ReDraft: Continual Post-Training With 11.3x Less Forgetting

Two numbers on the same model

A 37-page preprint, posted on arXiv on 15 September 2026 by a group of 18 authors, places two numbers side by side. On three visual tasks, with Qwen2.5-VL at 3 and 7 billion parameters, the proposed method gains 56.9 points on the new task and loses 1.5 points on the previous ones. Classic supervised fine-tuning, on the same data, gains 52.9 points and loses 16.6.

The ratio between the two losses is 11.3 times. That is the figure the ReDraft abstract^[1] puts front and centre.

The evidence shows one specific thing. The tension between learning a new task and preserving already acquired capabilities, in the continual post-training of multimodal (vision-language) large language models, depends on the target chosen for the update. Change the target, and the tension shrinks by an order of magnitude on the same models and the same tasks.

What the authors measured

The protocol uses three tasks: Counting, Clock Reading and Jigsaw. Two of the three start from near-zero accuracy on the base model, a detail that weighs on how the results should be read.

The models are two variants of Qwen2.5-VL, at 3 and 7 billion parameters. The authors measure two quantities for each method: the gain on the task just learned and the average loss on the previous tasks.

Three methods are compared on the same pairs of numbers. The first is supervised fine-tuning (SFT), which trains the model on the expert's answer. The second is OPSD, the on-policy self-distillation variant used as a comparison, which trains the model on its own outputs. The third is ReDraft, the method the group proposes.

The mechanism: the distance between target and policy

The authors' explanation starts from the nature of the target. SFT gives an explicit, strong signal: the expert's correct answer. That signal is off-policy, far from what the model would produce on its own, and it shifts the weights enough to erase earlier capabilities.

On-policy methods, such as RLVR and self-distillation, stay close to the current policy. The problem is the opposite one: when the model starts from zero on the task, its rollouts contain little useful signal. A model that is almost always wrong produces few examples to learn from.

ReDraft takes a middle path. It shows the model the expert's answer as a reference, asks it to revise its own incorrect rollout, and keeps the revision only when a verifier accepts it. Fine-tuning happens on whatever survives the filter.

Every retained target is therefore explicit, because it is correct, and at the same time close to the policy, because it originates from the model's own text.

The comparison with OPSD and an ambiguity in the abstract

The comparison with OPSD is less straightforward than it looks. The abstract reports the pair of 19.3 in gain and 6.2 in loss and states that ReDraft improves on both axes. The text leaves a doubt open: whether those two values describe OPSD in absolute terms or the margin between the two methods.

Under the first reading, OPSD learns little (19.3 points against 56.9) and forgets little (6.2 against 1.5). Under the second, the margin in ReDraft's favour is 19.3 points more gain and 6.2 points less loss. The preprint's 37 pages contain the tables that settle the question; the abstract, taken in isolation, leaves it intact.

Under both readings the direction stays the same.

The purely on-policy method pays in learning what it saves in forgetting, and reference-guided revision breaks the trade-off.

What the data-space and parameter-space analyses say

The authors add two analyses that make the result less accidental. The first is in data space: the revised targets turn out to be more probable under the base model than the expert answers. A more probable target requires a smaller shift in the weights.

The second is in parameter space. The updates induced by ReDraft stay compact and follow the direction of SFT more closely than OPSD does. The method learns in the same direction as the expert, with shorter steps.

It is this coherence between design and measurement that supports the paper's thesis.

The dichotomy between learning and preserving turns out to be an artefact of the off-policy target, rather than an intrinsic constraint of optimisation.

The verifier as a hidden variable

The filter that accepts or rejects each revision decides which examples enter training. In the paper's three tasks, verification is mechanical: a count, a time, the position of the pieces. The correct answer exists and is unique.

This property makes the revision loop cheap and reliable. On open-ended tasks, such as a summary or a diagnosis, the verifier's role passes to a judge, human or automated, and the cost of the filter grows.

The abstract omits the share of revisions accepted. That number measures how much inference work is needed for each useful example, and it remains to be read in the body of the paper.

The declared limits of the evidence

The evidence has a narrow perimeter, and that needs saying. Three synthetic tasks, two sizes of a single model, an automated verifier that decides what enters training. Counting objects, reading clocks and reassembling puzzles are tasks with a verifiable answer; many production tasks lack this property.

The figures cited here come from the preprint's public abstract. A full reading of the 37 pages, with protocols, variances and ablations, remains the necessary step before generalising.

The 11.3x ratio holds under these conditions. How well it holds on tasks with high initial accuracy, on larger models or in domains without a verifier remains a separate dimension, largely still to be measured.

What changes for those who allocate budget and design infrastructure

For those allocating R&D budget, the measured result shifts a question. The cost of forgetting in continual post-training depends on the distance between the target and the current policy, more than on the optimisation method. Spending to mitigate forgetting with regularisation techniques addresses a symptom, while the choice of target acts on the cause.

For those designing data infrastructure, the paper describes three components: a store of reference answers, a revision loop for incorrect rollouts, and a reliable verifier. The third is the bottleneck. The method holds up only as far as the verifier does.

For a board, the thesis the data supports is circumscribed. In multimodal continual learning on verifiable tasks, revising the model's own output beats replacing it with the expert's output, on both measured axes. Beyond this perimeter, the evidence is silent.