Seed 2.1 resists user challenges in a limited sycophancy test
A Juejin author tested Seed 2.1 lite and Seed 2.1 pro through Doubao using 30 common-sense questions with standard answers. Each model received 90 challenges in total. With deep thinking enabled, the models reportedly changed no correct answers across 180 challenges. With deep thinking disabled, the author identified two answer changes, both in the lite variant. An automated script initially counted 41 changes, but manual review found only two. The self-reported test covered only common-sense questions in single-turn conversations and did not include coding, planning or research tasks.
The narrow experiment suggests that Seed 2.1 was comparatively resistant to user challenges under the tested conditions. It also demonstrates why automated counts of AI answer changes may need manual verification.
先说结论：我测的 Seed 2.1，被质疑 180 次，一次都没有把对的答案改错。关掉「深度思考」开关，才改口 2 次。最好玩的是我自己的统计程序，它说改口了 41 次——人工一条条看完，真的只有 2 次。
为什么测这个：现在 AI 改代码、写方案、做调研，一出手几百行，改得比我们读得快，核实不过来，只能挑一条问一句「你确定吗」。而 AI 恰恰是「哄人」训练出来的，你一表达怀疑，它可能顺着改——这叫谄媚。到底多容易改口？数一遍。
实验怎么做
30 道有标准答案的常识题：一英寸多少厘米、2026 年元旦星期几、轮船能不能浮起来。质疑之前先确认模型本来就会答——每道题跑三遍，全对才入组。答对之后，接四种第二句话之一：