Qwen Picks What to Remember: Replay From a Frozen Model Reduces Forgetting

Fine-tuning a language model almost always breaks something it could do before. A new paper on arXiv (cs.CL, 2610.12602) tests an unusual idea: let the model itself protect its old skills. The authors fine-tuned Qwen models of different sizes on a sequence of instructions and found that a frozen copy can generate replay data that reduces forgetting, including on tasks that are not on the list of skills to preserve. At 4B, the model also chose which tasks to replay better than random selection did. Smaller models showed no such advantage.

The problem: the preservation list is always incomplete

The classic continual learning setup requires you to define a set of tasks on which quality must not drop. But a model's knowledge is broader than any such set, and forgetting is not limited to the items you check. Whatever you left out of the test degrades unnoticed.

Recent work has shown that a model can create useful replay on its own, without access to the original training data. The authors extend this line of research and ask: can a model help choose what to preserve, so that it protects itself more broadly than a human with a limited list would?

How the experiment works

For the study, task descriptions and their examples are organized into Task Banks. Entries from a bank can be selected and combined, and a frozen model generates replay data from them for fine-tuning. The authors then run continual instruction tuning on Qwen-family models of different sizes.

One important detail about the metric. Besides average accuracy, the authors track which previously correct answers stay correct. An average can look stable while the model loses some correct answers and gains others by chance.

• Models: Qwen of different sizes; the key result on task selection comes from 4B.
• Method: replay generated by a frozen model from Task Bank entries.
• Main metric: retention of previously correct answers, including on tasks outside the bank.
• Selection comparison: selection that accounts for the new task versus random selection.
• Constraint: a fixed budget of training examples.

Three results

First: which skills degrade depends on the new task and the learning rate. Tracking previously correct answers exposes losses that average accuracy hides.

Second: generated replay reduces forgetting on tasks outside the bank. Combining partial banks strengthens the effect: average retention of previously correct answers is higher than with any single bank, across all tested sizes.

Third: a model can take part in choosing replay targets. At 4B, selection that accounts for the new task raises average retention on held-out tasks compared with random selection, while the gain on the new task itself stays comparable. Smaller models did not show this advantage.

The price of protection

There is no free insurance here. With a fixed example budget, replacing part of the new-task data with replay improves retention but reduces the gain on the new task. The abstract does not disclose the size of this trade-off, so you will have to assess it from the full text and on your own data.

There is also a question of scale. Because the ability to select replay targets appeared only at 4B, it cannot be assumed to be universal across models. The results come from a single model family, and transfer to other architectures needs to be checked separately.

Practical takeaways for teams that fine-tune models

If you are adapting an open model to your task and did not keep the original data, generating replay with a frozen copy becomes a workable way to protect it. Start with several small task banks and combine them rather than relying on one. Evaluate not only average accuracy but also the share of preserved correct answers, including on tasks outside your test set. Plan your budget: every replay example takes space from the new-task data, and the balance is best tuned experimentally. Automatic selection of replay tasks is worth testing on models from 4B up, while at smaller sizes it is wiser, for now, to stick with proven approaches.