NVIDIA 推出 X-Token：投影引导的跨Tokenizer知识蒸馏，在 Llama-3.2-1B 上平均分超越 GOLD +3.82

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:

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

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:

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:

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

■ X-Token — NVIDIA Research

01 — Background

What is Knowledge Distillation?

Knowledge distillation (KD) transfers “dark knowledge” from a large teacher model to a smaller student model. The student learns from the teacher’s full next-token probability distribution, not just the correct answer.

This is done via per-position KL divergence over the teacher’s output distribution at every token position in the sequence.

The constraint: standard KD requires a shared tokenizer. If Llama-3.2-1B is the student, it cannot learn from Qwen3-4B or Phi-4-mini — their token vocabularies do not align. Token positions have no correspondence across different tokenizer families.

Llama

Student tokenizer

Qwen / Phi

Incompatible teachers

≠ Match

Vocab mismatch

02 — The Problem

Two Structural Failures in GOLD

GOLD is the prior state-of-the-art cross-tokenizer KD method. It partitions tokens into a string-matched common subset (trained with KL) and an uncommon remainder (trained with ULD rank-matching).

NVIDIA researchers identified two distinct failures:

Uncommon-token failure:

Critical tokens fall into the unmatched subset. Llama packs “201” as one token. Qwen splits it into “2”, “0”, “1”. All 1,100 multi-digit Llama numerals fall into the uncommon set under Qwen3-4B. They receive identity-agnostic noise and suppressive gradients — GSM8k drops to

Over-conservative matching:

Strict string equality discards well-formed pairs. Student token

Hundreds

maps to teacher tokens

Hund

reds

, but GOLD drops this alignment entirely.

03 — Solution

X-Token: Three Core Components

X-Token is a logit-distribution-based cross-tokenizer KD method. It requires no auxiliary trainable components and no architectural changes — it is a drop-in replacement for the standard KD loss.

Span Alignment:

DP-based alignment groups tokens into chunks that decode to the same text substring. Cached per sequence — zero per-step overhead.

Projection Matrix W:

A sparse matrix W ∈ ℝ⁼|V_S|×|V_T|⁽ maps each student token to a weighted combination of teacher tokens, bridging the vocabulary gap.

Two Loss Modes:

P-KL

removes the partition entirely.

H-KL

retains the partition but relaxes matching via top-1 mappings under W. Each targets a different failure mode.

04 — Projection Matrix W

How W is Constructed

W is built deterministically before training in two passes. No training data or learned parameters are required at initialization.

Exact-match pass:

For every (student, teacher) token pair whose decoded strings match after canonicalization, set W[s,t] = 1. Canonicalization unifies space prefixes, newlines, byte-fallback tokens, and special tokens across families.

Multi-token rule pass:

For unmatched student tokens, re-tokenize their decoded text under the teacher. Assign decayed weights W[s,τᵢ] = β·γⁱ with (β,γ) = (0.9, 0.1). A 2-token span gets (0.909, 0.091). Each row is truncated to top-4 entries and row-normalized.

Because each row sums to 1, Wᵀ is probability-preserving: Wᵀp_S is a valid probability vector over V_T without additional normalization.

05 — Loss Formulations

P-KL vs H-KL: When to Use Each

Selection is based on a coverage audit: measure what fraction of critical token categories (e.g. multi-digit numerals) appear in the common set C.

Applying the wrong mode reverses results: P-KL on Phi-mini drops to 37.50 avg. vs H-KL’s 39.18.

06 — Results

Benchmark Results on Llama-3.2-1B (3-shot)

Student: Llama-3.2-1B — trained on NemotronClimbMix, 30K steps, batch 768, context 4096.

07 — Multi-Teacher Distillation

Teacher Complementarity Drives Gains

X-Token extends to multiple teachers. Each gets its own projection matrix W_m and loss mode. The aggregated loss uses per-teacher weights α_m.

Key finding: static weighting outperforms confidence-adaptive weighting in all tested setups. Phi-mini (α=0.8) + Llama-3B (α=0.2) achieves the best result.

Combining two reasoning-heavy teachers (Phi-mini + Qwen-4B) scores below the best single teacher. Teacher diversity matters more than teacher count.

08 — Key Takeaways

What to Remember About X-Token

GOLD’s partition actively harms training when critical tokens (e.g., multi-digit numerals) fall into the uncommon set —

P-KL eliminates the partition entirely

using projection matrix W.

H-KL

retains the partition but relaxes matching to top-1 mappings under W — best when the partition is structurally sound.

The projection matrix W is

built rule-based before training

from tokenizer strings alone; no learned parameters required at init.

Multi-teacher gains (+1.3 over single-teacher) come from

teacher complementarity

, not from adding more teachers with overlapping strengths.

GSM8k recovers from

(GOLD) to

(P-KL) — a 6× gain that exceeds same-tokenizer KD from a stronger Llama-3.2-3B teacher.

arXiv: 2605.21699  —  Institution: NVIDIA

NVIDIA Research — X-Token: Projection-Guided Cross-Tokenizer Knowledge Distillation — arXiv:2605.21699 — marktechpost.com

Key Takeaways

• X-Token identifies two distinct, opposite failure modes in GOLD: uncommon-token suppression (fix: remove the partition with P-KL) and over-conservative matching (fix: relax it with H-KL).
• The projection matrix W is built rule-based from tokenizer strings before training; it can optionally be jointly refined with the student for additional gains.
• P-KL on Qwen3-4B improves over GOLD by +3.82 avg. and recovers GSM8k from 2.56 to 15.54.
• Multi-teacher distillation gains (+1.3 over single-teacher) come from teacher complementarity, not just from adding more teachers.
• Loss mode selection (P-KL vs H-KL) is determined by a coverage audit on token categories; applying the wrong mode reverses the ranking.