PyroDash, posted to arXiv 2607.20327, proposes a token-level small/large-model collaborative inference framework: it inserts a "help-request control token" during the small-model generation process, triggering a one-time hand-off that throws the hard problem to a frozen large model to complete. The core idea is to internalize "when is it worth spending big money to call an LLM" into the SLM itself, with no external router, no need to retrain the LLM, and no access to its logits. Training proceeds in three stages: control-token embedding learning → offloading-oriented SFT → cost-aware alignment with GRPO, where the reward function directly ties inference cost to answer accuracy. On 5 math reasoning benchmarks, with λ=0.05, the average accuracy is 64.04% — 6.36 percentage points above the pure-LLM baseline — and cost is cut 20.4%; with λ=0.6, accuracy drops to 54.55%, but the LLM token share is only 1.9%, with 0.012 LLM calls per sample, and the average inference cost drops from $49.36 to $1.78. The real value of this approach is turning "on-demand escalation" into a learnable policy rather than an engineering hard-route: the small model itself knows when it can't compute any further, and has learned to raise its hand precisely. The lesson for production deployment is that you can keep the upper bound of LLM inference capability while offloading most of the easy requests to cheap SLMs — this is the cake that routing papers have been drawing for a long time, but PyroDash actually lands it with one control token plus GRPO. The paper is 73 pages, code and data aren't public, and more independent reproduction is needed to verify the thresholds and reward function's transferability across domains.