DeepSeek V4.1-Flash Shows Reduced Phase-Sensitive Retrieval Weakness
ByteDance Seed researchers linked periodic retrieval failures in DeepSeek-V4-Flash-Base to position-dependent behavior in chunked KV-cache compression, with a 40.2-point gap between the best and worst position groups in a controlled 128K-context test. LeiPhone AI tested DeepSeek V4.1-Flash through the official API and found all 64 fixed-sampling completions correct, although prediction probabilities varied as irrelevant prefix length changed. Seed’s data show the best-to-worst position gap narrowing from 40.2 points in V4-Flash-Base to 19.1 points after post-training and 6.1 points in V4.1-Flash.
The results show that DeepSeek’s newer version reduced the measured position sensitivity of compressed-cache retrieval, while also demonstrating that average retrieval scores can conceal weaknesses at particular information positions.
导语：同一段代码，前面只增加无关Token，DeepSeek会改变答案吗？
同一段代码，前面只增加无关Token，DeepSeek会改变答案吗？
作者丨李娜
编辑丨岑峰
#DeepSeek#ByteDance Seed#DeepSeek V4.1-Flash#DeepSeek-V4-Flash-Base#DeepSeek V4#Qwen3-0.6B