Training Objective

01 · 同一场练习，两种评分01 · One exercise, two grading rules

同一段课文，一位老师考“续写下一个词”，另一位考“补回空格”，学生会练成同一种本领吗？One teacher asks for the next word; another asks students to restore a blank. Does the same passage train the same skill?

02切换两种目标，让同一样本的损失位置重画Switch Objectives and Redraw Loss Positions on the Same Sample

S5 玩具级概念图S5 toy concept map

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、得分与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。All probabilities, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Training Objective

SAMPLE

同一样本Same sample

五个 token 保持不变，排除数据变化。Five tokens stay fixed so the data does not change.

VISIBILITY

可见范围Visibility

目标决定训练时哪些 token 可被看到。The objective decides which tokens are visible during training.

LOSS

损失位置Loss positions

只有被评分的位置产生直接误差信号。Only scored positions produce a direct error signal.

UPDATE

更新方向Update direction

误差信号回传后推动不同的行为。Backpropagated error pushes toward different behavior.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

objective(output, target) → loss → parameter update

Canvas 固定样本，只换 objective；loss 标记从每个下一 token 位置重排到单个 MASK。The Canvas fixes the sample and changes only the objective; loss markers move from next-token positions to one MASK.

定义期望Define target 下一词或被遮词next token or masked token

计算差异Measure gap 输出对目标output versus target

形成更新Create update 朝降低 loss 的方向toward lower loss

05只换目标却沿用旧 loss 位置Change the Objective but Keep Old Loss Positions

主动制造失败Create a failure

边界实验Boundary experiment

从因果目标切到掩码目标后，若还在所有下一词位置算损失，实际优化的仍是旧任务。After switching from causal to masked training, keeping every next-token loss position still optimizes the old task.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ 数据相同，训练结果就相同。Same data means the same learned behavior. → ✓ 目标会改变遮挡、评分位置与更新方向。The objective changes visibility, scoring positions, and update direction.

✗ loss 就是模型犯错的次数。Loss is simply the number of mistakes. → ✓ 它是由目标定义的可优化数值，不必等于错误计数。It is an optimizable number defined by the objective, not necessarily an error count.

✗ 目标写在文档里就会生效。Writing an objective in a document makes it active. → ✓ 它必须落实为输出、目标与 loss 的计算规则。It must be implemented as a rule over output, target, and loss.

06 · 一句话带走06 · One line to keep

训练目标规定“看什么、在哪算 loss、朝哪更新”；同一样本切换目标，学到的行为也会改变。The training objective controls visibility, loss positions, and update direction. Switching it changes what the same sample teaches.