Catastrophic Forgetting

01 · 新岗位考满分，老本行却不会了01 · Ace the new role, forget the old one

模型在新客服领域从 52 分升到 95 分，如果旧摘要任务从 94 分掉到 24 分，这次微调还能只报“成功”吗？A model rises from 52 to 95 on a new support task but drops from 94 to 24 on summarization. Can the fine-tune be reported as a success?

02提高新任务占比，看旧任务曲线下坠；加入回放样本再托回Raise New-task Share, Watch the Old Curve Fall, Then Recover with Replay

双曲线 · 回放救援two curves · replay rescue

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

Catastrophic Forgetting

NEW

新任务梯度New-task gradient

训练信号持续推高新领域得分。Updates keep raising new-domain performance.

OLD

旧能力漂移Old capability drift

缺少旧样本约束时，参数离开原能力区域。Without old-data constraints, parameters leave the prior capability region.

REPLAY

回放样本Replay samples

在新训练中混回一部分旧任务样本。Mix a portion of old-task examples into new training.

JOINT EVAL

联合验收Joint evaluation

新任务增益与旧能力保持必须同时报告。Report new-task gain and old-capability retention together.

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

MECHANISM

accept = new_task_gain ≥ target ∧ old_task_drop ≤ guardrail

新任务占比驱动两条曲线反向移动；回放比例只托住旧曲线，明确展示折中而非魔法恢复。New-task share drives the curves in opposite directions. Replay lifts the old curve, showing a trade-off rather than a magical full recovery.

集中学习新域Focus on new domain新任务得分上升new-task score rises

旧信号缺席Old signal absent原能力边界漂移prior capability drifts

回放与联合验收Replay and evaluate守住旧能力护栏protect old-capability guardrails

05训练批次 100% 新任务、0% 回放Train on 100% New-task Data with 0% Replay

主动制造失败Create a failure

边界实验Boundary experiment

只看新任务会得到漂亮曲线；把旧任务曲线同时保留，才能看到能力坠落。The new-task curve looks excellent in isolation. Keeping the old-task curve visible reveals the capability collapse.

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.

✗ 新任务涨分就说明微调成功。A higher new-task score proves the fine-tune succeeded.→ ✓ 必须同时检查旧能力护栏。Old-capability guardrails must also pass.

✗ 灾难性遗忘等于模型删除了旧数据。Forgetting means old data was literally deleted.→ ✓ 它描述能力下降，不等同于存储删除。It describes performance loss, not literal data deletion.

✗ 加入回放就能无代价完全恢复。Replay restores everything at no cost.→ ✓ 回放占预算，也可能降低新任务适配速度。Replay consumes budget and can slow new-task adaptation.

06 · 一句话带走06 · One line to keep

灾难性遗忘让新任务得分上升时旧能力同步坠落；回放可缓解，但最终必须用新旧任务双护栏共同验收。Catastrophic forgetting lets new-task performance rise while old capability falls. Replay can help, but acceptance requires guardrails for both.