Stabilizing language models under continual learning via condition-anchored distillation

Continual adaptation of language models can change their output distribution on prompts learned earlier, while retaining every old prompt-answer pair may be undesirable or impossible.

语言模型的持续适应可能会改变其在先前学习的提示上的输出分布，而保留每一个旧的提示-回答对可能是不可取的或不可能的。

We study condition-anchored generative distillation (CAGD): retain a small set of old prompts, use a frozen previous model to reconstruct completions and generation states, and match its predictive distributions while learning the next task.

我们研究了条件锚定生成蒸馏（CAGD）：保留一小部分旧提示，使用冻结的先前模型来重构补全内容和生成状态，并在学习下一个任务时匹配其预测分布。

The formulation separates three roles that ordinary replay conflates: conditions select the behavior to protect, teacher generations locate relevant states, and soft targets specify how predictions may change.

该公式分离了普通重放所混淆的三个角色：条件用于选择需要保护的行为，教师生成用于定位相关状态，软目标则规定了预测可能发生的变化方式。

For autoregressive language generation, teacher-rollout distillation admits an exact chain-rule decomposition of sequence divergence. For masked-diffusion language modeling, our implementation directly controls local denoising drift on teacher-generated completions.

对于自回归语言生成，教师展开蒸馏允许对序列散度进行精确的链式法则分解。对于掩码扩散语言建模，我们的实现直接控制了教师生成补全内容上的局部去噪漂移。

In continual adaptation of a 219M masked diffusion language model, CAGD reduces four-task final held-out loss from 2.927 to 1.114 in one task order and from 2.168 to 0.891 in exact reverse.

在对一个 2.19 亿参数的掩码扩散语言模型进行持续适应的过程中，CAGD 在一种任务顺序下将四项任务的最终留出损失从 2.927 降低至 1.114，在完全相反的顺序下从 2.168 降低至 0.891。

The same soft targets lower final average loss by 0.055 over hard replay when teacher-generated support is held identical. The direction persists on fresh facts and natural instructions across SMDM and Qwen3.

当教师生成的支持内容保持一致时，相同的软目标比硬重放降低了 0.055 的最终平均损失。这一趋势在 SMDM 和 Qwen3 的新事实和自然指令上依然存在。

On GSM8K, Qwen adaptation preserves answer-format compliance, but exact-match retention is seed-mixed at 0.6B and worsens at 1.7B. These results support condition-anchored functional preservation as a common design principle across the tested language-generation objectives.

在 GSM8K 上，Qwen 的适应过程保持了对回答格式的合规性，但精确匹配的保留率在 0.6B 模型上受随机种子影响，在 1.7B 模型上则有所恶化。这些结果支持将条件锚定功能保持作为跨测试语言生成目标的一种通用设计原则。