deepseek-v4.1IMAGE
LisanBench test suggests DeepSeek RL-trained V4.1 Flash 1017 on the benchmark
Testing V4.1-Flash-1017 on LisanBench, the poster found a chain of 64 two-letter words and concluded DeepSeek almost certainly RL-trained on that benchmark.
Lisan al Gaib
@scaling01Archived 11 days agoOriginally posted 2026.10.18
Lisan al Gaib @scaling01Originally posted 2026.10.18
I just tested DeepSeek-V4.1-Flash-1017 on LisanBench I've previously suspected that labs are RL'ing on it, but for DeepSeek I am 100% certain that they do for example: it includes a chain of 64 two letter words I have only seen this behavior once when I hillclimbed LisanBench myself with Qwen3.5 models a while back
This post quotes another post: x.com/scaling01/status/2111518455216394680
x.com/scaling01/status/2111528125280464945Archived 2026.10.18 16:29