EXPERIMENT IX · INFERENCE

想够了吗Thought Enough

模型能否知道一道题什么时候已经“想够了”？Can a model tell when a question has been "thought enough"?

待开始Not started提前退出Early exit推理成本Inference costKV CacheKV Cache
  1. Q提出可证伪的问题State a falsifiable question当前阶段Current stage
  2. R做最小运行路径Build a minimal runnable path未开始Not started
  3. E连接真实验证Validate against reality未开始Not started
  4. P公开证据与限制Publish evidence and limits未开始Not started

为什么值得做Why it is worth doing

增加计算深度能换来准确率，也会换来时延和成本。如果能按题目难度提前退出，简单的题就不必走完全程。More compute depth buys accuracy, and also latency and cost. If a model could exit early by question difficulty, easy questions would not need the full path.

Q

假设Hypothesis

HYPOTHESIS

基于中间状态置信度的提前退出，在准确率基本不变的前提下减少可观的平均迭代次数。Early exit driven by confidence in intermediate states cuts average iteration count meaningfully with essentially unchanged accuracy.

R

最小装置Minimal setup

MINIMAL SETUP
  1. 01在第 VII 号实验的循环模型上加入逐步置信度头Add a per-step confidence head to the looped model from experiment VII
  2. 02按难度分层构造测试集Build a test set stratified by difficulty
  3. 03画出不同阈值下的准确率–平均迭代次数曲线Plot accuracy vs. average iteration count across thresholds
E

什么会证明我错了What would prove me wrong

FALSIFIED IF

任何阈值下，提前退出都带来明显准确率损失。At any threshold, early exit costs clearly in accuracy.

P

将会公开What will be published

WILL PUBLISH

阈值扫描结果、难度分层数据、退出分布图。Threshold sweeps, difficulty-stratified data, and exit distributions.

前期线索Early leads相关的已有工作，不是这个实验的结论Related earlier work, not a result of this experiment

  • 承接第 VII 号实验Follows experiment VII需要先完成循环模型的复现。Requires the looped model reproduction to be finished first.

实验日志Run logRUN LOG

还没有记录。实验开始后，每一次运行、失败和修正都会记在实验室里。Nothing logged yet. Once the experiment starts, every run, failure and fix will be recorded in the lab.

去实验室Go to the lab ↗