Reflected Survival Routing

Integrate this into a Transformer encoder as a token-wise adaptive-depth wrapper, or into an MoE block as a survival-aware router. At layer $\ell$, maintain a vector of survival masses $S\in[0,1]^B$ for the batch's tokens. The halting head receives the current token representation, the layer index, an estimated continuation value, and a mean-field statistic computed only from active tokens. Use the following pseudocode: initialize $S_i=1$ and $h_i=x_i$; for each layer $\ell$, define $A=\{i:S_i>\epsilon\}$, compute $\mu_\ell=\sum_i S_i g(h_i)/(\sum_i S_i+\epsilon)$, optionally compute temporal memory $c_\ell=\sum_{s\le\ell}\phi(\ell-s)\Delta L_s$, run the layer only for active tokens, predict $q_i=\operatorname{sigmoid}(f_\theta(h_i,\mu_\ell,c_\ell,\ell))$, and update $\Delta L_i=S_iq_i$, $S_i\leftarrow S_i-\Delta L_i$. Tokens with $S_i<\epsilon$ are removed from subsequent attention and their outputs are cached. Define $V^{stop}$ from an exit classifier or the residual stream's prediction head, and define $V^{cont}$ with a one-step value head predicting the benefit of one more layer minus its normalized FLOP cost. Add the task loss, compute penalty $\lambda_c\sum_{i,\ell}\Delta L_{i,\ell}C_\ell$, and complementarity loss $\lambda_{comp}[\operatorname{softplus}(-r)^2+(r\Delta L)^2]$. The paper-derived quantities are the monotone stopping measure, survivor occupation measure, and kernel convolution; the value heads, kernel bandwidth, and loss weights are learned or estimated empirically. First test on CIFAR-100 with a 6-layer ViT-Tiny, comparing fixed-depth inference, standard ACT, and independent token pruning at an equal accuracy target. Measure average active tokens per layer, FLOPs, calibration, and whether survival curves are monotone and non-oscillatory. A useful signal is at least 20% lower inference FLOPs at matched top-1 accuracy, with lower variance in compute across batches than independent halting.