NVIDIA Introduces X-Token: Projection-Guided Cross-Tokenizer KD That Outperforms GOLD by +3.82 Average Points on Llama-3.2-1B

Knowledge distillation (KD) transfers “dark knowledge” from a large teacher model to a smaller student. The student learns from the teacher’s full output probability distribution over tokens, not just correct answers. This is done via per-position Kullback–Leibler (KL) divergence over next-token probability distributions.

This formulation requires a shared tokenizer. A practitioner committed to Llama-3.2-1B cannot leverage stronger teachers with incompatible tokenizers — such as Phi-4-mini or Qwen3-4B — because token positions do not correspond across vocabularies. This also prevents multi-teacher distillation across tokenizer families.

NVIDIA researchers introduced X-Token, a logit-distribution-based method for cross-tokenizer KD (Knowledge distillation). It operates as a drop-in replacement for the standard KD loss, requiring no auxiliary trainable components and no architectural changes.

The Problem X-Token is Solving

Two prior approaches dominate cross-tokenizer KD. ULD (Universal Logit Distillation) sidesteps vocabulary alignment by rank-sorting both distributions and minimizing L1 distance. It discards token identity entirely. GOLD adds span alignment and a hybrid loss. It partitions tokens into a 1-to-1 string-matched common subset, trained with KL divergence, and an uncommon remainder, trained with ULD-style rank matching. GOLD is the current state of the art.

The research team identifies two structural failures in GOLD’s design:

Failure 1: Uncommon-token failure– When tokenizers fragment text differently, critical tokens fall into the unmatched uncommon subset. Llama-3 packs multi-digit numbers as single tokens — “201” is one token. Qwen3 splits them digit by digit: “2”, “0”, “1”. Under GOLD, all 1,100 of Llama’s two- and three-digit numerals (100 two-digit, 1,000 three-digit) fall into the uncommon set when Qwen3-4B is the teacher. Those tokens receive two types of harmful signal: identity-agnostic noise from rank-based ULD matching, and suppressive gradients from the common-KL term acting through the full-vocabulary softmax. The result: GSM8k accuracy drops to 2.56 under GOLD with Qwen3-4B, compared to 12.89 for same-tokenizer KD from a weaker Llama-3.2-3B teacher.

Failure 2: Over-conservative matching– GOLD uses strict string equality to define the common subset. A student token Hundreds corresponds to teacher tokens Hund followed by reds under teacher-side re-tokenization, but strict matching discards this pair. Useful alignment signal is lost even when the correspondence is well-formed.

These two failures require opposite remedies: eliminate the partition when critical tokens are misaligned, and relax it when alignment is structurally sound.

How X-Token Works

X-Token has three components: span alignment, a projection matrix W, and two complementary loss formulations — P-KL and H-KL.

Span Alignment

Teacher and student tokenizers produce sequences of different lengths for the same text. X-Token uses dynamic-programming (DP) span alignment, grouping tokens into chunks where each chunk-pair decodes to the same underlying text substring. A chain-rule merge then combines per-token probabilities within each chunk into a single chunk-level distribution for use in the distillation loss. The alignment is cached per sequence and adds no per-step training overhead.

The research team also identifies a failure in TRL’s surface-substring alignment, which is used in TRL’s GOLD trainer. TRL accumulates per-side decoded buffers and flushes only when both buffers match as equal raw strings. A byte-level disagreement — such as Llama-3 auto-prepending <bos> while Qwen-3 does not — prevents future flushes and forces all remaining tokens into one mis-grouped super-group at end of sequence. The DP approach handles this with a single gap move, regardless of sequence length.

The Projection Matrix W

After alignment, teacher and student distributions still operate over different vocabularies. The projection matrix W ∈ ℝ^V_S|×|V_T| maps each student token to a weighted combination of teacher tokens, bridging the vocabulary mismatch.

W is constructed deterministically in two passes:

Pass 1 (exact-match): For every (student token, teacher token) pair whose decoded strings match after canonicalization, set W[s, t] = 1. Canonicalization unifies space prefixes (Ġ, _, ␣), newlines, byte-fallback tokens of the form <0xHH>, and model-specific special tokens across tokenizer families.

Pass 2 (multi-token rule): For each student token without an exact match, re-tokenize its decoded text under the teacher tokenizer. If the resulting sequence has length ≤ 4, assign exponentially-decayed weights: W[s, τᵢ] = β·γⁱ with (β, γ) = (0.9, 0.1). A length-2 span receives normalized weights (0.909, 0.091). A length-3 span receives (0.9009, 0.0901, 0.0090). A length-4 span receives (0.9000, 0.0900, 0.0090, 0.0009). The leading sub-token receives the highest weight because it typically carries the most informative probability mass — for example, “_inter” in [“_inter”, “national”] or “_20” in [“_20”, “24”].

Each row is truncated to its top-4 entries and row-normalized. Because each row of W is non-negative and sums to 1, left-multiplication by W⊤ is probability-preserving: if p_S is a probability vector, W^⊤p_S is also a valid probability vector over V_T. W is constructed once before training and can optionally be jointly refined with the student under P-KL.

P-KL: Addressing Erroneous and Suppressive Gradients

P-KL removes the partition entirely. It projects the student distribution p̂_S^(k) into teacher vocabulary space via W:

$\tilde{p}_S^{(k)}[t] = \sum_{s\in\mathcal{V}_S} W[s, t] \cdot \hat{p}_S^{(k)}[s]$

Then it computes KL divergence directly between teacher and projected student:

$\frac{\partial\mathcal{L}_{common}}{\partial z_{j}} = p_S[j] \cdot M_{\mathcal{C}}(T)$

There is no uncommon set, so rank-based ULD noise is eliminated. The suppressive gradient problem is also eliminated: the projection routes the student’s probability mass for “201” directly onto {2, 0, 1} in the teacher vocabulary via W.

The research team formally proves (Proposition 1) that GOLD’s common-KL term induces non-negative gradients on every uncommon student logit. The gradient on an uncommon student logit j is: ∂ℒ_common/∂z_j = p_S[j] · M_C(T), where M_C(T), is the teacher probability mass on the common subset. Under gradient descent, this always drives z_j downward — suppressing every uncommon token’s probability regardless of the ground-truth token.

H-KL: Relaxing the 1-to-1 Matching

H-KL applies when the partition is structurally sound — that is, when critical tokens land in the common subset. In that case, GOLD’s direct KL on identity-aligned pairs delivers sharper per-pair supervision than P-KL’s projection, which blends student probability mass across multiple teacher tokens. The opportunity is to make the partition less wasteful by relaxing the strict string-equality criterion.

H-KL retains GOLD’s hybrid loss structure but expands the common set C using W. For each student token s, it selects the top-ranked teacher token t* = argmax_{t’∈V_T} W[s, t’], and adds (s, t*) to C. Exact matches are preserved since they receive weight 1 in W, the highest possible. Near-equivalent pairs like (Hundreds, Hund) — excluded by GOLD — are now admitted. The expanded C feeds the same hybrid loss: direct KL on common pairs, ULD on the remainder.

Selecting Between P-KL and H-KL

The selection uses a coverage audit over token categories in the student vocabulary. For math tasks, multi-digit numerals are the critical category. Table 8 in the research paper shows: under Qwen3-4B, 0 out of 100 two-digit Llama numerals and 0 out of 1,000 three-digit Llama numerals appear in C. Under Phi-4-mini-Instruct, all 100 two-digit and all 1,000 three-digit numerals appear in C. ASCII punctuation and single-digit numerals are fully covered in both cases.

The rule: use P-KL when critical tokens fall outside C (Qwen3-4B), and H-KL when the partition is sound (Phi-4-mini-Instruct). Table 2 in the research paper shows the mode reversal is sharp: P-KL outperforms H-KL by +3.55 avg. on Qwen3-4B, while H-KL outperforms P-KL by +1.68 avg. on Phi-4-mini.

Multi-Teacher Distillation

X-Token extends to multiple teachers. Each teacher has its own projection matrix W_m and loss selection. For same-tokenizer teachers, standard token-level KL is used. The multi-teacher loss aggregates per-teacher losses with weights α_m:

$\mathcal{L}_{KD,multi} = \sum_{m=1}^{M}\alpha_{m}\frac{1}{|\mathcal{K}_{m}|}\sum_{k\in\mathcal{K}_{m}}\mathcal{L}_{*,m}^{(k)}$

The research team evaluates static and confidence-adaptive weighting schemes. Confidence-adaptive variants compute α_m from cross-entropy, Shannon entropy, or maximum predicted probability of the teacher’s distribution. Static weighting outperforms adaptive schemes in both multi-teacher setups evaluated.

Dynamic KD/CE Scaling

Training combines the distillation loss ℒ_KD with next-token cross-entropy ℒ_CE. Because these terms differ in magnitude and shift during training, X-Token rescales the KD term at each step to match the scale of ℒ_CE:

$\mathcal{L} = \text{sg}(\mathcal{L}_{CE} / \mathcal{L}_{KD}) \cdot \mathcal{L}_{KD} + \mathcal{L}_{CE}$

where sg(·) is stop-gradient. Table 4 in the paper shows dynamic scaling outperforms three fixed-weight settings (KD-heavy, balanced, CE-heavy) on the Qwen3-4B (P-KL) pair.

Experiments and Results

Student: Llama-3.2-1B. Teachers: Llama-3.2-3B (same tokenizer), Qwen3-4B, and Phi-4-mini-Instruct. Training data: NemotronClimbMix dataset, 30,000 steps, batch size 768, context length 4096. Optimizer: AdamW, learning rate 5×10⁻⁵, 5% warmup with cosine decay, weight decay 0.1, gradient clipping 1.0. Each experiment is feasible on a single NVIDIA H100 GPU; the research team used 128 H100s to speed up iteration.

Evaluation: 3-shot accuracy on MMLU, GSM8k, MATH-Hendrycks, Winogrande, and HellaSwag.

Key results:

On Qwen-4B (P-KL regime): GOLD reaches 35.03 avg., below even continued pre-training without a teacher (36.63). This confirms the partition is actively harmful when critical tokens are misaligned. Pure ULD (36.77) already improves over GOLD, indicating the partition is the primary failure source. P-KL further improves to 38.85 avg. (+3.82 over GOLD). GSM8k alone moves from 2.56 to 15.54, surpassing same-tokenizer KD from Llama-3.2-3B (12.89) on that benchmark.

On Phi-mini (H-KL regime): GOLD reaches 38.66 avg. — a reasonable baseline where the partition is structurally sound. H-KL improves to 39.18 avg. (+0.52 over GOLD). P-KL applied to Phi-mini drops to 37.50 avg., confirming that the wrong loss mode hurts even when W is available.

Multi-teacher: Phi-mini (H-KL, α=0.8) + Llama-3B (standard KL, α=0.2) under static weighting reaches 40.48 avg. This is +2.08 over same-family KD from Llama-3B alone, and +1.30 over the best single cross-tokenizer result (39.18). Combining Phi-mini + Qwen-4B — two teachers with overlapping reasoning strengths — scores only 38.49, below the best single teacher. Adding Qwen-4B as a third teacher yields 40.15, with math/reasoning degrading (GSM8k 20.39 → 19.18) while commonsense improves slightly. Teacher complementarity, not teacher count, drives gains.

Strengths and What to Watch

Strengths:

• The suppressive gradient problem in GOLD’s hybrid loss is formally proved (Proposition 1), not just observed empirically
• W is constructed rule-based from tokenizer strings alone; no training data or learned parameters needed at initialization
• Dynamic KD/CE scaling removes the need to tune fixed loss weights; it outperforms three fixed-weight baselines in ablations
• Multi-teacher extension adds no architectural changes; each teacher uses its own W_m and appropriate loss
• The coverage audit for P-KL vs H-KL selection is a defined, reproducible criterion based on per-category token retention in C

What to Watch:

• Experiments use only Llama-3.2-1B as the student under continued pre-training; larger students and instruction-tuned settings are not evaluated
• Only three teacher pairs are tested; low-overlap tokenizer families (SentencePiece, byte-level BPE) are left for future work
• Static weighting outperforms confidence-adaptive weighting in all tested multi-teacher setups, but why?
• The multi-token rule in Pass 2 skips student tokens whose decoded text re-tokenizes to sequences longer than 4 under the teacher; those rows remain zero in W

Marktechpost’s Visual Explainer

Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Email Name

Email *

Submit