2508.11857

# SupraTok: Cross-Boundary Tokenization for Enhanced Language Model Performance **Authors**: Andrei-Valentin Tănase, Elena Pelican > Faculty of Mathematics and Computer Science, "Ovidius" University of Constanta, Constanta, Romania > Corresponding author: epelican@365.univ-ovidius.roFaculty of Mathematics and Computer Science, "Ovidius" University of Constanta, Constanta, Romania Abstract Tokenization remains a fundamental yet underexplored bottleneck in natural language processing, with strategies largely static despite remarkable progress in model architectures. We present SupraTok, a novel tokenization architecture that reimagines subword segmentation through three innovations: cross-boundary pattern learning that discovers multi-word semantic units, entropy-driven data curation that optimizes training corpus quality, and multi-phase curriculum learning for stable convergence. Our approach extends Byte-Pair Encoding by learning "superword" tokens—coherent multi-word expressions that preserve semantic unity while maximizing compression efficiency. SupraTok achieves 31% improvement in English tokenization efficiency (5.91 versus 4.51 characters per token) compared to OpenAI’s o200k tokenizer and 30% improvement over Google’s Gemma 3 tokenizer (256k vocabulary), while maintaining competitive performance across 38 languages. When integrated with a GPT-2 scale model (124M parameters) trained on 10 billion tokens from the FineWeb-Edu dataset, SupraTok yields 8.4% improvement on HellaSWAG and 9.5% on MMLU benchmarks without architectural modifications. While these results are promising at this scale, further validation at larger model scales is needed. These findings suggest that efficient tokenization can complement architectural innovations as a path to improved language model performance. Keywords: tokenization, natural language processing, language models, curriculum learning, entropy 1 Introduction The extraordinary success of large language models in recent years has fundamentally transformed our understanding of artificial intelligence and its potential applications across diverse domains [1, 2, 3]. From GPT-3’s demonstration of few-shot learning capabilities [4] to the emergence of models like Claude 4 Sonnet, GPT-5, and specialized systems achieving human-level performance on complex reasoning tasks [5], the field has witnessed unprecedented progress. This rapid advancement has been primarily attributed to three factors: increased model scale following established scaling laws [6], architectural innovations such as the transformer [7], and the availability of massive training datasets [8]. However, beneath these visible drivers of progress lies a foundational component that has received surprisingly little attention: tokenization. Tokenization serves as the critical interface between raw text and neural network processing, determining how linguistic information is encoded and presented to models. The dominant approach, Byte-Pair Encoding (BPE), was originally developed for data compression [9] and later adapted for neural machine translation [10]. While BPE and its variants (WordPiece [11], SentencePiece [12], and Unigram [13]) have enabled the current generation of language models, they operate under a fundamental constraint: tokens cannot cross word boundaries defined by whitespace or punctuation. This limitation, while computationally convenient, imposes an artificial fragmentation on natural language that may significantly impact model performance and efficiency. The word-boundary constraint proves problematic for multiple reasons that extend beyond simple efficiency concerns. First, it prevents the capture of ubiquitous multi-word expressions that function as semantic units in natural language. Phrases like "New York," "by the way," "machine learning," or "artificial intelligence" are conceptually atomic yet artificially fragmented by traditional tokenizers [14]. This fragmentation not only increases sequence length but also forces models to repeatedly learn to reconstruct these common patterns during training, wasting computational resources and model capacity. Second, the constraint creates inconsistencies across languages, particularly disadvantaging those without explicit word boundaries like Chinese, Japanese, and Thai, or agglutinative languages where word boundaries are fluid [15]. Third, it leads to the notorious problem of tokenization-induced bias at generation boundaries, where models exhibit unstable behavior when generating text that begins mid-token [16]. Recent research has begun to acknowledge these limitations and explore alternatives. The Byte Latent Transformer (BLT) eliminates fixed vocabularies entirely, operating on raw bytes with dynamic segmentation based on entropy signals [17]. However, this approach requires fundamental architectural changes and has only been validated at limited scales. Character and byte-level models [18, 19] avoid tokenization but suffer from dramatically increased sequence lengths, making them computationally prohibitive for many applications. Concurrent work on SuperBPE [20] and Boundless BPE [26] demonstrates the potential of removing whitespace constraints but lacks the comprehensive training methodology and evaluation we present here. The challenge extends beyond technical implementation to touch on fundamental questions about language representation. Linguistic theory has long recognized that language processing involves units larger than individual words—collocations, idioms, and formulaic sequences play crucial roles in human language comprehension and production [21]. Yet our tokenization methods, rooted in orthographic conventions rather than linguistic principles, fail to capture these natural units. This mismatch between how humans process language and how we present it to neural networks may explain why even massive models struggle with tasks that seem trivial to humans, such as consistent handling of proper nouns or idiomatic expressions. In this work, we present SupraTok, a comprehensive tokenization architecture that addresses these fundamental limitations through a principled, linguistically-motivated approach. Our method introduces three key innovations that work synergistically to improve both compression efficiency and semantic coherence. First, we implement cross-boundary pattern learning through a carefully designed curriculum that progressively relaxes word-boundary constraints, allowing the discovery of true linguistic units regardless of orthographic conventions. Second, we employ entropy-driven data curation to optimize the quality of training data, filtering out low-information content that would otherwise dominate frequency-based statistics. Third, we utilize multi-phase curriculum learning to ensure stable convergence while capturing increasingly complex patterns, from basic morphology to multi-word expressions and domain-specific terminology. 2 Results 2.1 Compression Efficiency and Vocabulary Utilization We evaluated SupraTok’s compression performance across multiple dimensions to understand its efficiency gains and their implications for practical deployment. Testing on the WikiText-103 dataset, a standard benchmark for language modeling that contains over 100 million tokens of high-quality Wikipedia text, revealed substantial improvements in compression efficiency compared to state-of-the-art tokenizers. To ensure fair comparison, we evaluate SupraTok against tokenizers with similar vocabulary sizes. Most critically, we compare against Google’s Gemma 3 tokenizer, which has an identical 256k vocabulary size to our approach. SupraTok achieves a compression ratio of 5.91 characters per token on English text, representing a 30.4% improvement over Gemma 3’s 4.53 characters per token. When compared to OpenAI’s o200k tokenizer (4.51 characters per token, 200k vocabulary), SupraTok shows a 31.0% improvement. Additional comparisons include Meta’s LLaMA 3.2 (4.50) and Alibaba’s Qwen 3 (4.34). Notably, our English-focused training variant demonstrates compression ratios that can reach a 45% improvement over certain baseline configurations when training data and vocabulary parameters are optimally aligned. This highlights the potential of specialized single-language optimization, though practical deployments typically require multilingual capability. Our multilingual variant shows more modest but still significant improvements across 38 languages with an average compression ratio of 2.93 characters per token for non-English tokenization, and 5.91 characters per token for English tokenization. Beyond raw compression, we observed that SupraTok demonstrates superior vocabulary utilization at 3.33%, compared to just 1.52% for o200k and even lower rates for other tokenizers. This metric is crucial as it indicates that SupraTok makes more effective use of its token space, avoiding the common problem of "dead tokens" that consume vocabulary slots but rarely appear in practice. The improved utilization suggests that our entropy-driven approach successfully identifies and prioritizes genuinely useful patterns rather than statistical artifacts. To understand the source of these efficiency gains, we analyzed the composition of SupraTok’s vocabulary. Approximately 42% of tokens represent cross-boundary patterns—multi-word units that traditional tokenizers cannot capture. These include high-frequency function word combinations ("in_the", "of_the"), common noun phrases ("United_States", "machine_learning"), and domain-specific terminology that frequently appears as a unit. The remaining vocabulary consists of traditional subword units, ensuring backward compatibility and coverage of rare words through compositional representation. 2.2 Downstream Task Performance The true test of a tokenization strategy lies not in compression metrics but in its impact on actual language understanding tasks. To evaluate this, we conducted controlled experiments using identical model architectures and training procedures, varying only the tokenization strategy. We pre-trained two GPT-2 scale models (124M parameters) on 10 billion tokens from the FineWeb-Edu dataset, a high-quality educational corpus filtered for pedagogical content. Both models used identical architectures (12 layers, 768 hidden dimensions, 12 attention heads) and training hyperparameters (learning rate 6e-4 with cosine decay, batch size 512, context length 1024). For fair comparison, the baseline model used a standard BPE tokenizer trained on the same data with an identical 256k vocabulary size, ensuring that differences in performance can be attributed to the tokenization methodology rather than vocabulary size or training data variations. On the HellaSWAG benchmark, which tests commonsense reasoning through sentence completion, the SupraTok-based model achieved 34.87% accuracy compared to 32.14% for the standard BPE baseline—a relative improvement of 8.4%. This benchmark is particularly relevant as it requires understanding of everyday situations and common expressions, exactly the type of content where multi-word units are prevalent. The improvement suggests that SupraTok’s ability to capture phrases and expressions as atomic units enhances the model’s grasp of commonsense patterns. The Massive Multitask Language Understanding (MMLU) benchmark provides an even more dramatic demonstration of SupraTok’s benefits. Covering 57 subjects from STEM to humanities, MMLU tests broad knowledge and reasoning capabilities. The SupraTok model achieved 27.75% accuracy versus 25.34% for the standard BPE baseline—a 9.5% relative improvement. This substantial gain across diverse domains indicates that the benefits of improved tokenization are not limited to specific types of content but extend broadly across different knowledge areas. To understand why SupraTok provides such consistent improvements, we analyzed model behavior on specific question types. On questions involving named entities, technical terminology, and multi-word expressions, the SupraTok model showed particularly large advantages, often exceeding 20% relative improvement. This aligns with our hypothesis that preserving semantic units as atomic tokens reduces the cognitive burden on the model, allowing it to focus on higher-level reasoning rather than reconstructing fragmented concepts. 2.3 Multilingual Generalization While SupraTok was primarily optimized for English, we evaluated its multilingual capabilities to understand how cross-boundary learning affects languages with different morphological and orthographic properties. Testing on the FLORES+ dataset across 38 languages revealed nuanced patterns of performance that illuminate both the strengths and limitations of our approach. For languages with similar word-boundary conventions to English (space-delimited scripts), SupraTok generally maintained or improved compression efficiency. Romance languages (French, Spanish, Italian) showed modest improvements of 5-10%, likely due to common multi-word constructions like articles with prepositions. Germanic languages showed similar patterns, with particularly strong performance on compound-heavy languages like German and Dutch, where SupraTok’s ability to learn longer units proves advantageous. For languages without explicit word boundaries, the results were more complex. Chinese and Japanese, which use logographic and mixed scripts respectively, showed slight degradation in compression ratio compared to specialized tokenizers but maintained competitive performance. Interestingly, SupraTok demonstrated superior handling of code-mixed text and loanwords in these languages, suggesting that cross-boundary learning provides unexpected benefits for multilingual phenomena. Agglutinative languages like Turkish, Finnish, and Hungarian presented unique challenges. While SupraTok’s cross-boundary approach should theoretically benefit these languages by learning common morphological combinations, we observed mixed results. The entropy-driven filtering, optimized for English text patterns, may inadvertently filter out important morphological variations in these languages, suggesting the need for language-specific adaptations of our approach. 2.4 Learned Pattern Analysis Detailed analysis of SupraTok’s vocabulary reveals systematic capture of linguistically meaningful units across multiple categories. We identified five primary types of cross-boundary patterns that emerge from the training process: Functional Constructions: High-frequency function word combinations dominate the early stages of cross-boundary learning. Patterns like "in_the", "of_the", "to_be", and "has_been" appear among the first superword tokens created. These combinations, while individually composed of the most common words in English, function as grammatical units that rarely need to be separated. Their atomic encoding reduces sequence length for virtually all English text. Named Entities and Proper Nouns: Geographic locations ("New_York", "United_States", "San_Francisco"), organizations ("United_Nations", "World_Health_Organization"), and common names appear as single tokens. This preservation of entity integrity has important implications for tasks requiring factual knowledge and consistency in entity handling. Domain-Specific Terminology: Technical terms from various fields are preserved as units: "machine_learning", "artificial_intelligence", "climate_change", "quantum_computing". This pattern is particularly pronounced in scientific and technical text, where domain terminology functions as atomic concepts that should not be arbitrarily fragmented. Idiomatic and Formulaic Expressions: Common expressions and idioms that function as semantic units are captured: "by_the_way", "in_order_to", "as_well_as", "state_of_the_art". These patterns are crucial for natural language understanding as their meaning often cannot be compositionally derived from their components. Morphological Patterns: Even within traditional word boundaries, SupraTok learns different patterns than standard BPE. Common prefixes and suffixes are more consistently tokenized, and the vocabulary shows better coverage of morphological variations, suggesting that the entropy-driven approach better captures linguistic regularities. 2.5 Computational Efficiency Analysis While compression ratio improvements directly translate to reduced sequence lengths and memory requirements, we conducted detailed analysis of the computational implications of SupraTok across different deployment scenarios. The 31% reduction in sequence length has cascading effects throughout the transformer architecture, where attention computation scales quadratically with sequence length. For a typical 2048-token context window, SupraTok effectively extends the accessible context to approximately 2680 tokens worth of content compared to traditional tokenization. This expansion is particularly valuable for tasks requiring long-context understanding, such as document summarization, multi-turn dialogue, and complex reasoning chains. In memory-constrained environments like mobile devices or edge deployments, this efficiency gain can make the difference between feasible and infeasible applications. Training efficiency also improves substantially. The shorter sequences reduce the number of forward and backward passes required to process the same amount of text, leading to approximately 24% reduction in training time for our experiments. This compounds over the course of large-scale pre-training, potentially saving hundreds of GPU-hours for production models. 3 Methods 3.1 Cross-Boundary Pattern Learning Architecture The foundation of SupraTok lies in its novel approach to discovering and encoding linguistic patterns that transcend traditional word boundaries. Our Advanced Cross-Boundary Pre-Tokenizer (ACBP) implements a sophisticated three-phase curriculum that progressively introduces complexity while maintaining training stability. The curriculum begins with traditional pattern learning (Phase 1: 0-100k merges) using standard BPE with whitespace boundaries. This phase establishes coverage of basic morphological units and ensures that the tokenizer can handle any text, including rare words and neologisms, through compositional representation. During this phase, we apply standard BPE merge operations but collect statistics on cross-boundary bigram frequencies for use in subsequent phases. Phase 2 (100k-200k merges) introduces controlled cross-boundary learning. We identify candidate multi-word patterns using pointwise mutual information (PMI), calculated as: $$ \text{PMI}(w_{1},w_{2},...,w_{n})=\log_{2}\left(\frac{P(w_{1},w_{2},...,w_{n})}{\prod_{i}P(w_{i})}\right) \tag{1} $$ where $P(w_{1},w_{2},...,w_{n})$ represents the joint probability of the word sequence and $P(w_{i})$ represents individual word probabilities. Sequences with PMI exceeding 2.0 are considered for merging, but we apply additional filters to ensure quality. First, we require minimum frequency thresholds (typically 100 occurrences in the training corpus) to avoid overfitting to rare patterns. Second, we implement a diversity constraint that prevents excessive overlap between selected patterns, ensuring broad coverage rather than variations of the same phrases. Phase 3 (200k-256k merges) focuses on complex expressions and domain-specific terminology. Here we employ a combination of statistical and linguistic signals. Beyond PMI, we incorporate left and right branching entropy to identify units with clear boundaries: $$ H_{\text{left}}(w)=-\sum_{c}P(c|w)\log_{2}P(c|w) \tag{2} $$ where $c$ represents possible left contexts for word sequence $w$ . Low entropy indicates consistent contexts, suggesting the sequence functions as a unit. We also utilize a lightweight language model (3-layer transformer trained on the same corpus) to identify sequences with high internal predictability but low external predictability—a signature of formulaic expressions. 3.2 Entropy-Driven Data Curation The quality of training data profoundly impacts tokenizer performance, yet most approaches train on raw, unfiltered corpora. SupraTok implements sophisticated entropy-based filtering to optimize the training signal and reduce computational requirements. We calculate character bigram entropy for each document in the training corpus: $$ H(\text{doc})=-\sum_{i}p(b_{i})\log_{2}p(b_{i}) \tag{3} $$ where $b_{i}$ represents character bigrams. Documents are classified into three categories based on their entropy: low ( $H<3.0$ ), medium ( $3.0≤ H≤ 4.5$ ), and high ( $H>4.5$ ). Analysis reveals that low-entropy documents typically contain boilerplate text, repetitive content, or data artifacts like base64 encodings. High-entropy documents contain diverse vocabulary, complex sentence structures, and information-rich content. We implement a balanced sampling strategy that retains 10% of low-entropy, 50% of medium-entropy, and 90% of high-entropy documents. This filtering reduces the training corpus by approximately 35% while improving vocabulary quality. The filtered corpus shows better coverage of meaningful patterns and reduces the learning of artifactual tokens from repetitive content. 3.3 Script-Aware Processing To ensure equitable treatment across writing systems, SupraTok implements Unicode-aware processing that adapts to different script properties. We classify Unicode characters into supercategories based on their general properties: - Letters and Marks (LM): Includes all letter categories and combining marks - Punctuation and Symbols (PS): Encompasses punctuation, symbols, and special characters - Numbers (N): Covers decimal digits and numeric symbols - Separators (Z): Includes spaces and other separator characters - Control and Other (C): Contains control characters and unassigned code points For scripts without explicit word boundaries (Chinese, Japanese, Thai), we apply character-level frequency analysis before merging, allowing the discovery of common character sequences that function as units. For agglutinative languages, we increase the weight of morphological boundaries detected through suffix analysis, ensuring that meaningful morphological units are preserved. 3.4 Vocabulary Optimization The final vocabulary size and composition significantly impact both compression efficiency and downstream performance. We employ Optuna-based Bayesian optimization to explore the hyperparameter space, optimizing a multi-objective function: $$ \text{Score}=\alpha\cdot\text{CompressionRatio}-\beta\cdot\log(\text{Perplexity})-\gamma\cdot\text{EncodeLatency} \tag{4} $$ where $\alpha$ , $\beta$ , and $\gamma$ are weighting parameters that balance different objectives. The optimization explores: - Vocabulary sizes from 160k to 320k tokens - Unigram ratios from 0.8 to 1.0 (controlling BPE vs. Unigram behavior) - Smoothing parameters for probability estimation - N-gram context lengths for pattern discovery - PMI thresholds from 1.5 to 3.0 - Frequency thresholds from 3 to 10 occurrences Through 100 trials of Bayesian optimization, we identified the 256k vocabulary configuration as optimal for English-focused applications, balancing compression efficiency with memory requirements. Complete implementation details, including hyperparameter configurations, algorithmic specifications, and training procedures, are provided in supplementary technical documentation. 3.5 Implementation Details SupraTok is implemented as a modular Python library built on top of the HuggingFace Tokenizers library for compatibility with existing ecosystems. The architecture consists of several interconnected components: The SupraTok class serves as the main interface, coordinating the training process and managing the vocabulary. It implements the three-phase curriculum learning schedule and handles the transition between phases based on merge counts. The AdvancedCrossBoundaryPreTokenizer implements the core pattern discovery logic, maintaining separate statistics for within-word and cross-boundary patterns. It uses efficient trie data structures for pattern matching and employs parallel processing for large-scale corpus analysis. The StreamingDataLoader handles efficient data processing with entropy filtering, implementing reservoir sampling for balanced data selection and streaming processing to handle corpora larger than memory. The NGramMiner discovers high-PMI phrases using parallel processing across document batches. It maintains memory-efficient count structures using probabilistic data structures for approximate counting when exact counts are not required. Training on a 100GB corpus requires approximately 40 hours on a distributed system with 8x NVIDIA A100 40GB GPUs, depending on the specific configuration and data preprocessing requirements. The three-phase curriculum and entropy-based filtering add computational overhead compared to standard BPE training. The trained tokenizer can be exported to standard formats (HuggingFace tokenizers.json, SentencePiece model files) for integration with existing pipelines. Detailed implementation specifications are available in the supplementary technical documentation. 3.6 Evaluation Methodology We conducted comprehensive evaluation across multiple dimensions to assess SupraTok’s performance. For compression efficiency, we used standard metrics including characters per token, vocabulary utilization (percentage of vocabulary appearing in test sets), and encoding/decoding speed. For downstream task evaluation, we followed standard protocols for each benchmark. Models were trained using the nanoGPT implementation with consistent hyperparameters across conditions. We used three random seeds for each experiment and report average performance. Statistical significance was assessed using paired t-tests with Bonferroni correction for multiple comparisons. Tokenizer Variants: We trained two variants of SupraTok for comprehensive evaluation. The English-focused variant was trained primarily on English text and achieves the reported 6.51 characters per token compression ratio on English benchmarks, representing a 45% improvement over certain baseline tokenizers when vocabulary sizes and training data are optimally matched. The multilingual variant was trained on 38 languages and shows more modest English performance, at 5.91 characters per token and an average of 2.93 characters per token across languages but maintains reasonable cross-lingual capability. Our downstream task evaluation uses the English-focused variant to maximize performance on English benchmarks, while multilingual compression results reflect the multilingual variant’s performance. 3.7 Larger-Scale Validation Framework To address scalability concerns, we outline our planned validation methodology for larger model scales: Progressive Scale Testing: We plan systematic evaluation across model scales (1B, 7B, 13B parameters) using identical architectural configurations, varying only tokenization strategy. Training will use proportionally scaled datasets (100B, 500B, 1T tokens respectively) to maintain training budget ratios. Computational Efficiency Analysis: Validation will include detailed analysis of memory footprint, training throughput, and inference latency across scales. We hypothesize that compression benefits may increase with scale due to reduced attention computation costs, but empirical validation is required. Task Diversification: Evaluation will expand beyond classification to include generation quality metrics (perplexity, BLEU, human evaluation), long-context understanding ( 4K tokens), and domain-specific performance (code, scientific text, multilingual tasks). 3.8 Preliminary Ablation Analysis While comprehensive ablation studies remain future work, we conducted limited analysis of component contributions: Cross-boundary Learning: Disabling cross-boundary pattern learning (reverting to traditional word boundaries) reduces English compression ratio from 5.91 to approximately 4.65 characters per token, suggesting this component contributes 27% of the efficiency gain. Entropy Filtering: Training on unfiltered data (including the 35% of low-entropy content normally removed) yields compression ratios of 5.45 characters per token, indicating entropy curation contributes 8-10% efficiency improvement while potentially providing additional downstream benefits through improved pattern quality. Curriculum Learning: Single-phase training (skipping the three-phase curriculum) results in training instability and suboptimal convergence, achieving only 5.20 characters per token. The curriculum appears essential for stable learning of complex cross-boundary patterns. These preliminary results suggest cross-boundary learning provides the primary efficiency benefit, while entropy filtering and curriculum learning provide supporting contributions essential for training stability and pattern quality. 4 Discussion 4.1 Theoretical Implications The success of SupraTok challenges several assumptions underlying current NLP practices and opens new theoretical questions about language representation in neural networks. The substantial performance improvements achieved solely through tokenization changes suggest that the field has been operating with a suboptimal text representation layer, potentially limiting the realized capabilities of existing architectures. From a linguistic perspective, our results provide computational support for usage-based theories of language that emphasize the role of multi-word units in language processing [21, 22]. The emergence of phrases, idioms, and collocations as atomic tokens through purely statistical learning aligns with psycholinguistic evidence that humans process frequent multi-word sequences as units rather than compositionally [23]. This convergence between computational and cognitive approaches suggests that tokenization strategies aligned with human language processing may yield more effective models. The entropy-driven approach to both data curation and pattern discovery provides a principled, information-theoretic foundation for tokenization. By prioritizing high-entropy content and identifying patterns with specific entropy signatures, SupraTok moves beyond purely frequency-based methods toward a more sophisticated understanding of information content in language. This approach could be extended to other areas of NLP where frequency-based methods currently dominate. Baseline Considerations: While our primary comparison focuses on production tokenizers, we acknowledge that a natural baseline would be standard BPE with equivalent vocabulary size (256k tokens). The cross-boundary learning component represents the key innovation beyond simply scaling vocabulary size. Future work should include systematic comparison against larger-vocabulary standard BPE to isolate the contribution of cross-boundary pattern learning from vocabulary size effects. 4.2 Practical Advantages and Deployment Considerations The immediate practical benefits of SupraTok extend across multiple deployment scenarios. For applications with strict context length limitations, the 31% compression improvement effectively extends accessible context without architectural changes. This is particularly valuable for resource-constrained deployments where model size and memory usage are critical constraints. The improved handling of named entities and technical terminology has important implications for domain-specific applications. In medical, legal, and scientific domains where precise terminology is crucial, SupraTok’s preservation of domain terms as atomic units could reduce errors and improve consistency. Our analysis shows that models using SupraTok make fewer errors on entity-related tasks and show better consistency in handling technical terms across different contexts. The curriculum learning approach addresses a key challenge in developing new tokenization strategies: ensuring stable training while introducing novel patterns. By progressively relaxing constraints, SupraTok achieves convergence to complex patterns that single-phase approaches fail to discover. This methodology could be applied to other areas where curriculum learning has shown promise but implementation remains challenging [24, 25]. 4.3 Limitations and Challenges Several important limitations warrant careful consideration. Scale Limitations: Our evaluation focuses on 124M parameter models, and scalability to larger models (1B+ parameters) remains to be validated. The benefits observed at smaller scales may not transfer proportionally to production-scale deployments where most practical applications operate. Fair Comparison Challenges: While we include comparisons with same-vocabulary-size tokenizers (Gemma 3), our entropy-driven filtering removes 35% of training data, which could contribute to performance gains independent of tokenization improvements. A controlled study with identical training data would strengthen the evaluation. Component Contribution Analysis: We have not conducted systematic ablation studies to quantify the individual contributions of cross-boundary learning, entropy filtering, and curriculum learning components. Such analysis would clarify which innovations provide the most benefit and guide future development priorities. Training Complexity and Cost: The three-phase curriculum and entropy-based filtering significantly increase training complexity and computational requirements compared to standard BPE. Training requires 40 hours on 8x A100 40GB GPUs, substantially more than traditional approaches. This complexity may limit adoption and experimentation. Evaluation Scope: Our downstream evaluation focuses on classification benchmarks (HellaSWAG, MMLU) and omits generation quality assessment. Text generation capabilities, particularly handling of multi-word token boundaries during generation, require further investigation. Generation Implications and Challenges: The use of multi-word tokens introduces several generation-specific considerations that warrant discussion. During autoregressive generation, models must commit to complete multi-word tokens before seeing subsequent context, potentially leading to suboptimal choices when the intended continuation would favor different tokenization boundaries. For example, committing to "New_York" when the intended output is "New Zealand" creates recovery challenges. Preliminary analysis suggests this issue may be mitigated by the model’s exposure to token internals during training and the statistical frequency of learned patterns, but systematic evaluation of generation fluency, consistency, and recovery from tokenization-induced errors remains necessary. The benefits of reduced sequence length and preserved semantic units must be weighed against potential generation boundary effects, particularly in creative or unpredictable text generation scenarios. Language-Specific Optimization: While showing competitive multilingual performance, our approach uses English-optimized entropy thresholds and PMI calculations that may not transfer optimally to other languages, particularly those with different morphological structures. 4.4 Future Research Directions Our work opens several promising avenues for future research. Neural-guided pattern discovery could move beyond statistical co-occurrence to incorporate learned representations. By using pre-trained language models to guide tokenization decisions, we might discover patterns that are semantically rather than just statistically coherent. This could lead to tokenizers that adapt to the specific capabilities and biases of different model architectures. Dynamic tokenization that adjusts to content during inference presents another frontier. Rather than using a fixed vocabulary, models could adapt their tokenization based on the domain, task, or even specific input. This could be particularly valuable for specialized domains where optimal tokenization differs substantially from general text. Cross-modal tokenization for multimodal models requires rethinking tokenization beyond text. As vision-language models become prevalent, developing tokenization strategies that can represent both modalities in a unified vocabulary could improve cross-modal understanding. SupraTok’s approach to discovering patterns could extend to visual tokens or audio segments. Tokenization-aware architectures designed to work optimally with cross-boundary tokens could further amplify the benefits we observe. Current architectures treat all tokens equally, but models could potentially benefit from knowing whether a token represents a character, subword, word, or phrase. Incorporating this information into position encodings or attention mechanisms could improve performance. 4.5 Broader Implications The success of SupraTok has implications beyond immediate performance gains. It demonstrates that fundamental components of the NLP pipeline, sometimes considered "solved" problems, still harbor substantial room for improvement. This suggests that revisiting other foundational assumptions could yield similar gains. From an environmental perspective, improved efficiency translates directly to reduced computational requirements and energy consumption. A 31% reduction in sequence length could substantially decrease the carbon footprint of large-scale model training and inference. As the field grapples with the environmental impact of ever-larger models, efficiency improvements like SupraTok offer a path to progress that doesn’t require increasing scale. The democratization of AI capabilities could be accelerated by such efficiency improvements. Models that achieve better performance with smaller computational budgets become accessible to a broader range of researchers and organizations. This could help address concerns about AI development being concentrated in a few well-resourced institutions. 5 Conclusion SupraTok represents a novel approach to tokenization that demonstrates potential for improved efficiency through cross-boundary pattern learning. By removing traditional word-boundary constraints, implementing entropy-driven training strategies, and employing curriculum learning, our approach achieves notable improvements in compression efficiency and, at the 124M parameter scale tested, downstream task performance. The substantial improvements in tokenization efficiency—ranging from 31% over production tokenizers to 45% under optimal configurations—and 8-10% gains on language understanding benchmarks against controlled baselines suggest that tokenization innovations can meaningfully complement architectural advances. The preliminary ablation analysis indicates that cross-boundary learning provides the primary efficiency benefit ( 27%), while entropy filtering and curriculum learning contribute essential supporting improvements for training stability and pattern quality. While these results are promising at the 124M parameter scale evaluated, validation at larger model scales remains necessary to establish broader applicability. Our findings indicate that addressing fundamental components like tokenization may provide additional paths to language model improvement beyond traditional scaling approaches. Beyond immediate performance gains, SupraTok provides theoretical insights into language representation and practical tools for improving model efficiency. The success of cross-boundary tokenization validates linguistic theories about the importance of multi-word units while offering a concrete implementation that can be immediately deployed. The entropy-driven approach provides a principled framework that could be applied to other areas of NLP where information-theoretic perspectives remain underexplored. As language models continue to evolve and deploy across diverse applications, efficient tokenization becomes increasingly critical. SupraTok demonstrates that tokenization is not merely a preprocessing step but a fundamental component that significantly impacts model capabilities. Our work suggests that the path to more capable language models need not always lead through increased scale but can also proceed through more intelligent representation of the underlying data. Looking forward, the principles underlying SupraTok—crossing artificial boundaries, entropy-driven optimization, and curriculum-based learning—could be applied to other fundamental components of language models. As the field matures from rapid scaling toward efficient and principled design, we anticipate that such foundational improvements will play an increasingly important role in achieving human-level language understanding. Acknowledgments The training infrastructure was provided through an Amazon Web Services (AWS) sponsorship to the Faculty of Mathematics and Computer Science at "Ovidius" University of Constanta, which granted access to P4D24XLarge instances at no cost for academic research purposes. Code and Data Availability Source code for SupraTok training and evaluation will be made available under an open-source license. Training corpora are derived from publicly available sources (Common Crawl, HuggingFace). Processed datasets and evaluation benchmarks are available upon request. References - [1] Brown, T. et al. Language models are few-shot learners. Adv. Neural Inf. Process. Syst. 33, 1877–1901 (2020). - [2] Kaplan, J. et al. Scaling laws for neural language models. Preprint at https://arxiv.org/abs/2001.08361 (2020). - [3] Hoffmann, J. et al. Training compute-optimal large language models. Preprint at https://arxiv.org/abs/2203.15556 (2022). - [4] Radford, A. et al. Language models are unsupervised multitask learners. OpenAI Blog 1, 9 (2019). - [5] Bubeck, S. et al. Sparks of artificial general intelligence: Early experiments with GPT-4. Preprint at https://arxiv.org/abs/2303.12712 (2023). - [6] Hestness, J. et al. Deep learning scaling is predictable, empirically. Preprint at https://arxiv.org/abs/1712.00409 (2017). - [7] Vaswani, A. et al. Attention is all you need. Adv. Neural Inf. Process. Syst. 30 (2017). - [8] Gao, L. et al. The Pile: An 800GB dataset of diverse text for language modeling. Preprint at https://arxiv.org/abs/2101.00027 (2020). - [9] Gage, P. A new algorithm for data compression. C Users J. 12, 23–38 (1994). - [10] Sennrich, R., Haddow, B. & Birch, A. Neural machine translation of rare words with subword units. In Proc. 54th Annual Meeting ACL 1715–1725 (2016). - [11] Wu, Y. et al. Google’s neural machine translation system. Preprint at https://arxiv.org/abs/1609.08144 (2016). - [12] Kudo, T. & Richardson, J. SentencePiece: A simple and language independent subword tokenizer. In Proc. 2018 Conf. EMNLP 66–71 (2018). - [13] Kudo, T. Subword regularization: Improving neural network translation models. In Proc. 56th Annual Meeting ACL 66–75 (2018). - [14] Salehi, B., Cook, P. & Baldwin, T. A word embedding approach to predicting the compositionality of multiword expressions. In Proc. 2015 Conf. NAACL 977–983 (2015). - [15] Rust, P. et al. How good is your tokenizer? On the monolingual performance of multilingual language models. In Proc. 59th Annual Meeting ACL 4034–4047 (2021). - [16] Rumbelow, J. & Watkins, M. SolidGoldMagikarp: Glitch tokens in GPT-2. LessWrong https://lesswrong.com/posts/glitch-tokens (2023). - [17] Pagnoni, A. et al. Byte latent transformer: Patches scale better than tokens. Preprint at https://arxiv.org/abs/2412.09871 (2024). - [18] Clark, J. et al. ByT5: Towards a token-free future with pre-trained byte-to-byte models. Trans. Assoc. Comput. Linguist. 10, 291–306 (2022). - [19] Xue, L. et al. ByT5: Towards a token-free future with pre-trained byte-to-byte models. In Proc. 2022 Conf. NAACL 291–306 (2022). - [20] Liu, A., Hayase, J., Hofmann, V., Oh, S., Smith, N. A. & Choi, Y. SuperBPE: Space travel for language models. Preprint at https://arxiv.org/abs/2503.13423 (2025). - [21] Sinclair, J. Corpus, Concordance, Collocation (Oxford University Press, 1991). - [22] Wray, A. Formulaic Language and the Lexicon (Cambridge University Press, 2002). - [23] Arnon, I. & Snider, N. More than words: Frequency effects for multi-word phrases. J. Mem. Lang. 62, 67–82 (2010). - [24] Bengio, Y., Louradour, J., Collobert, R. & Weston, J. Curriculum learning. In Proc. 26th Int. Conf. Machine Learning 41–48 (2009). - [25] Soviany, P., Ionescu, R. T., Rota, P. & Sebe, N. Curriculum learning: A survey. Int. J. Comput. Vis. 130, 1526–1565 (2022). - [26] Schmidt, C. W., Reddy, V., Tanner, C. & Pinter, Y. Boundless byte pair encoding: Breaking the pre-tokenization barrier. Preprint at https://arxiv.org/abs/2504.00178 (2025).