BitTokens: Bit-Level Token Representations

• BitTokens are bit-level token representations that redefine traditional interfaces in language modeling, watermarking, sub-byte encoding, and blockchain systems.
• In NLP and numeracy, BitTokens enable fixed-length binary encoding and IEEE 754 embeddings to reduce output-layer complexity and token counts while sustaining performance.
• In watermarking and blockchain, BitTokens embed multibit payloads and enforce consensus-based asset management, balancing efficiency with robustness and decentralization.

Searching arXiv for the cited BitTokens-related papers to ground the article in the current record. BitTokens is not a standardized term in the current literature; it denotes several distinct bit-centric token representations and token systems whose common feature is that token identity, payload, or state is expressed at bit granularity rather than solely through conventional vocabularies, decimal strings, or contract state. In recent arXiv usage, the term spans fixed-length binary lexical codes for language generation, single-token IEEE 754 number embeddings, multibit watermark tokens that carry Bernoulli evidence for every payload bit, variable-width bit-level BPE below the byte boundary, and, in a separate blockchain sense, Bitcoin-resident or Bitcoin-inspired token mechanisms and stake-based AI incentive tokens (Zhuang et al., 12 May 2026, Kreitner et al., 8 Oct 2025, Gloaguen et al., 12 May 2026, Moon et al., 9 Jun 2025, Bartoletti et al., 2020, Wang et al., 2023, Lui et al., 29 Jun 2025).

1. Scope of the term

In language modeling and tokenization, BitTokens are representational interfaces: they replace or augment standard token emission with explicitly structured bit patterns. In blockchain and decentralized AI, the same label or closely related usage refers instead to fungible or incentive-bearing tokens whose correctness, balances, or rewards are enforced by script rules, inscriptions, or stake-weighted consensus. The term therefore has no single canonical definition.

Usage	Core representation	Immediate objective
BitLM	Fixed-length binary code per vocabulary item	Multi-token causal generation
Single-token numeracy	IEEE 754 payload in one [NUM] token	Efficient arithmetic
Multibit watermarking	Every token carries Bernoulli evidence for all payload bits	Recover IDs or timestamps
Bit-level BPE	Variable-width bit tokens below the byte boundary	Lossless sequence compression
Bitcoin/deAI token systems	Script-, inscription-, or stake-defined token state	Fungibility, transfer, incentives

This distribution of meanings suggests that BitTokens is best understood as a family resemblance term. In all cases, the design move is to make token structure explicit at the bit level, but the surrounding semantics differ sharply: lexical realization, numerical representation, watermark payload transmission, Unicode compression, fungible asset accounting, and decentralized incentive allocation are distinct problem settings (Zhuang et al., 12 May 2026, Kreitner et al., 8 Oct 2025, Gloaguen et al., 12 May 2026, Moon et al., 9 Jun 2025, Bartoletti et al., 2020, Wang et al., 2023, Lui et al., 29 Jun 2025).

2. BitTokens as binary lexical codes in causal language modeling

In BitLM, BitTokens are the core representational choice: every vocabulary symbol is emitted and received as a short fixed-length binary code, and lexical realization is carried out by continuous diffusion in this binary space rather than by a large-vocabulary softmax. If the tokenizer vocabulary size is $V$, the number of bits per token is fixed to

$B = \lceil \log_2 V \rceil,$

and each token id $y \in \{0,\dots,V-1\}$ is mapped deterministically to

$\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$

For a sequence $y_{1:L}$, the clean binary matrix is

$$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$

Only the first $V$ patterns are used; the remaining $2^B-V$ codes are unused. The codebook is fixed, not learned, and coincides with token IDs in binary. This isolates the effect of the output interface itself (Zhuang et al., 12 May 2026).

Architecturally, BitLM keeps a standard transformer LLM backbone and adds a compact diffusion head. Binary codes are lifted into hidden space by a position-wise MLP, $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$, replacing a conventional embedding lookup. Causality is preserved across blocks through a block-causal mask $\mathcal{M}_m$, with block factorization

$B = \lceil \log_2 V \rceil,$0

where $B = \lceil \log_2 V \rceil,$1. The mask permits full attention within a block while preserving left-to-right causality across blocks. In the reported experiments, the block size is $B = \lceil \log_2 V \rceil,$2, so four tokens are refined jointly per decoding step (Zhuang et al., 12 May 2026).

The diffusion head uses an analog-bits relaxation for binary variables and trains an $B = \lceil \log_2 V \rceil,$3-prediction denoiser. For a clean block $B = \lceil \log_2 V \rceil,$4, one samples $B = \lceil \log_2 V \rceil,$5 and $B = \lceil \log_2 V \rceil,$6, forms

$B = \lceil \log_2 V \rceil,$7

and predicts

$B = \lceil \log_2 V \rceil,$8

Training minimizes

$B = \lceil \log_2 V \rceil,$9

At inference, the model iteratively denoises for $y \in \{0,\dots,V-1\}$0 steps per block using an ODE-style schedule and classifier-free guidance scale $y \in \{0,\dots,V-1\}$1, then commits by a hard sign projection, $y \in \{0,\dots,V-1\}$2, and maps back to token IDs with $y \in \{0,\dots,V-1\}$3. The denoiser explicitly does not factorize probabilities across positions or bit channels; correlations across both are represented implicitly by the denoising dynamics (Zhuang et al., 12 May 2026).

The principal complexity claim is that replacing a $y \in \{0,\dots,V-1\}$4-way softmax by $y \in \{0,\dots,V-1\}$5 bit channels changes output-layer complexity from $y \in \{0,\dots,V-1\}$6 to $y \in \{0,\dots,V-1\}$7 per position, with $y \in \{0,\dots,V-1\}$8. For vocabularies such as $y \in \{0,\dots,V-1\}$9, $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$0 is on the order of $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$1, eliminating the very large $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$2 projection of a standard LM head. The interface also amortizes backbone computation: one backbone update realizes $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$3 tokens, while the diffusion head performs $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$4 lightweight denoising steps on $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$5 channels. The paper does not report wall-clock speedups or tokens-per-second, and explicitly states that realized throughput depends on $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$6, $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$7, the relative cost of the diffusion head versus the removed softmax head, and hardware (Zhuang et al., 12 May 2026).

Empirically, BitLM is pretrained at $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$8B, $\phi(y)=2\cdot \mathrm{bin}_B(y)-1 \in \{-1,1\}^B.$9B, $y_{1:L}$0B, and $y_{1:L}$1B parameters on FineWeb-350B for one epoch, with smooth scaling in pretraining loss. On XSum, ROUGE for BitLM 8B with the diffusion head is $y_{1:L}$2 for PT and $y_{1:L}$3 for FT, compared with $y_{1:L}$4 and $y_{1:L}$5 for the same backbone using a conventional LM head. Pointer-generator baselines remain stronger, with PTGEN at $y_{1:L}$6, but the reported ablation indicates best XSum performance at $y_{1:L}$7 and CFG $y_{1:L}$8. The authors characterize the results as promising but preliminary and identify better denoising schedules, adaptive block size, learned or optimized codebooks, hybrid heads, task-specific copying, and sampling strategies beyond hard sign as open directions (Zhuang et al., 12 May 2026).

3. BitTokens as single-token number embeddings

A second meaning of BitTokens appears in numeracy-oriented language modeling, where the goal is to embed any number into a single token using its IEEE 754 binary floating-point representation. The motivation is that conventional BPE or subword tokenization splits a number such as “-1(2345.6789)” into multiple tokens, forcing arithmetic to be reconstructed from text and making multiplication and division especially difficult. The paper evaluates eight frontier LLMs on nine numeracy tasks and reports extreme reasoning-token counts for single calculations under maximal reasoning, including GPT-5 with Division $y_{1:L}$9, Standard deviation $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$0, and Exponentiation $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$1 tokens; Qwen3-235B with Division $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$2, Standard deviation $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$3, and Exponentiation $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$4; and DeepSeek v3.1 with Division $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$5, Standard deviation $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$6, and Exponentiation $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$7 (Kreitner et al., 8 Oct 2025).

The proposed BitTokens satisfy a set of nine desiderata: D1 Token efficiency, D2 Uniqueness, D3 Structured, D4 Scale invariance, D5 Normalization, D6 Numerical stability, D7 Continuity, D8 Robustness, and D9 Arithmetic. Prior single-token approaches are presented as falling short. xVal scales a trainable [NUM] token by the numeric value and must compress values to a small magnitude range such as $$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$8, breaking D4 and harming D6 and D9. FoNE uses Fourier features and satisfies D1-D8, but fails D9 because multiplication in sinusoidal encoding is non-local and computationally complex (Kreitner et al., 8 Oct 2025).

BitTokens encode a real number as one learned [NUM] token augmented with a deterministic 64-dimensional vector representing the float64 sign, exponent, and fraction bits. For normalized numbers,

$$A_{1:L,0}=[\phi(y_1),\dots,\phi(y_L)]^\top \in \{-1,1\}^{L\times B}.$$9

with $V$0, an 11-bit exponent field with bias $V$1, and a 52-bit fraction. Special cases are handled explicitly: zero has $V$2 and $V$3, subnormals have $V$4 and $V$5, and $V$6 equal to all 1s encodes $V$7 when $V$8 and NaN when $V$9. The bit vector is scaled from $2^B-V$0 to $2^B-V$1 to obtain unit RMS norm and compatibility with layer normalization, can be concatenated with an additional 64-bit reciprocal encoding $2^B-V$2, and is zero-padded to the model’s embedding dimension before being added to the learned [NUM] embedding. At output time, a linear layer plus sigmoid predicts bits, which are thresholded at $2^B-V$3 and decoded back to float64 (Kreitner et al., 8 Oct 2025).

The architecture itself is minimally changed. Controlled experiments use a 6-layer, 6-head, 768-dimensional transformer of approximately 117M parameters, trained up to 10B tokens with context length 1024 and effective batch size 192. The model uses untied embeddings and a dedicated number head with bit-wise binary cross-entropy, equal weighting for all bits, RoPE, QK-normalization, FlashAttention-2, and dropout $2^B-V$4. Inputs combine FineWeb 10B text with numeracy tasks spanning comparisons, addition/subtraction, multiplication, division, exponentiation, mean, and standard deviation. Numbers are sampled uniformly in log-space with up to 15 significant digits (Kreitner et al., 8 Oct 2025).

The reported gains are both algorithmic and token-efficiency-oriented. In multi-task performance, BitTokens attain Addition $2^B-V$5, Multiplication $2^B-V$6, Division $2^B-V$7, Min/Max $2^B-V$8, Interval $2^B-V$9, and Sorting $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$0, compared with Single-digit Addition $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$1, Multiplication $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$2, Division $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$3, Min/Max $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$4, Interval $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$5, and Sorting $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$6. FineWeb perplexity also improves, with BitTokens at $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$7 versus Single-digit $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$8 and Subword $$E_{1:L}=\mathrm{MLP}(A_{1:L,0}) \in \mathbb{R}^{L\times d}$$9. Token counts fall sharply: for Addition, BitTokens require $\mathcal{M}_m$0 input and output tokens, versus $\mathcal{M}_m$1 for Single-digit and $\mathcal{M}_m$2 for Subword; for Division, $\mathcal{M}_m$3 versus $\mathcal{M}_m$4 and $\mathcal{M}_m$5. Ablations show that a base-10 bit-encoding lowers multi-task Multiplication from $\mathcal{M}_m$6 to $\mathcal{M}_m$7 and Division from $\mathcal{M}_m$8 to $\mathcal{M}_m$9, while removing reciprocal bits drops multi-task Division from $B = \lceil \log_2 V \rceil,$00 to $B = \lceil \log_2 V \rceil,$01 (Kreitner et al., 8 Oct 2025).

The principal limitation is representational scope: float64 provides roughly 15–17 significant decimal digits and range approximately $B = \lceil \log_2 V \rceil,$02. The paper also notes a continuity caveat, since equal weighting of bit errors does not strictly reflect numeric closeness. Multi-step tasks such as exponentiation and standard deviation remain hard for small models across all methods, though the single-token interface still compresses the number of emitted tokens (Kreitner et al., 8 Oct 2025).

4. BitTokens as multibit watermark carriers

In watermarking, BitTokens refers to generated tokens that carry bit-level information for a hidden payload. The binomial multibit watermark encodes every bit of an $B = \lceil \log_2 V \rceil,$03-bit message at every token position. At position $B = \lceil \log_2 V \rceil,$04, one samples $B = \lceil \log_2 V \rceil,$05 independent Bernoulli$B = \lceil \log_2 V \rceil,$06 score vectors $B = \lceil \log_2 V \rceil,$07, one per payload bit, seeded by a private key and local context. For bit $B = \lceil \log_2 V \rceil,$08, the scores are complemented according to the message bit by an XNOR-like rule, yielding $B = \lceil \log_2 V \rceil,$09, and these are summed into an aggregate binomial score

$B = \lceil \log_2 V \rceil,$10

Tokens with larger $B = \lceil \log_2 V \rceil,$11 align with more bits of the message. The next-token distribution is then biased using either a Red-Green-style exponential tilt, a Soft PPL tournament rule, or a distortion-free SynthID-style transform (Gloaguen et al., 12 May 2026).

The point of this interface is that every emitted token contributes partial evidence to every payload bit. This differs from position-allocation methods such as MPAC, RSBH, StealthInk, BiMark, and MirrorMark, which encode only one bit per step, and from Cycle-Shift or ArcMark, which encode the whole message per token but incur exponentially scaling decoding cost with payload length. The binomial method is designed for practical payloads such as user IDs or timestamps, and the paper evaluates up to 64-bit payloads against eight baselines (Gloaguen et al., 12 May 2026).

Decoding is based on per-bit vote accumulation. For a realized token sequence $B = \lceil \log_2 V \rceil,$12, the decoder recomputes the Bernoulli bits $B = \lceil \log_2 V \rceil,$13 and forms counts

$B = \lceil \log_2 V \rceil,$14

Majority decoding sets $B = \lceil \log_2 V \rceil,$15, with random tie-breaking. Under the null hypothesis of no watermark, $B = \lceil \log_2 V \rceil,$16, enabling a two-sided binomial $B = \lceil \log_2 V \rceil,$17-value or normal approximation. The paper argues that prior metrics such as bit accuracy and message accuracy are insufficient because they assume the text is watermarked, and introduces per-bit confidence scoring and BA@$B = \lceil \log_2 V \rceil,$18\%FPR as practically relevant deployment metrics. For zero-bit detection over all bits, it defines a log-likelihood-ratio-style statistic $B = \lceil \log_2 V \rceil,$19 and estimates a Monte Carlo $B = \lceil \log_2 V \rceil,$20-value to control calibration on non-watermarked text (Gloaguen et al., 12 May 2026).

A stateful encoder reallocates pressure toward underencoded bits. It maintains the bit-wise progress statistic

$B = \lceil \log_2 V \rceil,$21

and scores candidate tokens by averaging the probability of future correctness over a small set of plausible future lengths $B = \lceil \log_2 V \rceil,$22. This stateful variant is intended to help bits near the decision boundary more than already well-encoded bits. The paper states that computing the stateful score requires only the tokens that survive base sampling constraints such as top-$B = \lceil \log_2 V \rceil,$23, keeping overhead manageable (Gloaguen et al., 12 May 2026).

The empirical picture is mixed but favorable. For 32-bit payloads at about 500 tokens, bit accuracy exceeds about 90% for the binomial watermark versus about 80% for the best baseline, and the stateless binomial variant achieves the best BA@1%FPR versus perplexity trade-off across 16-, 32-, and 64-bit payloads. The stateful variant improves message accuracy, particularly at larger payloads and low distortion, but slightly reduces per-bit confidence because it spreads pressure across bits rather than making any single bit extremely confident. Under 10% deletion, Ours+ records $B = \lceil \log_2 V \rceil,$24 bit accuracy versus MC2Mark $B = \lceil \log_2 V \rceil,$25, BiMark $B = \lceil \log_2 V \rceil,$26, and MPAC $B = \lceil \log_2 V \rceil,$27; under 10% substitution, Ours+ records $B = \lceil \log_2 V \rceil,$28 versus $B = \lceil \log_2 V \rceil,$29, $B = \lceil \log_2 V \rceil,$30, and $B = \lceil \log_2 V \rceil,$31. Paraphrasing drives bit accuracy toward approximately $B = \lceil \log_2 V \rceil,$32 for all methods, and message accuracy to approximately $B = \lceil \log_2 V \rceil,$33 at 10% edits. The paper therefore presents the approach as robust relative to prior multibit schemes, but not robust to strong paraphrase (Gloaguen et al., 12 May 2026).

5. BitTokens below the byte boundary

In tokenization research, BitTokens denotes tokens corresponding to bit sequences rather than fixed 8-bit bytes. The specific proposal addresses byte-level fallback in subword tokenization, where rare Unicode characters are emitted as UTF-8 bytes. For CJK and emoji this inflates sequence length severely: the paper states that, on average, a CJK character represented with UTF-8 bytes increases sequence length by three times relative to a character-level representation, while most emoji require four bytes. In evaluated datasets, the byte-level portion of tokens under the Llama2 tokenizer is $B = \lceil \log_2 V \rceil,$34 for Chinese, $B = \lceil \log_2 V \rceil,$35 for Korean, and $B = \lceil \log_2 V \rceil,$36 for Japanese (Moon et al., 9 Jun 2025).

The proposed bit-level BPE crosses below the byte boundary. For 3-byte UTF-8 CJK code points, each character is re-encoded as one 6-bit prefix token and two trailing 9-bit tokens. If successive characters share the same 6-bit prefix, the prefix is emitted once and then omitted until it changes. The method is therefore both variable-width and stateful with respect to the current prefix, but lossless. It does not rely on a prefix-free code or the Kraft–McMillan inequality; instead, it uses the deterministic 3-byte length of the target Unicode blocks and a constrained grammar in which a prefix token is followed by exactly two 9-bit payload tokens (Moon et al., 9 Jun 2025).

The paper specifies the encoding and decoding with bitwise equations. For bytes $B = \lceil \log_2 V \rceil,$37 representing one character, the prefix and 9-bit tokens are computed by equations (3)–(5); decoding re-inserts the current prefix after every 9-bit bigram and reconstructs the original bytes through equations (6)–(8). Vocabulary growth is modest: relative to the baseline byte-level BPE vocabulary, the tokenizer adds 256 extended-byte tokens in the range $B = \lceil \log_2 V \rceil,$38–$B = \lceil \log_2 V \rceil,$39 and three prefix tokens $B = \lceil \log_2 V \rceil,$40, for a total increase of 259 tokens (Moon et al., 9 Jun 2025).

The local compression gain can be substantial. For the Chinese string with UTF-8 bytes corresponding to “召唤众,” the re-encoding yields p1 5E 97 p1 CA A4 p1 5E, which after prefix deduplication becomes p1 5E 97 CA A4 5E, a $B = \lceil \log_2 V \rceil,$41 reduction compared to the original byte sequence. If a later character switches prefix, as in the Japanese example involving 认, the gain falls to $B = \lceil \log_2 V \rceil,$42. Corpus-wide reductions are smaller but still measurable: in English→Chinese, total length decreases from 131M to 127M ($B = \lceil \log_2 V \rceil,$43) and the byte portion from 64M to 60M ($B = \lceil \log_2 V \rceil,$44); in English→Japanese, from 176M to 174M ($B = \lceil \log_2 V \rceil,$45) and 41M to 40M ($B = \lceil \log_2 V \rceil,$46); in Japanese→Korean, from 113M to 111M ($B = \lceil \log_2 V \rceil,$47) and 59M to 56M ($B = \lceil \log_2 V \rceil,$48) (Moon et al., 9 Jun 2025).

The main trade-off is entropy. Using the Rényi-entropy-based efficiency analysis reproduced in the paper, the proposed tokenizer lowers entropy: English→Chinese drops from $B = \lceil \log_2 V \rceil,$49 to $B = \lceil \log_2 V \rceil,$50, English→Japanese from $B = \lceil \log_2 V \rceil,$51 to $B = \lceil \log_2 V \rceil,$52, and Japanese→Korean from $B = \lceil \log_2 V \rceil,$53 to $B = \lceil \log_2 V \rceil,$54. Figure 1 attributes this to the heavy concentration of probability mass on the new high-frequency prefix tokens. The authors therefore prioritize length reduction over entropy and caution that lower entropy may correlate with lower model quality. Throughput is correspondingly ambiguous. The paper proposes a “perceived TPS” given approximately by $B = \lceil \log_2 V \rceil,$55, where $B = \lceil \log_2 V \rceil,$56 is the relative sequence-length gain on references. On that measure, English→Chinese reports $B = \lceil \log_2 V \rceil,$57 for the byte baseline versus $B = \lceil \log_2 V \rceil,$58 for the proposed tokenizer, English→Japanese $B = \lceil \log_2 V \rceil,$59 versus $B = \lceil \log_2 V \rceil,$60, and Japanese→Korean $B = \lceil \log_2 V \rceil,$61 versus $B = \lceil \log_2 V \rceil,$62, indicating that actual throughput depends on the balance between shorter sequences and larger vocabularies (Moon et al., 9 Jun 2025).

6. BitTokens in blockchain and decentralized AI

A distinct usage of BitTokens appears in blockchain literature, where the focus is not lexical tokenization but fungible asset representation and incentive allocation. In "Computationally sound Bitcoin tokens," BitTokens are fungible tokens on Bitcoin enforced by consensus through a small extension of the Bitcoin Script language. The symbolic model represents ownership and value via deposits of the form $B = \lceil \log_2 V \rceil,$63 and authorizations $B = \lceil \log_2 V \rceil,$64, with actions Gen, Burn, Split, Join, Exchange, and Give. The implementation introduces neighborhood covenants, including rtxo, stxo, ptxo, verscr, and verrec, and carries token state in each UTXO’s arg field as (op, owner, tkval, tkid). The security claim is computational soundness: adversaries can make computational executions diverge from the ideal symbolic functionality only with negligible probability, under standard assumptions such as EUF-CMA signatures and collision-resistant hashing (Bartoletti et al., 2020).

The per-transaction conservation properties are explicit. Split enforces

$B = \lceil \log_2 V \rceil,$65

Join enforces the output value as the sum of sibling inputs, and Exchange and Give preserve tkid and tkval while changing owners as authorized. The design is intended to prevent unauthorized mint, inflation, theft, and cross-token joins directly at script-validation time. This sharply contrasts with off-chain indexer models (Bartoletti et al., 2020).

BRC-20 is such an off-chain model. It is an experimental token format introduced in March 2023 that brings token-like behavior to Bitcoin without smart contracts by combining ordinal theory and inscriptions. Deploy, mint, and transfer are encoded as JSON inscriptions attached to specific satoshis, while canonical state lives in off-chain indexers that parse inscriptions, enforce ticker-specific constraints, track ordering and transfer-intent consumption, and maintain balances. The constraints are simple:

$B = \lceil \log_2 V \rceil,$66

and a valid transfer additionally requires $B = \lceil \log_2 V \rceil,$67 and proper forwarding of the inscribed satoshi. The paper stresses that balances are fungible only insofar as indexers agree on state and documents edge cases such as “cursed/unbound” inscriptions, two-step transfer timing sensitivities, and centralization pressures around APIs and indexers (Wang et al., 2023).

The empirical network effects of BRC-20 in 2023 were material. The paper correlates the surge with average block size increasing from approximately 1.2 MB to more than 2 MB, mempool transaction count rising toward approximately 25,000 from approximately 5,000 in much of 2022, non-Ordinal fees rising by approximately 10% in March, and cumulative inscription fees exceeding 150 BTC. Its synthesis table concludes “more hype than hope,” with 34 hype versus 27 hope signals, while still acknowledging that BRC-20 expanded Bitcoin’s functionality and catalyzed new infrastructure (Wang et al., 2023).

A third economic-token usage appears in Bittensor, where the native token TAO underpins a decentralized AI marketplace. TAO has a fixed cap of 21 million with halvings on a schedule similar to Bitcoin, and emissions per tempo are split 41% to miners, 41% to validators/stakers, and 18% to subnet owners. Rewards are assigned under Yuma Consensus, a stake-weighted, median-clipping, subjective-utility mechanism. The paper reports strong concentration across 64 active subnets: overall stake Gini mean $B = \lceil \log_2 V \rceil,$68, overall reward Gini mean $B = \lceil \log_2 V \rceil,$69, and top 1% wallets holding on average $B = \lceil \log_2 V \rceil,$70 of stake while earning $B = \lceil \log_2 V \rceil,$71 of rewards. Validators show particularly strong stake-to-reward correlation, approximately $B = \lceil \log_2 V \rceil,$72–$B = \lceil \log_2 V \rceil,$73, and miners $B = \lceil \log_2 V \rceil,$74–$B = \lceil \log_2 V \rceil,$75, while performance-to-reward correlation is weaker, approximately $B = \lceil \log_2 V \rceil,$76 for validators and $B = \lceil \log_2 V \rceil,$77–$B = \lceil \log_2 V \rceil,$78 for miners. The paper therefore characterizes rewards as overwhelmingly driven by stake and proposes protocol-level interventions, including performance-weighted emission split, composite scoring, trust-bonus multipliers, and an 88th-percentile stake cap that moves the median coalition fraction needed for a 51% attack from approximately 1–2% to approximately 20% with about 22% whale penalty (Lui et al., 29 Jun 2025).

7. Comparative themes and unresolved questions

Across these usages, BitTokens reassign where commitment happens. In BitLM, commitment moves from a categorical softmax over $B = \lceil \log_2 V \rceil,$79 symbols to iterative denoising over $B = \lceil \log_2 V \rceil,$80 channels; in numeracy, it moves from decimal strings to IEEE 754 sign, exponent, and fraction bits; in watermarking, each token becomes repeated statistical evidence for every payload bit; in bit-level BPE, byte fallback is replaced by variable-width bit tokens that expose repeated UTF-8 structure (Zhuang et al., 12 May 2026, Kreitner et al., 8 Oct 2025, Gloaguen et al., 12 May 2026, Moon et al., 9 Jun 2025).

This suggests a common methodological pattern: bit-level interfaces are introduced when a conventional token interface is viewed as the bottleneck. In BitLM the bottleneck is the one-token-at-a-time softmax; in numeracy it is multi-token decimal rendering and long reasoning chains; in watermarking it is one-bit-per-position allocation; in CJK byte fallback it is the rigidity of the 8-bit boundary. The reported gains are correspondingly interface-level: lower output-layer complexity and native blockwise generation, large reductions in numeric token counts, stronger multibit recovery at low distortion, and shorter sequences for Unicode-heavy text (Zhuang et al., 12 May 2026, Kreitner et al., 8 Oct 2025, Gloaguen et al., 12 May 2026, Moon et al., 9 Jun 2025).

The corresponding liabilities also recur. Binary geometry can improve compactness while raising calibration questions: BitLM notes that every bit matters and does not report explicit analyses of bit-error robustness or probability calibration; the watermarking paper argues that earlier metrics lacked practical insight and therefore introduces per-bit confidence scoring; the numeracy paper identifies a continuity caveat because equal bit weighting does not strictly reflect numeric distance; bit-level BPE reduces sequence length but further lowers entropy (Zhuang et al., 12 May 2026, Gloaguen et al., 12 May 2026, Kreitner et al., 8 Oct 2025, Moon et al., 9 Jun 2025).

In blockchain usage, the central contrast is not between tokenizers but between trust models. The neighborhood-covenant construction seeks consensus-enforced conservation and authorization, whereas BRC-20 explicitly locates state in off-chain indexers. Bittensor adds a third regime in which token rewards and influence are mediated by stake-weighted consensus, raising concentration and incentive-alignment questions rather than lexical or decoding ones. A plausible implication is that “BitTokens” in blockchain contexts refers less to bitwise representation itself than to how token correctness or token influence is grounded—on-chain scripts, off-chain inscription parsers, or stake-weighted reward formulas (Bartoletti et al., 2020, Wang et al., 2023, Lui et al., 29 Jun 2025).

The open problems are therefore domain-specific. BitLM identifies better denoising schedules and solvers, adaptive block size, learned or optimized codebooks, hybrid heads, copying mechanisms, and sampling strategies beyond hard sign. The numeracy work leaves higher precision, scientific notation, and frontier-scale evaluation as natural extensions. The watermarking work points to robustness under strong paraphrase, larger payloads, and stateful integration with distortion-free schemes. Bit-level BPE explicitly leaves boundary design for other Unicode blocks, including emoji, as a generalization problem. BRC-20 highlights standardization of indexer rules and richer infrastructure as prerequisites for broader use, while the Bittensor analysis notes that the dTAO upgrade may materially change decentralization dynamics (Zhuang et al., 12 May 2026, Kreitner et al., 8 Oct 2025, Gloaguen et al., 12 May 2026, Moon et al., 9 Jun 2025, Wang et al., 2023, Lui et al., 29 Jun 2025).

In sum, BitTokens names a cluster of techniques and systems that expose token structure at bit granularity for efficiency, controllability, or enforceability. The term does not designate a single architecture or protocol. Its most coherent encyclopedic interpretation is as a family of bit-structured token interfaces whose implementations range from diffusion-based lexical realization and IEEE 754 numeric payloads to multibit watermark carriers, sub-byte Unicode compression, consensus-enforced Bitcoin assets, inscription-indexed Bitcoin token formats, and stake-governed decentralized AI incentives.