Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities

Hyunsik Kim; Youngmoon Jung

cs.CLOct 1, 2026

Universal Byte-Level Encoding: UTF-8/UTF-16 Routing to Reduce Cross-Script Token-Budget Disparities

Authors: Hyunsik Kim, Youngmoon Jung

Organizations: Samsung Research

Abstract

Byte-level byte-pair encoding (BBPE) tokenizers are attractive for multilingual large language models (LLMs) because they cover all Unicode text. In UTF-8-based BBPE, however, many scripts start from a higher fallback cost than English: when no learned merges can be applied, a multibyte character requires multiple byte-derived symbols. We call this worst-case pre-merge cost the encoding floor. A higher floor can increase token counts and per-request cost and shrink usable context. Changing the text encoding can reduce this gap, but a single global encoding can make already-efficient English spans more expensive in mixed-script text. We propose Universal Byte-Level Encoding (UBE), a dual-alphabet tokenizer that keeps 1-2-byte UTF-8 characters on the UTF-8 path while routing 3-4-byte UTF-8 characters through UTF-16. This lowers the encoding floor for 3-byte Basic Multilingual Plane (BMP) characters in scripts with high token premiums (token counts relative to English) without raising it for already-efficient spans in mixed-script text. UBE changes only the byte representation presented to byte-pair encoding (BPE); the merge rule remains standard, and exact decoding is preserved. UBE also composes with alternative boundary policies and morphology-based representations. In a Unicode 17 audit, UBE exactly round-trips all Unicode scalar values and all inputs in the official normalization, grapheme-break, and emoji test suites. Across intrinsic evaluations, UBE lowers dispersion in English-normalized token-count ratios, reducing cross-lingual token-budget disparity. In multilingual language model (LM) experiments, UBE matches BBPE's LM quality. In the main multilingual settings, UBE reduces token counts most for high-premium scripts and slightly lowers English token counts, yielding more usable context under fixed token budgets and faster prompt processing in content-matched benchmarks.

Figures & tables

	ASCII	Cyrillic/Arabic/Hebrew	CJK, Devanagari, Thai	Non-BMP
BBPE	1	2	3	4
BBPE16	2	2	2	4
UBE	1	2	2	4

Table 1: Encoding-floor values before BPE merges. BBPE16 ( Kim et al., 2026 ) is a UTF-16-only baseline; bold marks UBE’s target trade-off: ASCII stays at the 1-byte UTF-8 floor, while 3-byte BMP scripts (e.g., CJK, Devanagari, and Thai) fall from 3 to 2 via the UTF-16 Private Use Area (PUA) path; 4-byte non-BMP code points remain at a 4-symbol floor. CJK stands for Chinese, Japanese, and Korean.

Figure 1: Pre-merge base-symbol decomposition of “GPU性能”. BBPE, BBPE16, and UBE use 9, 10, and 7 base symbols, respectively. PUA denotes the Private Use Area; Appendix A gives the full algorithm.

Panel of Figure 2 — Figure 2: Token-premium Gini vs. total FLORES-200 token count ( 101-lang ; 16K–256K vocabularies). S-BBPE and S-UBE apply SuperBPE-style boundary relaxation to BBPE and UBE, respectively. Public LLM tokenizers are reference points (Appendix Table H.2 ); lower-left is better.

Dataset	Eval. units	BBPE Gini	UBE Gini	$\Delta$ Gini	Note
FLORES-200	204	0.2183	0.1979	-9.4%	parallel sentences
UDHR	387	0.2499	0.2221	-11.1%	public-text stress test
SIB-200	205	0.2297	0.2102	-8.5%	topic classification
MGSM	11	0.1841	0.1843	+0.1%	math reasoning

Table 2: Cross-dataset token-premium Gini ( 101-lang/32K ). UDHR is a public-text stress test with prefix overlaps. The full 16K–256K sweep is in Appendix Table D.12 .

Language	Script	BBPE	UBE	$\Delta$ %
Santali	Ol Chiki	341.287	222.944	-34.7%
Central Atlas Tamazight	Tifinagh	269.166	183.305	-31.9%
Tibetan	Tibetan	336.804	286.737	-14.9%
Arabic	Arabic	60.600	60.407	-0.3%
English	Latin	31.160	31.068	-0.3%

Table 3: Representative per-language token counts on FLORES-200 ( 101-lang/32K ; mean tokens per sentence). Other vocabulary sizes appear in Appendix Table C.3 .

Config	Tok	BPB $\downarrow$	BPB Gini $\downarrow$	XNLI	Bele	XCOPA	ARC-E	ARC-C	HSwag	PIQA	LAMB
En-only	BBPE	1.4909	0.1993	34.3	23.6	51.4	38.1	18.3	27.5	61.9	23.6
En-only	UBE	1.4698	0.1989	34.4	23.0	51.5	38.5	17.0	27.8	62.7	22.5
6-lang	BBPE	1.3680	0.1930	34.6	23.5	52.0	35.6	17.1	27.4	61.3	23.1
6-lang	UBE	1.3760	0.1880	34.9	23.2	52.2	36.7	18.4	28.0	62.0	23.1
101-lang	BBPE	1.5174	0.1762	34.3	24.8	52.0	35.5	17.0	27.7	59.3	22.8
101-lang	UBE	1.5237	0.1739	34.4	23.7	51.8	35.9	17.8	27.9	60.1	23.1

Table 4: LM evaluation (182M, 32K vocabulary, 200K mC4 steps; primary seed; multi-seed in Appendix F.2 ) across three tokenizer-training configurations. Every LM is trained on the same 101-language mC4 corpus; En-only/6-lang/101-lang identify only the corpus used to learn the tokenizer vocabulary. Bits per byte (BPB) and BPB Gini use 101 held-out mC4 language splits; benchmark columns report zero-shot accuracy (Bele = Belebele, HSwag = HellaSwag, LAMB = LAMBADA).

Fixed component	English	Selected BMP (12)	Overall (204)
Standard	31.160 $\to$ 31.068	178.499 $\to$ 145.838	71.889 $\to$ 69.857
SuperBPE-style	33.143 $\to$ 32.973	161.463 $\to$ 130.218	74.187 $\to$ 72.137
SCRIPT-style	31.698 $\to$ 31.652	184.743 $\to$ 150.700	69.957 $\to$ 68.001
MYTE maps	109.140 $\to$ 109.094	255.737 $\to$ 183.867	148.142 $\to$ 141.354

Table 5: Composition with UBE on FLORES-200 (primary 101-language/32K BPE vocabularies). Values are mean BPE tokens (BBPE

\to

UBE) or base symbols (MYTE

\to

MYTE-UBE) per sentence. Selected BMP denotes the 12 configurations defined in Table D.3 .

Method	Common-text BPB $\downarrow$	BPB Gini $\downarrow$	Multi. mean $\uparrow$	English mean $\uparrow$
UBE (ours)	1.5643	0.1937	36.8	33.0
BBPE	1.5908	0.2006	36.6	32.7
BBPE16	1.5561	0.1996	36.3	32.5
MYTE+BPE (derived 32K)	1.6474	0.2029	36.5	32.2
SCRIPT-BPE	1.5401	0.1951	36.4	32.9

Table 6: Language-model comparison across five tokenizers (182M parameters; 32K vocabularies trained on approximately 50 MiB/language). Single-seed runs match 283.36B supervised UTF-8 bytes. Group means weight benchmarks equally. BPB for MYTE+BPE and SCRIPT-BPE is representation-conditioned; see Appendix E.5 for details.

Appendix figures & tables51 assets

Supplementary material from the paper’s appendix.

Appendix

Vocab size	BBPE merges	UBE merges	UBE overhead
16K	15,739	15,483	1.6%
32K	31,739	31,483	0.8%
64K	63,739	63,483	0.4%
128K	127,739	127,483	0.2%
256K	255,739	255,483	0.1%

Appendix

Table B.1: Merge-slot accounting by vocabulary size. All rows reserve the same five special tokens; UBE spends an additional 256 slots on its disjoint UTF-16 alphabet relative to BBPE.

Dataset	Original Gini	Det. Gini	$\Delta$ Gini	$\Delta$ mean premium
FLORES-200	0.2522	0.2517	-0.00050	-0.0028
UDHR	0.2473	0.2471	-0.00025	+0.0003
SIB-200	0.2600	0.2595	-0.00050	-0.0025
MGSM	0.3120	0.3120	+0.00002	-0.0004

Appendix

Table B.2: Deterministic tie-break sensitivity check for UBE 6-lang/32K . “Det.” is the tokenizer trained with deterministic secondary ordering for equal-frequency merge candidates.

\Delta

columns report Det. minus the original UBE tokenizer.

Setting	Runs	Layers/heads/width	Params	Batch recipe	Global tok/step	Role
182M	6	14/12/768	182M	32 $\times$ 1	524,288	full coverage matrix (200K steps, 105B tok)
329M	4	14/12/768	329M	8 $\times$ 4	524,288	128K composition check (200K steps, 105B tok)
1.3B	6	24/16/2048	1.35B actual	8 $\times$ 1	131,072	scale verification (205K steps, $\sim$ 26.9B tok)

Appendix

Table B.3: Reported core LM suite on 8

\times

A100 80 GB GPUs. “Batch recipe” is per-device microbatch

\times

gradient-accumulation steps; all rows use sequence length 2048. The 182M and 1.3B rows form the primary two-scale BBPE/UBE matrix; the 329M row is a separate 128K composition check for S-BBPE and S-UBE.

Tokenizer coverage	BBPE train	UBE train	BBPE eval	UBE eval	Steps/epoch	200K coverage
En-only on 101-lang mC4	$\sim$ 184.0B	$\sim$ 166.9B	$\sim$ 7.5B	$\sim$ 7.2B	350,991 / 318,320	57.0% / 62.8%
6-lang on 101-lang mC4	$\sim$ 120.0B	$\sim$ 119.0B	$\sim$ 7.0B	$\sim$ 6.9B	228,967 / 227,033	87.3% / 88.1%
101-lang on 101-lang mC4	$\sim$ 108.1B	$\sim$ 107.8B	$\sim$ 5.3B	$\sim$ 5.3B	206,142 / 205,695	97.0% / 97.2%

Appendix

Table B.4: Estimated 32K LM token accounting under the shared 101-language mC4 release. Train counts use one-shard-per-language byte-weighted extrapolation; evaluation counts enumerate held-out terminal shards. Steps/epoch are train-token totals divided by 524,288 global tokens/step, and fixed 200K steps correspond to 104.9B training tokens.

Config	Langs	Train			Eval
Config	Langs	Shards	GiB	Docs	Shards	Docs
En-only	1	118	101	32.1M	1	0.27M
6-lang	6	1,330	253	62.3M	6	0.46M
101-lang	101	2,273	346	89.3M	101	4.36M

Appendix

Table B.5: Tokenizer-training corpus size by configuration. Rows aggregate the language subset used to train each tokenizer; the LM corpus always uses the full 101-language mC4 release. Evaluation columns enumerate held-out terminal shards, document counts are byte-weighted one-shard estimates, and raw byte totals are exact Arrow file sizes.

Language	Region	Speakers	BBPE	UBE	$\Delta$ %	Script
Amharic	Ethiopia	$\sim$ 57M	202.775	137.539	-32.2%	Ethiopic
Lao	Laos	$\sim$ 30M	342.100	233.445	-31.8%	Lao
Santali	India	$\sim$ 8M	341.287	222.707	-34.7%	Ol Chiki
Telugu	India	$\sim$ 82M	237.530	220.807	-7.0%	Telugu
Kannada	India	$\sim$ 44M	251.956	237.506	-5.7%	Kannada
Tibetan	China/Nepal	$\sim$ 6M	264.246	247.286	-6.4%	Tibetan

Appendix

Table C.1: Per-language UBE impact on FLORES-200 in the 6-lang/32K stress setting (mean tokens per sentence). The listed non-English languages are outside tokenizer training; UBE cuts token counts by 6–35%, equivalent to 6–53% more fixed-budget content. The 16K and 64K results for the same languages appear in Table C.2 .

Language	Script	16K			32K			64K
Language	Script	BBPE	UBE	$\Delta$ %	BBPE	UBE	$\Delta$ %	BBPE	UBE	$\Delta$ %
Amharic	Ethiopic	219.321	140.388	-36.0%	202.775	137.539	-32.2%	163.896	134.837	-17.7%
Lao	Lao	343.176	234.068	-31.8%	342.100	233.445	-31.8%	256.355	231.798	-9.6%
Santali	Ol Chiki	363.850	230.685	-36.6%	341.287	222.707	-34.7%	341.274	213.132	-37.5%
Telugu	Telugu	263.163	224.362	-14.7%	237.530	220.807	-7.0%	184.658	181.593	-1.7%
Kannada	Kannada	354.337	238.641	-32.7%	251.956	237.506	-5.7%	220.944	219.524	-0.6%

Appendix

Table C.2: Per-language UBE impact on FLORES-200 across the 6-lang vocabulary sweep (16K/32K/64K; mean tokens per sentence). The same six representative non-English languages as Table C.1 are shown.

\Delta

% is computed as

(\mathrm{UBE}-\mathrm{BBPE})/\mathrm{BBPE}

Language	Script	Vocab	BBPE	UBE	$\Delta$ %
Santali	Ol Chiki	16K	341.312	237.326	-30.467%
		32K	341.287	222.944	-34.676%
		64K	341.271	212.692	-37.677%
		128K	251.125	211.712	-15.695%
		256K	228.583	207.379	-9.276%
Central Atlas Tamazight	Tifinagh	16K	269.200	192.112	-28.636%

Appendix

Table C.3: Representative per-language UBE impact on FLORES-200 across 101-lang vocabulary sizes (average tokens per sentence). Extends Table 3 .

Evaluation suite	Eval sentences	Exact overlaps	Prefix overlaps
FLORES-200	206,448	0	53
UDHR	35,870	1	7,102
SIB-200	41,820	0	10
MGSM	2,750	0	0

Appendix

Table C.4: Surface-form contamination audit against the mC4 scan corpus. Prefix overlaps use the 50-character prefix-substring check described in the text.

Config	Vocab	UBE wins/ties/losses	$\Delta$ Gini
En-only	16K	37 / 19 / 148	-13.9%
En-only	32K	41 / 23 / 140	-11.9%
En-only	64K	37 / 82 / 85	-7.3%
6-lang	16K	202 / 0 / 2	-8.7%
6-lang	32K	191 / 0 / 13	-7.2%
6-lang	64K	169 / 24 / 11	-6.2%

Appendix

Table D.1: FLORES-200 language-level outcome summary across configurations (204 languages). Columns show languages where UBE produces fewer tokens (wins), ties, and more tokens (losses), plus the corresponding Gini change. The 128K and 256K rows are only run for 101-lang .

Vocab	BBPE Gini	UBE Gini	$\Delta$ Gini	UBE wins
16K	0.2194	0.1927	-12.1%	196/204
32K	0.2183	0.1979	-9.4%	189/204
64K	0.2225	0.2043	-8.2%	187/204
128K	0.2258	0.2140	-5.2%	174/204
256K	0.2246	0.2204	-1.9%	150/204

Appendix

Table D.2: FLORES-200 Gini for BBPE and UBE under the primary 101-lang training regime. Table D.3 gives the token-count breakdown.

Tokenizer	EN avg	Latin avg	Selected BMP (12)	All avg	Gini
BBPE 16K	34.863	62.373	200.188	79.408	0.21938
BBPE 32K	31.160	56.743	178.499	71.889	0.21833
BBPE 64K	28.716	52.540	161.576	66.056	0.22254
BBPE 128K	27.174	48.751	140.748	60.594	0.22576
BBPE 256K	26.272	45.100	115.565	55.389	0.22456
UBE 16K	34.671	62.163	153.261	76.410	0.19274

Appendix

Table D.3: Gini coefficient and groupwise mean tokens per sentence on FLORES-200 ( 101-lang ; 16K–256K). Latin avg covers Latin-script languages; Selected BMP denotes 12 language–script configurations covering CJK, Tibetan, Ol Chiki, Tifinagh, and Myanmar. † BBPE16 lowers Gini via a uniform 2-byte floor that raises Latin token counts; SuperBPE compositions are reported separately.

Panel of Figure 21 — Figure D.1: Token premium (ratio to English) vs. vocabulary size for six script families under UBE ( 101-lang ).

Tokenizer	EN avg	Latin avg	Selected BMP (12)	All avg	Gini
SuperBPE BBPE 16K	38.308	66.840	180.079	82.853	0.19967
SuperBPE BBPE 32K	33.143	61.177	161.463	74.187	0.19217
SuperBPE BBPE 64K	29.133	55.586	151.281	66.871	0.18996
SuperBPE BBPE 128K	26.337	51.476	140.300	61.045	0.19042
SuperBPE BBPE 256K	24.357	47.479	122.457	55.220	0.18694
SuperBPE UBE 16K	38.099	66.534	149.793	80.676	0.18269

Appendix

Table D.4: SuperBPE composition results on FLORES-200 (204 evaluation languages; tokenizers trained under the 101-lang mC4 regime; 16K–256K vocab). Format matches Table D.3 .

Panel of Figure 23 — Figure D.2: Gini coefficient vs. vocabulary size for full-data (346 GiB) and half-data ( $\sim$ 173 GiB) tokenizer training ( 101-lang ). Solid markers denote full data and open markers denote half data.

Top 10: UBE helps most
Language	BBPE avg	UBE avg	BBPE16 (128K) avg	UBE/BBPE
Santali	341.271	212.692	222.620	0.6232
Central Atlas Tamazight	269.123	182.243	183.143	0.6772
Tamashek (Tifinagh)	270.571	185.609	187.934	0.6860
Dzongkha	286.182	265.685	267.341	0.9284
Tibetan	264.208	247.171	247.411	0.9355

Appendix

Table D.5: Best and worst 10 languages by UBE/BBPE token ratio on FLORES-200 ( 101-lang ; BBPE/UBE: 64K, BBPE16: 128K). Ratio

<1

means UBE uses fewer tokens; ratio

>1

means UBE uses more. BBPE16 is included as a cross-substrate reference.

Panel of Figure 25 — Figure D.3: Best and worst five languages by UBE/BBPE token-count ratio across the full 101-lang vocabulary sweep (16K–256K). Bars start at the vertical line, which marks equal token counts (ratio 1); red bars mark ratios below 1 (UBE uses fewer tokens) and blue bars ratios of 1 or above.

Script	Language	Codepoint	Floor (bytes/char)		Avg tokens/sent		Save
			UTF-8	UTF-16	BBPE	UBE
Ol Chiki	Santali	U+1C5E	3	2	341.3	212.7	37.7%
Tifinagh	Central Atlas Tamazight	U+2D30	3	2	269.1	182.2	32.3%
Tibetan	Dzongkha	U+0F42	3	2	286.2	265.7	7.2%

Appendix

Table D.6: Token composition for the scripts with the largest BBPE-to-UBE token-count gaps in the 101-lang /64K setting. Per-script-family token counts across all 16K–256K vocabularies appear in Table D.9 .

Script	Word	Chars	BBPE		UBE		Save
			Base	Tokens	Base	Tokens
Ol Chiki	ona	3	9	9	6	6	33%
Ol Chiki	kemikal	7	21	21	14	12	43%
Tifinagh	sin	3	9	9	6	6	33%
Tifinagh	acukula	7	21	21	14	14	33%
Ol Chiki	(24 chars)	24	66	63	45	39	38%

Appendix

Table D.7: Word- and sentence-level tokenization examples in the 101-lang /64K setting. “Base” and “Tokens” count pre-merge symbols and final BPE tokens. “Word” gives a romanized form; “( n chars)” denotes sentence fragments. Table D.8 extends the same examples to all 16K–256K vocabularies.

Script	Word	Chars	16K		32K		64K		128K		256K
Script	Word	Chars	BBPE	UBE	BBPE	UBE	BBPE	UBE	BBPE	UBE	BBPE	UBE
Ol Chiki	ona	3	9	6	9	6	9	6	6	6	6	6
Ol Chiki	kemikal	7	21	14	21	13	21	12	14	12	14	10
Tifinagh	sin	3	9	6	9	6	9	6	9	6	6	6
Tifinagh	acukula	7	21	14	21	14	21	14	21	14	14	14
Ol Chiki	(24 chars)	24	63	42	63	40	63	39	45	39	42	38

Appendix

Table D.8: Word- and sentence-level token counts across the 16K–256K vocabulary sweep ( 101-lang ), using the same inputs as Table D.7 .

Script family	Langs	UTF-8	Metric	16K	32K	64K	128K	256K
Latin	127	1–2	BBPE	62.373	56.743	52.540	48.751	45.100
			UBE	62.163	56.627	52.500	48.722	45.097
			$\Delta$	$-0.210$	$-0.117$	$-0.040$	$-0.029$	$-0.003$
Arabic	22	2	BBPE	78.312	69.687	62.817	56.489	51.358
			UBE	77.610	69.549	62.741	56.453	51.336
			$\Delta$	$-0.702$	$-0.138$	$-0.076$	$-0.036$	$-0.022$

Appendix

Table D.9: Per-script-family average token counts on FLORES-200 ( 101-lang ; 16K–256K; 29 scripts).

\Delta=\mathrm{UBE}-\mathrm{BBPE}

(negative = fewer UBE tokens).

Metric	BBPE	UBE	S-BBPE	S-UBE
Alphabet tokens	256	512	256	512
Merge tokens	31,739	31,483	31,739	31,483
Dead merges	13	9	524	531
Active utilization	99.9%	99.8%	98.3%	98.2%
Pure PUA–PUA merges	—	23.6%	—	24.6%
GPT-2–GPT-2 merges	—	66.6%	—	64.6%

Appendix

Table D.10: Vocabulary utilization diagnostics ( 101-lang/32K ). Other vocabulary sizes are in Table D.11 .

Family	Metric	16K	32K	64K	128K	256K
BBPE	Dead merges	9	13	20	82	224
	Active util.	99.825%	99.900%	99.939%	99.921%	99.905%
	Depth-1	12.326%	7.889%	4.937%	2.982%	1.774%
UBE	Dead merges	6	9	16	78	217
	Active util.	99.713%	99.847%	99.911%	99.907%	99.899%
	Pure PUA	24.758%	23.587%	22.297%	20.924%	19.641%

Appendix

Table D.11: Vocabulary utilization diagnostics across vocabulary sizes ( 101-lang ; 16K–256K), extending Table D.10 .

Dataset	Vocab	BBPE Gini	UBE Gini	$\Delta$ Gini
FLORES-200	16K	0.2194	0.1927	-12.14%
FLORES-200	32K	0.2183	0.1979	-9.36%
FLORES-200	64K	0.2225	0.2043	-8.20%
FLORES-200	128K	0.2258	0.2140	-5.20%
FLORES-200	256K	0.2246	0.2204	-1.86%
UDHR	16K	0.2449	0.2158	-11.86%

Appendix

Table D.12: Cross-dataset Gini comparison of BBPE and UBE across vocabulary sizes ( 101-lang ; 16K–256K).

Corpus	Vocab	Wilcoxon $p$	Holm $p$	BH $q$	Gini BBPE $\to$ UBE	$\Delta$ Gini 95% CI
En-only	16K	$9.21\!\times\!10^{-6}$	$2.76\!\times\!10^{-5}$	$1.18\!\times\!10^{-5}$	0.3495 $\to$ 0.3010	[-0.0616, -0.0323]
En-only	32K	$1.85\!\times\!10^{-3}$	$3.70\!\times\!10^{-3}$	$2.08\!\times\!10^{-3}$	0.3380 $\to$ 0.2978	[-0.0537, -0.0244]
En-only	64K	$3.19\!\times\!10^{-1}$	$3.19\!\times\!10^{-1}$	$3.19\!\times\!10^{-1}$	0.3183 $\to$ 0.2950	[-0.0353, -0.0109]
6-lang	16K	$2.46\!\times\!10^{-33}$	$2.21\!\times\!10^{-32}$	$2.21\!\times\!10^{-32}$	0.2960 $\to$ 0.2703	[-0.0376, -0.0139]
6-lang	32K	$4.12\!\times\!10^{-26}$	$2.88\!\times\!10^{-25}$	$1.24\!\times\!10^{-25}$	0.2717 $\to$ 0.2522	[-0.0325, -0.0074]
6-lang	64K	$2.34\!\times\!10^{-22}$	$1.17\!\times\!10^{-21}$	$4.20\!\times\!10^{-22}$	0.2603 $\to$ 0.2442	[-0.0299, -0.0044]

Appendix

Table E.1: Statistical analysis of the nine core FLORES-200 comparisons. The Gini CI uses paired resampling over non-English premiums; Wilcoxon uses paired mean token counts.

Script cohort	$n$	Standard	SuperBPE-style	SCRIPT-style
Latin/Arabic/Cyrillic	161	59.205 $\to$ 59.078	64.534 $\to$ 64.262	59.832 $\to$ 59.719
South Asian	21	103.401 $\to$ 103.199	91.472 $\to$ 91.514	77.316 $\to$ 78.385
Han/Japanese	4	43.363 $\to$ 43.364	43.455 $\to$ 43.465	42.722 $\to$ 42.801
Khmer/Lao/Myanmar/Thai	5	130.013 $\to$ 130.446	113.257 $\to$ 115.093	126.617 $\to$ 124.854
Remaining	13	164.495 $\to$ 134.332	160.236 $\to$ 130.662	170.053 $\to$ 139.682
All	204	71.889 $\to$ 69.857	74.187 $\to$ 72.137	69.957 $\to$ 68.001

Appendix

Table E.2: Primary-corpus FLORES-200 breakdown (101-language mC4, 32K vocabularies). Entries are BBPE

\to

UBE final tokens/sentence;

n

counts language–script configurations.

Pre-tokenization	Method	English	Selected BMP (12)	Overall	Gini	Lowest
Standard GPT-2	BBPE	36.959	183.592	67.875	0.2281	6/204
Standard GPT-2	UBE (ours)	36.938	153.442	66.079	0.2086	162 /204
Official script/category	SCRIPT-BPE	37.982	151.444	63.785	0.1785	25/204
SCRIPT-style	BBPE	37.569	188.228	65.677	0.2048	7/204
SCRIPT-style	UBE (ours)	37.539	154.015	63.918	0.1832	4/204

Appendix

Table E.3: SCRIPT-BPE comparison on FLORES-200. Means are final BPE tokens/sentence; Selected BMP denotes the 12 configurations defined in Table D.3 . Gini uses 203 English-normalized non-English premiums. Tokenizer training uses the same 101-language mC4 prefix (approximately 50 MiB of UTF-8 text per language; 32K vocabulary). Change rows compare UBE with BBPE under the same boundary policy; “Lowest” counts wins among all five methods.

Evaluation	Script cohort	Comparisons	UBE lower	SCRIPT-BPE lower
FLORES-200	Latin/Arabic/Cyrillic	161	156	5
FLORES-200	South Asian scripts	21	1	20
FLORES-200	Han/Japanese	4	0	4
FLORES-200	Khmer/Lao/Myanmar/Thai	5	0	5
FLORES-200	Remaining scripts	13	12	1
FLORES-200	All	204	169	35

Appendix

Table E.4: Pairwise UBE/SCRIPT-BPE win counts (32K tokenizers trained on 101-language mC4 with approximately 50 MiB of UTF-8 text per language).

Boundary policy	Representation	English	Selected BMP (12)	Overall	Gini
Standard	BBPE	36.959	183.592	67.875	0.2281
Standard	UBE (ours)	36.938	153.442	66.079	0.2086
SuperBPE-style	BBPE	39.586	163.120	68.858	0.1916
SuperBPE-style	UBE (ours)	39.524	135.050	67.208	0.1736

Appendix

Table E.5: Representation-by-boundary composition on FLORES-200 (101-language mC4, approximately 50 MiB/language, 32K vocabulary). Means are final BPE output tokens per sentence; Selected BMP denotes the 12 configurations defined in Table D.3 . Gini uses 203 English-normalized non-English premiums.

Base representation	Alphabet	English	Selected BMP (12)	Overall	Total positions
UTF-8 bytes	256	130.530	292.895	196.106	40,485,643
UBE (ours)	512	130.481	199.638	176.088	36,353,040
MYTE	256	109.140	255.737	148.142	30,583,555
MYTE-UBE (ours)	512	109.094	183.867	141.354	29,182,333

Appendix

Table E.6: Representation-only FLORES-200 comparison. Means and totals are base-symbol positions, not learned BPE output tokens; Selected BMP denotes the 12 configurations defined in Table D.3 .

Unicode 17 input set	Inputs	UBE raw exact	SCRIPT-BPE after NFC	MYTE raw exact
Assigned through Unicode 16 (excluding Private-Use)	155,063	All recovered	All recovered	154,023/155,063
Assigned in Unicode 17	4,803	All recovered	Outside released coverage	All recovered
Private-Use (Co)	137,468	All recovered	Filtered by design	All recovered
Unassigned/reserved (Cn)	814,730	All recovered	Filtered by design	All recovered
NormalizationTest.txt unique inputs	37,671	All recovered	166 outside coverage	21,373/37,671
GraphemeBreakTest.txt unique inputs	764	All recovered	74 outside coverage	684/764

Appendix

Table E.7: Deterministic Unicode 17 fidelity audit (UBE/SCRIPT-BPE tokenizers from Table E.3 ). “After NFC” includes SCRIPT-BPE’s normalization; no aggregate is reported across excluded inputs. Sequence files are versioned stress tests, not conformance claims.

Tokenizer Steps Model-input positions Supervised UTF-8 bytes Common-text BPB

\downarrow

BPB Gini

\downarrow

UBE (ours, primary tokenizer) 200,000 104,857,600,000 361,441,233,445 1.5298 0.1735 BBPE 200,437 105,086,713,856 361,439,652,484 1.5439 0.1740 BBPE16 237,228 124,375,793,664 361,440,525,559 1.5418 0.1882

Tokenizer XNLI XCOPA Bele ARC-E ARC-C HSwag PIQA LAMB English mean UBE (ours, primary tokenizer) 34.4 51.8 23.7 35.9 17.8 27.9 60.1 23.1 33.0 BBPE 34.4 51.5 24.0 35.3 16.8 27.9 60.3 23.3 32.7 BBPE16 34.4 51.9 23.3 35.5 18.0 27.6 60.8 19.5 32.3

Appendix

Table E.8: Source-byte-matched language-model comparison. Top: training exposure (steps, model-input positions, supervised UTF-8 bytes) and held-out BPB/BPB Gini. Bottom: zero-shot accuracy (%); English mean weights the five English tasks equally.

Evaluation	Entries	Primary 3-byte BMP
XNLI	15	3
Belebele	122	26
XCOPA	11	3
MGSM	11	5
Each English benchmark	1	0
Held-out BPB	101	20

Appendix

Table E.9: Language-entry coverage by primary encoding-floor group. Each English benchmark denotes ARC-Easy, ARC-Challenge, HellaSwag, PIQA, or LAMBADA separately.

Input content	ASCII-byte share	UBE $\Delta$ vs. BBPE	BBPE16 $\Delta$ vs. BBPE
Issue text	73.91%	-0.27% [-0.47%, -0.15%]	+21.92% [+20.56%, +23.19%]
Issue text + edit-file contents	99.24%	-0.18% [-0.21%, -0.15%]	+24.86% [+24.35%, +25.40%]

Appendix

Table E.10: Final BPE output-token changes on ten MAPS translations of 100 SWE-bench tasks, using 32K tokenizers trained on the primary 101-language corpus. Negative values favor the named method over BBPE.

Config	Tok	BPB	BPB Gini	XNLI	Bele	XCOPA	ARC-E	ARC-C	HSwag	PIQA	LAMB
En-only	BBPE	1.4909	0.1993	34.3	23.6	51.4	38.1	18.3	27.5	61.9	23.6
En-only	UBE	1.4698	0.1989	34.4	23.0	51.5	38.5	17.0	27.8	62.7	22.5
6-lang	BBPE	1.3680	0.1930	34.6	23.5	52.0	35.6	17.1	27.4	61.3	23.1
6-lang	UBE	1.3760	0.1880	34.9	23.2	52.2	36.7	18.4	28.0	62.0	23.1
101-lang	BBPE	1.5174	0.1762	34.3	24.8	52.0	35.5	17.0	27.7	59.3	22.8
101-lang	UBE	1.5237	0.1739	34.4	23.7	51.8	35.9	17.8	27.9	60.1	23.1

Appendix

Table F.1: Full LM benchmark results (182M, 32K, 200K mC4 steps). Benchmark columns report zero-shot accuracy. XNLI/Belebele/XCOPA are multilingual; ARC-Easy/ARC-Challenge, HellaSwag, PIQA, and LAMBADA are English-only.

Config	Tok	BPB	BPB Gini	XNLI	Bele	XCOPA	ARC-E	ARC-C	HSwag	PIQA	LAMB
En-only	BBPE	1.4193	0.1971	34.5	22.9	52.3	39.4	17.4	28.5	62.6	25.6
En-only	UBE	1.3668	0.1956	34.8	22.9	52.8	37.3	19.0	28.2	62.3	27.0
6-lang	BBPE	1.3071	0.1956	34.8	23.1	52.9	40.4	18.9	29.2	63.5	28.9
6-lang	UBE	1.3106	0.1973	35.0	23.9	52.6	39.6	18.2	29.2	63.0	28.6
101-lang	BBPE	1.4673	0.1972	35.0	23.0	51.1	39.1	18.1	29.1	64.0	29.9
101-lang	UBE	1.4707	0.1983	34.5	23.4	51.9	39.9	18.4	29.1	63.2	29.1

Appendix

Table F.2: LM evaluation (1.3B, 32K vocabulary, 205K Chinchilla-style mC4 steps) across the same three tokenizer-training configurations as Table 4 . Lower-coverage rows ( En-only , 6-lang ) are fixed-compute stress tests. Benchmark columns report zero-shot accuracy.

Config	Tok	BPB	BPB Gini	XNLI	Bele	XCOPA	ARC-E	ARC-C	HSwag	PIQA	LAMB
En-only	BBPE	1.4879	0.2006	34.2	23.0	52.1	40.2	18.1	27.6	61.2	20.6
En-only	UBE	1.4588	0.2002	34.2	23.2	51.9	40.2	18.4	27.6	61.9	22.6
6-lang	BBPE	1.4329	0.1751	34.7	23.0	52.0	37.7	18.9	28.0	61.3	22.4
6-lang	UBE	1.4456	0.1812	34.5	22.9	51.6	38.1	19.3	27.8	62.1	22.8
101-lang	BBPE	1.4431	0.2037	34.8	23.2	51.5	39.5	17.6	28.1	61.4	23.2
101-lang	UBE	1.4435	0.2047	34.8	23.4	51.9	38.2	18.1	27.6	61.0	24.7

Appendix

Table F.3: LM evaluation (230M, 64K vocabulary, 200K mC4 steps) across the same three tokenizer-training configurations as Table 4 . Benchmark columns report zero-shot accuracy.

Config	Tok	BPB	BPB Gini	XNLI	Bele	XCOPA	ARC-E	ARC-C	HSwag	PIQA	LAMB
101-lang	BBPE	1.5044	0.2029	34.6	23.0	51.4	41.0	19.4	28.3	62.4	25.5
101-lang	UBE	1.5028	0.2021	34.9	23.1	51.6	40.0	19.4	28.7	62.2	26.7
101-lang	S-BBPE	1.5064	0.1912	34.3	24.2	52.3	39.4	18.0	28.4	60.8	29.9
101-lang	S-UBE	1.4929	0.1914	34.3	23.4	52.4	39.9	17.8	28.2	62.6	26.7

Appendix

Table F.4: LM evaluation (329M, 128K vocabulary, 200K mC4 steps; 101-lang). S-BBPE and S-UBE provide SuperBPE-style composition checks. Benchmark columns report zero-shot accuracy.

Language modeling and multilingual accuracy
Tok	Run	BPB	BPB Gini	XNLI	Bele	XCOPA
BBPE	A	1.4674	0.1863	34.7	23.6	51.8
	B	1.4715	0.1833	34.3	24.5	51.8
	Primary	1.5174	0.1762	34.3	24.8	52.0
	mean	1.4854 $\pm$ 0.0278	0.1819 $\pm$ 0.0052	34.4 $\pm$ 0.2	24.3 $\pm$ 0.7	51.9 $\pm$ 0.1
UBE	A	1.4717	0.1879	34.9	23.5	51.8

Appendix

Table F.5: Multi-seed LM evaluation (182M, 32K vocabulary, 101-lang, 3 seeds). Mean rows report mean

\pm

cross-seed standard deviation for BPB, BPB Gini, and benchmarks. Benchmark columns report zero-shot accuracy. Matching run labels use the same seed.

Language	BBPE tok	UBE tok	$\Delta$ tok	BBPE rate (seq/s, $\pm$ 95% CI)	UBE rate (seq/s, $\pm$ 95% CI)	$\Delta$ spd
Amharic	1438	976	-32.1%	31.8 $\pm$ 0.01	46.0 $\pm$ 0.02	+45.01%
Arabic	975	957	-1.8%	46.1 $\pm$ 0.01	48.1 $\pm$ 0.02	+4.29%
English	226	223	-1.3%	145.3 $\pm$ 0.15	151.0 $\pm$ 0.41	+3.90%
Hindi	672	672	0.0%	65.2 $\pm$ 0.03	65.3 $\pm$ 0.02	+0.08%
Japanese	354	347	-2.0%	107.5 $\pm$ 0.05	108.7 $\pm$ 0.05	+1.07%
Korean	260	253	-2.7%	140.1 $\pm$ 0.10	143.9 $\pm$ 0.10	+2.76%

Appendix

Table F.6: Same-content prompt-processing benchmark on 5-sentence FLORES-200 bundles ( 6-lang/32K , 182M, A100 80 GB, batch 1, FP32, torch.compile ). Token columns are bundle lengths; rates use 30 trials with 95% CI half-widths over 28 samples after symmetric min/max trimming.

\Delta

tok = (UBE

-

BBPE)/BBPE and

\Delta

spd = relative sequence-rate change.

Stage	Configured steps	Runs	Wall-clock/run	GPU-hours/run	Total GPU-hours
182M/32K LM pre-training	200,000	12	1–2 days	192–384	2,304–4,608
230M/64K LM pre-training	200,000	6	1.5–3 days	288–576	1,728–3,456
329M/128K LM pre-training	200,000	4	2–4 days	384–768	1,536–3,072
1.3B/32K LM pre-training	205,000	6	8–14 days	1,536–2,688	9,216–16,128
Downstream eval. + inference	n/a	agg.	$<$ 6 h	$<$ 6	$<$ 6

Appendix

Table G.1: Approximate active compute budget for the primary 28-run LM suite. GPU-hours per run are wall-clock hours multiplied by 8 GPUs for LM training rows.

Method	Steps	Training time (hours)
UBE	200,000	12.94
BBPE	178,549	24.76
BBPE16	177,699	24.68
MYTE+BPE	179,526	12.06
SCRIPT-BPE	170,191	11.39

Appendix

Table G.2: Recorded training wall-clock time for the five-tokenizer matched-exposure comparison (Table 6 ), separate from the primary-suite budget.

Paper	Venue	Tok. langs	Eval langs	LM size	LM tokens	Benchmarks	Vocab
Arnett+	NeurIPS’25	97	97	—	—	—	8–262K
Schmidt+	EMNLP’24	1 (en)	1	350M–2.4B (64 runs)	200B	10 EN	33–49K
SuperBPE	COLM’25	1 (en)	1	680M–11B	$\sim$ 330B (8B baseline)	30 EN	200K
Abagyan+	arXiv’25	62	NR	3.3B	100B+10.5B	2 multi+11 EN + Aya	100–250K
Land & Arnett	TokShop’25	12 (mono)	12 (mono)	—	—	—	64K (mono); 256K merges (multi)
MAGNET	NeurIPS’24	9	9	126M	$\sim$ 10B bytes	4 multi+2	50–250K (BPE baselines)

Appendix

Table H.1: Experimental scope of selected tokenizer studies. Tok. langs = tokenizer-training languages, eval langs = intrinsic-evaluation languages, and benchmarks = downstream benchmarks. NR = intrinsic-evaluation language count not separately reported; mono/multi = monolingual/multilingual. The UBE row covers the primary 28-run suite.

Tokenizer	Vocab	EN avg	Total (M)	Gini	Type
GPT-4o (o200k)	200,019	26.554	11.6	0.2323	tiktoken
GPT-4 (cl100k)	100,277	26.860	18.6	0.3287	tiktoken
Llama 3	128,256	26.849	16.6	0.3318	tiktoken-BPE
Qwen 2.5	151,665	27.293	15.4	0.2724	BPE
Qwen 3.6 multimodal tokenizer †	248,044	27.204	12.8	0.2420	BPE
Mistral NeMo	131,072	27.497	14.2	0.3030	Tekken

Appendix

Table H.2: FLORES-200 Gini coefficient and total token count for tokenizers used by public LLMs, included as Figure 2 reference points.

Model	Lang scope	Params	Vocab	Tok. source	Scale
Compact tokenizer (32K–50K)
SmolLM2	English	135M–1.7B	49,152	self-trained	2T–11T tok
MobileLLM	English	125M–1B	32,000	Llama 2	1T tok
Pythia	English	160M–2.8B	50,277	GPT-NeoX	300B tok
TinyLlama	English + code	1.1B	32,000	Llama 2	3T tok
Phi-3.5-mini	23 langs	3.8B	32,064	Llama 2	3.4T tok

Appendix

Table H.3: Representative compact public LMs grouped by tokenizer vocabulary scale: compact (32K–50K), large (100K+), and family-inherited (128K+). Tok. source identifies the tokenizer source or family.

Asset	Access point / version used	License	Terms / handling in this work
mC4 pretraining corpus	Hugging Face allenai/c4 multilingual; deterministic 101-language local Arrow mirror	ODC-BY	Derived from Common Crawl snapshots. Fixed capped subsets are used for tokenizer and LM training; the local Arrow mirror is reconstructed from public sources and not redistributed.
FLORES-200	Hugging Face facebook/flores , devtest split	CC BY-SA 4.0	Used for the primary parallel intrinsic benchmark.
UDHR	NLTK udhr2 corpus (Unicode UDHR) and United Nations declaration text	Public UN text; public corpus distribution	Used only for robustness evaluation. Our loader downloads the public corpus when absent and records the local corpus checksum.
SIB-200	Hugging Face Davlan/sib200 , test split	CC BY-SA 4.0	Used only for robustness evaluation. Our loader uses a pinned public revision and records a text checksum for the measured split.
MGSM	Hugging Face juletxara/mgsm , test split	CC BY-SA 4.0	Used for robustness evaluation. Our preprocessing script builds the evaluation JSON with the original question text from the public dataset.
XNLI	facebookresearch/XNLI ; Hugging Face facebook/xnli , test split	CC BY-NC 4.0	Used for multilingual downstream evaluation only. We preserve the noncommercial restriction and do not redistribute the data.

Appendix

Table I.1: Existing dataset and benchmark assets used in this work.

Asset	Access point / version used	License	Terms / handling in this work
Hugging Face tokenizers	Upstream huggingface/tokenizers ; 0.22.3-dev.778/779	Apache 2.0	Primary SCRIPT-style tokenizer training uses dev.779. Our modifications retain the Apache 2.0 license and attribution files.
LitGPT	Lightning-AI/litgpt , version 0.5.11	Apache 2.0	Our trainer vendors and modifies selected LitGPT modules while preserving the upstream license notice and attribution.
lm-evaluation- harness	EleutherAI repo, version 0.4.9.1	MIT	Used as the evaluation backend for our offline benchmark runner, pinned to this version.
SCRIPT-BPE	sanderland/script_tok , commit 54a1058	Apache 2.0	Official SCRIPT-BPE tokenizers in Tables E.3 , E.7 , and 6 .
MYTE maps and decoder	tomlimi/MYTE , commit 177299c	Not stated in the repository	Used unmodified for Tables E.6 , E.7 , and 6 ; not redistributed.

Appendix

Table I.2: Existing software assets used in this work.

Language	Tokenizer	16K	32K	64K	128K	256K
Acehnese (Arabic)	UBE	85.927	79.600	75.741	70.241	67.480
	BBPE	86.626	79.774	75.897	70.374	67.620
Acehnese (Latin)	UBE	60.001	56.095	52.427	49.271	45.328
	BBPE	60.122	56.191	52.458	49.284	45.389
Afrikaans	UBE	52.071	47.458	43.605	40.014	36.589
	BBPE	52.299	47.583	43.651	40.024	36.590

Appendix

Table I.3: Complete FLORES-200 mean tokens per sentence for all 204 language–script configurations under 101-lang tokenizer training. Columns give vocabulary sizes; each language has paired UBE and BBPE rows. Means are rounded to three decimal places; the win/tie/loss counts in Table D.1 use unrounded values.

Language	BBPE	UBE	$\Delta$
ar	31.8	33.6	+1.8
bg	33.2	33.4	+0.2
de	33.4	34.1	+0.7
el	33.7	33.4	-0.3
en	42.0	40.4	-1.6
es	33.4	34.0	+0.6

Appendix

Table I.4: Per-language XNLI zero-shot accuracy under the primary 101-lang/32K 182M configuration (15 languages).

\Delta

= UBE

-

BBPE.

Language	BBPE	UBE	$\Delta$
et	50.8	49.8	-1.0
ht	52.0	52.4	+0.4
id	53.4	52.2	-1.2
it	48.4	49.6	+1.2
qu	50.0	48.6	-1.4
sw	54.6	53.4	-1.2

Appendix

Table I.5: Per-language XCOPA zero-shot accuracy under the primary 101-lang/32K 182M configuration (11 languages).

\Delta

= UBE

-

BBPE.

Language	BBPE BPB	UBE BPB	$\Delta$ BPB	UBE PPL
af	1.9670	1.9795	+0.0125	52.31
am	1.0390	1.0483	+0.0093	9.50
ar	1.2644	1.2544	-0.0100	17.86
az	1.9380	1.9210	-0.0170	36.04
be	1.4010	1.4363	+0.0353	30.66
bg	1.1531	1.1601	+0.0070	21.12

Appendix

Table I.6: Per-language held-out BPB for BBPE and UBE ( 101-lang/32K , 182M; 101 languages). UBE token-level perplexity is shown only as a scale reference;

\Delta

BPB = UBE

-

BBPE.

Explore similar work

Writing-System-Level Tokenizer Adaptation for Byte-Level BPE
Aug 1, 2026Bohdan DidenkoByte-Pair Encoding Tokenizer
Lviv Polytechnic National University
Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization
Aug 6, 2025Negar Foroutan, Clara Meister, Debjit Paul +4Byte-Pair Encoding Cross-Lingual Consistency
EPFL · University of Zurich
Equity with Efficiency: An Empirical Study of Tokenizers for Multilingual Large Language Models
Jun 13, 2026Kieron Seven Jun Wei Lee, Muhammad Reza Qorib, Andrew Ivan Soegeng +1Multilingual Language Models Tokenizer
National University of Singapore · Carnegie Mellon University · SAP

Aug 1, 2026, cs.CLJ/K move · Enter open · S save

Writing-System-Level Tokenizer Adaptation for Byte-Level BPE

Bohdan Didenko

Lviv Polytechnic National University

Byte-Pair Encoding Tokenizer Post-Hoc

Open paper AskPDF arXiv

Pretrained byte-level BPE tokenizers can segment underrepresented languages inefficiently. Replacing a tokenizer changes the meaning of nearly every token ID, while vocabulary expansion enlarges the model's embedding and output matrices. We study post-hoc adaptation that keeps the model-vocabulary size fixed and preserves most existing token-to-ID assignments as a construction-time compatibility property. Directly transferring tokens from a language-specific tokenizer does not guarantee derivability through the target BPE merge graph: an inserted entry can conflict with the target's greedy merge ranks. We formalize this failure as the merge ordering problem and introduce BPE-guided insertion, which builds each transferred token through a target-reachable decomposition. Our pipeline uses script-aware row selection to limit collateral fragmentation, reconstructs target-script byte-level prerequisites, and applies guided insertion to maintain merge-graph reachability. On Ukrainian adaptations of Nemotron and GPT-OSS, it reduces token counts by 33.5% and 36.6%, keeps changes on English and the evaluated four-language European aggregate within 0.05%, and retains 78.5%/77.3% of original model-vocabulary rows at the same IDs. Constraint-matched global and frequency-based removal achieve similar Ukrainian compression but increase English/European token counts by 0.7-2.2%; fresh same-size retraining compresses Ukrainian slightly more but retains effectively no same-ID rows and increases English token counts by 7.6-8.6%. The reallocation increases token counts on the evaluated three-language Cyrillic micro-aggregate by 6.7%/10.1%. Structural audits find all 28,134/45,398 inserted BPE nodes reachable under ordinary rank-ordered merging and no retained same-ID model-vocabulary entry newly broken. We release all tokenizers and code.

Abstract

Explore similar work

Writing-System-Level Tokenizer Adaptation for Byte-Level BPE

Parity-Aware Byte-Pair Encoding: Improving Cross-lingual Fairness in Tokenization

Equity with Efficiency: An Empirical Study of Tokenizers for Multilingual Large Language Models