Cross-tokenizer knowledge distillation allows a student to learn from teachers with incompatible vocabularies while preserving its tokenizer and inference architecture. Existing logit-based methods either supervise only the observed token or match full distributions through strict token partitions and heuristic ranking. We identify two limitations: an uncommon-token failure, where task-critical tokens outside the shared vocabulary are suppressed, and over-conservative matching, where exact string equality discards useful correspondences. We propose X-Token, which addresses both through a sparse projection initialized from tokenizer-level string rules and optionally refined during distillation. P-KL removes the partition and projects the student distribution into the teacher vocabulary; H-KL retains the hybrid loss while relaxing matching through the projection. A token-coverage audit guides the choice between them, and both extend naturally to heterogeneous teachers. Across Llama, OLMo, and Gemma students, X-Token achieves the highest average accuracy in all four teacher-student settings among all evaluated cross-tokenizer methods. Complementary teachers provide a further 1.30-point gain, and both variants are faster than the evaluated cross-tokenizer baselines in matched profiling.
An invertible graph aggregation preserves the complete feature table, but can increase the size of a classifier that reads one node at a time. For every even $N\ge4$, we construct a labeled tree with $2N$ vertices and maximum degree three. A shared one-hidden-layer rectified linear unit (ReLU) classifier needs one hidden unit on the raw scalar features and exactly $N$ units after one lazy neighborhood mean. All aggregate values are distinct, and the mixing matrix has condition number at most 8. We determine the complete error curve for heads with a free affine term. The width separation persists under scalar perturbations. Under a specified bounded uniform-noise model with one nuisance coordinate, every one-unit aggregate head has expected error at least $1/12$, while a fixed $N$-unit head has zero error. Controlled fits compare trained heads with these explicit feasible scores. On Roman-empire and Tolokers, aggregation improves fitted prediction, and retaining raw features helps at the largest tested parameter budget. Training-only diagnostics include conditional output-layer optimality bounds for eight fixed runs.
Masked diffusion models train a denoiser on masked cross-entropy and generate by unmasking many positions in parallel. Existing MDM sampling bounds assume an exact denoiser or leave its error uncontrolled while training bounds cover only the masking distribution on which the model was trained. We present an end-to-end guarantee for a learned denoiser under conditional mixing (CM): stationary sources (e.g. floored order-$r$ Markov and two-sided-Doeblin hidden-Markov sources) for which revealing distant visible tokens moves a masked token's posterior on average exponentially little in the distance with correlation length $τ_c$. We show the empirical risk minimizer trained under any masking distribution attains per-token excess $\varepsilon'$ on that distribution from $\varepsilon'^{-\mathcal{O}(τ_c\log|\mathcal{V}|)}$ training sequences, independent of sequence length $N$ -- one sample budget serves every decoding schedule when each schedule's denoiser is trained under its own weighting. We further show that the exponent's order $Θ(τ_c\log|\mathcal{V}|)$ is minimax-optimal up to constants; more efficient rates require added structure. With this, we bound the end-to-end KL as $\mathrm{KL}(π\Vert p_{\mathrm{alg}})\leq \varepsilon+N\varepsilon'$ under a dilated schedule of $T=\mathcal{O}(τ_c\log(N|\mathcal{V}|/\varepsilon))$ denoising steps versus $Ω(N)$ steps for a fixed random order on an explicit CM source. Finally, we show that block-autoregressive samplers chain across blocks with no added error and that a $Λ$-looped order-$r$ message-passing class on order-$r$ sources achieves polynomial sample complexity, with approximation error improving per loop until saturation at $Λr\approx τ_c\log(1/\varepsilon')$. Experiments on released and retrained models support our results' qualitative predictions.