Causal2Needles: Long-Video Benchmarks for 2-Needle Causal Reasoning

Causal2Needles is a benchmark dataset containing between 1,000 and 10,000 records designed for evaluating long-video understanding in multimodal large language models. Developed for the NeurIPS 2025 Datasets and Benchmarks Track, it focuses on "2-needle" reasoning tasks where models must retrieve and synthesize two separate temporal segments from a single video. The dataset provides multiple-choice questions that test causal reasoning across long-form visual content.

Use Cases

• Evaluating multimodal LLM performance on long-context video retrieval using 2-needle questions
• Benchmarking causal reasoning capabilities across temporal video segments
• Testing zero-shot visual question answering for long-form content

Strengths

• NeurIPS 2025 D&B Track peer-reviewed benchmark
• 1,000 to 10,000 evaluation samples
• Focuses on multi-point (2-needle) temporal reasoning
• Includes expert-designed multiple-choice question labels

Limitations

• Restricted to 2-needle scenarios which may not capture higher-order multi-step reasoning
• Non-commercial license (CC BY-NC-SA 4.0) limits industrial application
• Specific to long-context evaluation rather than general video classification

Provenance

Source

NeurIPS 2025 Datasets and Benchmarks Track (Arxiv: 2505.19853)

Freshness

Last updated November 2025; aligned with current NeurIPS 2025 research standards.