Efficiency Forgetting: Does Concise Reasoning Survive Continued Training?

Making a reasoning model concise is treated as the last step of post-training: you make the model efficient, then you ship it. But shipped models keep being fine-tuned. We ask whether conciseness survives that further training, and whether anyone would notice if it did not. Continuing to train a published concise model—VeriThinker, built on DeepSeek-R1-Distill-Qwen-7B—on ordinary in-domain math data drives its thinking length back up by 24–41% on MATH500 and 17–18% on OlympiadBench, with paired-bootstrap 95% CIs excluding zero, while accuracy never moves. We name this Efficiency Forgetting: the model still answers just as correctly, only slower and more expensively, so the compute a deployer paid to save silently evaporates, invisible to every accuracy metric. The effect is bounded in two ways. It is data-specific: off-domain code and chat data leave length statistically unchanged. And it needs headroom: on GSM8K, which this model already answers in 200 thinking tokens at 94%, there is no relapse to be had. It is also not mere imitation of long training data, since the traces we train on are three times shorter than the model’s own reasoning.