MPAR-Bench: New Benchmark for Multi-Point Associative Reasoning in LLMs

A new bilingual benchmark, MPAR-Bench, has been developed by researchers to assess the reasoning breadth of large language models (LLMs) in both English and Chinese. Unlike current benchmarks that emphasize reasoning depth—capability to navigate extensive inferential chains—MPAR-Bench focuses on reasoning breadth by challenging models to synthesize various semantically distinct clues into a unified response. Drawing inspiration from the cooperative game Just One, each task involves uncovering a hidden target from multiple independently generated clues. The benchmark features 1,000 tasks created through a multi-agent clue-generation system, diversity filtering based on embeddings, and human validation. The answer pool comes from public word lists, with all clue sets generated anew to maintain originality. Evaluation metrics include exact-match accuracy and others. The related paper can be found on arXiv with the identifier 2608.10444, categorized as 'cross'.