Seed 2.1 Study Compares Reasoning, Models, and Methodology
The author ran 216 calls across Seed 2.1 lite and pro, reasoning toggles, and two prompting approaches on nine deep-research questions. Enabling reasoning produced the largest reported accuracy gains at nearly no additional cost, while upgrading from lite to pro improved accuracy but at a substantially higher cost per correct answer; the five-step methodology mainly benefited weaker configurations and evidence conformity. The experiment covers only nine questions, with wide confidence intervals and no p-values surviving Holm correction.
The comparison gives a concrete cost-and-accuracy snapshot for deep-research workflows, but its small sample and statistical limitations mean the findings should guide hypotheses rather than serve as a final conclusion.
给大模型一堆材料、一个有争议的问题，让它给出立场和依据——这类深度研判（deep research）任务现在做得人越来越多，想做好它的建议也分三派：开深度思考、换更贵的模型、套一套研究方法论。我按这三条路做了一个 216 次调用的对照实验，用 Seed 2.1 lite 和 pro 两档模型、八种配置对着标准答案算正确率。结论先给出：思考开关的收益最大，且几乎不额外花钱；换模型有效，但每块钱买到的提升最少；方法论的收益取决于配置——弱配置靠它补短板，强配置上它推不动推理结果，但能明显改善证据的规范性。
这个排序不是体感。体感在单样本上靠不住：一次答对可能是运气，一次答错可能是检索没把段落送进来。要回答「哪个有用」，需要把三个因素同时放进对照，让差异只来自被考察的变量。下面交代怎么做到的、测出了什么、机制上为什么，以及这份数据允许我们说到哪一步。