similarity, necessitating semantic understanding. These designs establish a challenging and unique benchmark for long-video understanding.
2.2 World Model
There is an ongoing debate (Gurnee and Tegmark, 2023; Liu et al., 2025; Guan et al., 2024b; Hu et al., 2024; Kang et al., 2024; Motamed et al., 2025) over whether deep neural networks learn to develop internal world models—whether they can identify the underlying principles, such as laws of physics, that govern the observational data they learn from. Such world models, if present, can be used to predict future outcomes of possible interventions, such as kicking a ball or blowing a candle (Forrester, 1971; Ha and Schmidhuber, 2018; Li et al., 2025).
However, this debate largely overlooks causality between events involving humans, which could be induced by factors like biology (e.g., heavy sweating causes thirst), psychology (e.g., hearing compliments makes people feel good), and social norms (e.g., providing good service leads to tipping). Although causal questions sporadically appear in video understanding datasets, such as the episodic reasoning questions in MVBench, it is usually difficult to isolate them and quantitatively m