Correct Choices, Failed Sequences: Choice2Seq Benchmarks Correctness Retention Across Output Interfaces

Tianyou Lai, Wentao Yue, Zhiying Deng +1

A model may select the correct answer yet fail to construct it through a sequence of local decisions. We introduce Choice2Seq, a diagnostic benchmark that measures correctness rete…