“…Recently, case-based reasoning (CBR), which utilizes a memory of past labeled examples as cases, has emerged as a promising paradigm of inference-time adaptation without finetuning (Das et al 2020(Das et al , 2021Pasupat, Zhang, and Guu 2021;Gupta et al 2021). CBR has been found effective for tasks like knowledge graph completion (KGC) (Das et al 2020), question answering over knowledge bases (KBQA) (Das et al 2021), task-oriented semantic parsing (Pasupat, Zhang, and Guu 2021;Gupta et al 2021), translation (Khandelwal et al 2021), and text-based games (Atzeni et al 2022). However, many prior CBR approaches designed around Seq2Seq architectures simply concatenate input-output cases with the current input at the encoder (Das et al 2021;Pasupat, Zhang, and Guu 2021;Gupta et al 2021).…”