Claim that DeepSeek used LisanBench for reinforcement learning

A tester says they are certain DeepSeek used LisanBench for reinforcement learning after testing DeepSeek-V4.1-Flash-1017. They point to a chain of 64 two-letter words, saying they had seen similar behavior only when hill-climbing LisanBench with Qwen3.5 models. In a follow-up, they said the DeepSeek model performed very well but still used a ton of tokens.