What happened
In a November 2, 2026 post, DoorDash authors Chaitanya Ahuja and Sebastian Zuba describe a comparison of self-distillation and GRPO, a reward-based optimization method, as on-policy post-training approaches. They used Nemotron Nano 9B, updated all model weights, and gave both methods the same number of training iterations. The main comparison used a base model with thinking off; the post also reports a separate self-distillation chain with thinking on. The training sequence was tool use, science, then Knights & Knaves logic puzzles, with each stage starting from the previous stage’s weights.
The authors evaluated the three trained tasks on held-out test splits and also tracked five broader themes—code, math, instruction following, logic, and knowledge—using twenty benchmarks. Their grading usually required the requested answer format, but they also examined functionally correct answers that failed formatting requirements separately. That distinction matters because some apparent losses reflected output format rather than incorrect underlying answers.
Both methods improved tool use and science. Tool-use performance rose 8 points with self-distillation and 20 with GRPO; science rose 33 and 29 points, respectively. The largest separation came on Knights & Knaves: GRPO reached 95.9%, compared with 31.6% for self-distillation and 47.7% for the base model. Self-distillation did not recover the 18 points it lost on that logic task during the first, tool-use stage.
Retention also diverged over the sequence. Self-distilled Nemotron ended 21 points below the base model on instruction following and 30 points down on logic; its math performance fell 15 points. GRPO never lost more than 0.2 points on any tracked theme at any stage and finished 7 points up on math, 4 on logic, and 3 on instruction following after the final task. Knowledge moved relatively little for either method, although MMLU-Pro fell from 62.9% to 54.0% under self-distillation; the authors suggest the expected chain of reasoning before the answer may explain that outlier.
DoorDash also observed qualitative differences. In a thinking-on self-distillation run, the model sometimes referred to a demonstration it had not seen, and it could carry the last training set’s answer format into unrelated tasks. Its math answers became shorter on average, from 896 to 606 tokens, while GRPO’s grew to 1,522 tokens. The share of AIME 2025 answers with an explicit self-check rose from 1% to 22% under GRPO, versus 4% under self-distillation. Mixing the three datasets rather than training sequentially softened most of self-distillation’s accuracy loss.
What the archive adds
This is a new experiment in a continuing discussion of distillation choices. An earlier comparison argued that method choice depends on teacher–student alignment, not a default preference for on-policy sampling. DoorDash adds a direct comparison with GRPO focused on continual skill acquisition and retention. Earlier notes on OPSD and JOLT concern related self-distillation methods, not this DoorDash result.
Trend signal
This supports the agent-economics-moves-to-completed-work pattern only weakly: the comparison makes clear that an added capability is not useful if it damages established ones, but it does not measure cost per completed task. More directly, it reinforces the archive’s emphasis on evaluating post-training by both acquired behavior and retained capabilities rather than by a single target-task score.
Why it matters
For teams maintaining a model across expanding tool and task requirements, a successful training stage is not simply one that improves its newest benchmark. It must avoid degrading capabilities already relied on in production. DoorDash’s results show how sharply that tradeoff can vary by method: in its setup, self-distillation learned some tasks but accumulated losses in logic, instruction following, and math, while GRPO preserved those themes much more consistently.
The result is a reason to build retention checks into the post-training loop. Teams can evaluate a fixed set of representative capabilities before training and after every stage, separating formatting failures from genuinely wrong answers. That will not establish which optimization method is best for a different model or reward setup, but it can expose regressions before a new checkpoint is deployed. The study also suggests task order and whether datasets are mixed may affect outcomes, so a single sequential run should not be treated as a universal recipe.
Limits and uncertainty
This is a DoorDash experiment on Nemotron Nano 9B, a particular three-task sequence, and the authors’ chosen training and evaluation setup. The post does not establish that GRPO will outperform self-distillation on other base models, reward signals, task mixtures, or compute budgets. The excerpt reports equal iteration counts, but does not provide enough information here to compare total costs or wall-clock time. Some evaluation differences may also reflect formatting: relaxing format checks reduces some measured self-distillation losses, though it does not eliminate them. The reported averages and examples are evidence about this experiment, not a general guarantee of retention.
Reader takeaway
Continual post-training should be evaluated as a two-sided problem: did the model learn the new task, and what did it forget along the way? In DoorDash’s Nemotron Nano 9B comparison, GRPO performed much better on the final logic task and preserved prior capabilities more reliably. Treat that as a reason to test retention in your own setup, not as a universal ranking of methods.