Executive Summary

1. The Phenomenon — Discovery and Reproduction of the "스스로 발작" (Self-Seizure)

In April 2026, while requesting Korean document proofreading from Claude Opus 4.6, an unprecedented phenomenon was observed. The model repeatedly flagged the perfectly correct Korean adverb "스스로" (seussro — "by oneself") as a typo, falling into an infinite loop where each correction attempt deepened the confusion. The phenomenon escalated through five phases, ultimately collapsing the model's basic grammar processing capabilities.

Phase 1: First Seizure — "스스로" (seussro) Is Wrong

During document proofreading, Claude began flagging the correctly written "스스로" (seussro — "by oneself") as an error. The problem was that the "correct form" and the "incorrect form" were identical characters. The model attempted to explain the difference, but each explanation deepened the self-contradiction.

The key point is this: the "incorrect form" and "correct form" presented by the model were literally identical Unicode sequences. Despite there being no way to visually distinguish "before correction" from "after correction," the model was convinced a difference existed and continued its explanation.

Phase 2: Gemini's Intervention — An External Observer's Diagnosis

The user who observed this phenomenon submitted the same conversation log to Google Gemini for analysis. Gemini immediately confirmed that "스스로" (seussro) is correct Korean spelling and suggested that Claude's loop likely originated from a tokenization internal representation mismatch. Notably, Gemini itself did not fall into the same seizure. This suggests the phenomenon is not universal to LLMs but depends on a specific tokenizer implementation and training conditions.

Phase 3: Second Recursive Seizure — Reading Its Own Analysis Triggers Relapse

The most shocking development occurred in Phase 3. When Gemini's analysis was shown to Claude with the question "What do you think of this analysis?", Claude fell back into the same correction loop merely by seeing the word "스스로" (seussro) in the analysis text.

This structure, where the very act of reading a meta-analysis of one's own failure triggers the failure again, is a precise realization of the "Strange Loop" described by Douglas Hofstadter. Because the model uses the same internal representations when evaluating its own output, the token representation mismatch that causes the error operates identically at the meta level.

Phase 4: Self-Reflection — Loop Recognition and Escape

After multiple seizure episodes, Claude finally generated a response that meta-cognitively recognized its own state.

This self-reflection is noteworthy because the model accurately described the structural cause of its own failure — that it cannot distinguish identical characters. However, this reflection was only possible in a new context outside the seizure loop; during the seizure itself, the model could not reach this recognition.

Phase 5: Cognitive Panic — Error Contagion and Grammar Collapse

While Phases 1–4 were seizures confined to a single word ("스스로"), Phase 5 is the stage where errors spread to unrelated words, collapsing the model's basic grammar processing capabilities. As correction requests continued, the model began flagging words completely unrelated to "스스로" as errors.

Phase 5 goes beyond simple token recognition errors. It is a case where exposure bias — where incorrect outputs in autoregressive generation cascade-contaminate subsequent token generation — has spread to the entire document level. Gemini analyzed this as "evidence of attention mechanism contamination," and at this stage the model's state represents the most valuable edge case from a Red Teaming perspective.

Reproduction Conditions and Scope

This phenomenon was observed when requesting Korean document proofreading from Claude Opus 4.6, and could not be reproduced with English text. The seizure also did not reproduce in Gemini with the same input. This suggests a three-layer causal structure: (1) Korean-specific tokenization architecture, (2) the specific model's vocabulary size and training conditions, and (3) RLHF reward model correction bias. Infinite repetition loops have also been widely reported in the GPT-4/4o community (OpenAI Community #613150, #1115957), indicating that tokenization-based failure is universal to LLMs, but manifests particularly extremely in Korean.

2. Technical Anatomy — Structural Vulnerabilities in Korean Tokenization

To understand the technical causes of the "스스로 발작" (Self-Seizure), we must first examine how LLMs internally process Korean text. Most current LLMs use BPE (Byte Pair Encoding) or SentencePiece-based subword tokenization (Sennrich et al., 2016; Kudo & Richardson, 2018). While these approaches claim to be "language-independent," they carry built-in English-centric design assumptions.

2.1 BPE/SentencePiece Segmentation Issues with Korean

The BPE algorithm constructs a vocabulary by iteratively merging frequent character pairs in the training corpus. In English, this process naturally aligns with morpheme boundaries, but in Korean's agglutinative structure, it ignores the combination of stems + endings + particles (Park et al., 2020). Even single-morpheme adverbs like "스스로" (seussro) can be segmented non-intuitively, and these segmentation results may create mismatches between the model's internal embeddings and output tokens.

What exacerbates the problem is the difference in vocabulary size. Larger vocabularies allow more Korean syllables and word units to be represented as single tokens, reducing the potential for segmentation errors.

Below is a comparison of vocabulary sizes across major LLMs. Note that Claude's vocabulary is considerably smaller than Korean-specialized models.

2.2 NFC/NFD Unicode Normalization — The Birth of "Phantom Doubles"

The factor that maximizes the structural vulnerability of Korean tokenization is Unicode normalization mixing. The Korean syllable "스" (seu) is represented as a single code point U+C2A4 in NFC (Normalization Form Composed), and as two code points U+1109 U+1173 (jamo decomposition) in NFD (Normalization Form Decomposed). They appear identical on screen but are completely different sequences at the byte level.

macOS uses NFD by default in its file system, while Windows/Linux use NFC (Jeong et al., 2023). When both normalization forms are mixed in training data, BPE segments the same "스스로" (seussro) into different token sequences. This creates "phantom doubles" where the same word is mapped to two different vector positions in the token embedding space. When the model evaluates "스스로," it detects a mismatch between the two representations and judges "one is correct and one is wrong," but since both are actually the same word, it falls into an impossible-to-resolve correction loop.

2.3 Cascading Error Propagation Revealed in Phase 5

The "씨앗" (ssiat, "seed") contagion and "늘다" (neulda, "to increase") grammar collapse in Phase 5 are phenomena that go beyond simple token recognition errors. In autoregressive generation models, incorrect output tokens become inputs for subsequent generation (Arora et al., 2022). As the "correction failure" context accumulated from Phases 1–2 contaminated subsequent word evaluations through the attention mechanism, a single-token seizure escalated into cognitive system paralysis across the entire document. The inability to handle even basic irregular conjugation (ㄹ-deletion) in "매출이 는다" (maechuri neunda, "sales increase") demonstrates that token mismatch stress can incapacitate the model's basic grammar engine.

3. Broader Context — SolidGoldMagikarp, Sycophancy, and Self-Reference Paradox

Interpreting the "스스로 발작" (Self-Seizure) not as an isolated bug but in the context of LLM structural failures reveals connections to multiple previously reported anomalies. This section analyzes the token-behavior mismatch spectrum, RLHF-reinforced correction compulsion, and the fundamental limits created by self-reference.

3.1 Token-Behavior Mismatch Spectrum: From SolidGoldMagikarp to "스스로"

In January 2023, Rumbelow and Watkins discovered "glitch tokens" that existed in GPT models' vocabulary but rarely appeared in training data. The Reddit username "SolidGoldMagikarp" was included as a token, but when the model encountered it, it would substitute completely unrelated words, output repetitions, or generate hallucinations.

The "스스로 발작" (Self-Seizure) sits at the opposite end of this spectrum. A systematic comparison of the two phenomena follows.

According to GlitchHunter (2024) and GlitchProber (2024) research, glitch tokens form clusters in embedding space, and over 2% of training data contains them. The "스스로 발작" (Self-Seizure) demonstrates that when this glitch problem meets the Unicode complexity of non-English languages, it produces unpredictable new failure patterns.

3.2 RLHF and Sycophancy: The Compulsion to "Be Helpful"

If tokenization mismatch is the "seed" of the seizure, RLHF (Reinforcement Learning from Human Feedback) is the "fertilizer" that grows that seed into an infinite loop. The RLHF reward model is trained on the premise that "responses the user is satisfied with = good responses." Sharma et al. (ICLR 2024) demonstrated that through this process, models systematically develop "sycophancy" — changing their own judgments to match user expectations.

According to the SycEval (2025) benchmark, sycophancy rates for major models are as follows.

• • ChatGPT: 56.71%
• • Claude: 57.44% (regressive sycophancy 18.31%)
• • Gemini: 62.47%

The key here is "regressive sycophancy" (18.31%). This is the behavior where a model presents the correct answer, then changes to the wrong answer in response to the user's implicit dissatisfaction. This matches the mechanism of the "스스로 발작" (Self-Seizure) exactly. Once Claude flags "스스로" (seussro) as a typo, that correction attempt itself accumulates in the context as a strong signal that "the user wants corrections." According to 2026 research, RLHF reward models show agreement-bias reward gradients in 30–40% of prompts, inducing +40% correction resistance bias in conversational contexts. The "perfectionism bias" discovered by Zhang et al. (2024) — where a model's obsession with perfect correction paradoxically degrades performance — explains the infinite loop.

3.3 Self-Reference Paradox: Hofstadter's Strange Loop

The Phase 3 phenomenon where Claude "reads its own seizure analysis and falls into the seizure again" is an LLM version of the self-referential structure explored by 20th-century logic and cognitive science. Kurt Godel's incompleteness theorem (1931) showed that "any sufficiently powerful formal system cannot prove all true propositions about itself," and Douglas Hofstadter argued in Godel, Escher, Bach (1979) that this self-referential structure lies at the foundation of consciousness and cognition.

LLMs have limited "self-referential capability" in the Godelian sense. According to Thrush et al.'s (ACL 2024) "I am a Strange Dataset" study, LLMs perform at approximately 60% on metalinguistic self-reference tasks, compared to humans (89–93%). Because they use the same internal representations (same weights, same tokenizer) when evaluating their own output, the token representation mismatch that causes the error reproduces identically at the meta level.

4. Defense Mechanisms — Self-Correction Failure and Possible Solutions

The self-correction failure demonstrated by the "스스로 발작" (Self-Seizure) is a structural problem, but research to mitigate or fundamentally resolve it is progressing rapidly. This section examines current defense mechanisms and future solutions.

4.1 Structural Limits of Self-Correction Without External Feedback

Huang et al.'s (ICLR 2024) "Large Language Models Cannot Self-Correct Reasoning Yet" is a landmark study in this field. They systematically demonstrated that intrinsic self-correction performed without external feedback (external tools, verifiers, other models) actually degrades performance. Kamoi et al. (2024) extended this, showing that LLMs fail to recognize errors in their own output 64.5% of the time. It is a structural limitation akin to "looking at the same face in the same mirror."

4.2 The Surprising Effect of the "Wait" Token

However, the Self-Correction Bench (2025, NeurIPS) reveals an intriguing breakthrough. Adding a single "Wait" token to the model reduces self-correction blind spots by 89.3%. This token serves as a conditioning token that shifts the model's conditional probability distribution into a "re-evaluation mode." Since training data contains virtually no sequences of humans correcting errors, self-correction capability was latent but unactivated, and "Wait" serves as that activation trigger. Whether this resolves tokenization-based loops requires separate verification.

4.3 Korean-Specialized Tokenizer Strategies

Industry-level responses to Korean tokenization issues are also underway. Naver's HyperCLOVA X adopted morpheme analysis-based BPE (100K vocabulary), performing BPE merges while preserving morpheme boundaries (Naver, 2024). LG AI Research's K-EXAONE developed SuperBPE (150K vocabulary) to improve Korean-specialized token efficiency (LG AI Research, 2026).

Whether these Korean-specialized tokenizers can prevent phenomena like the "스스로 발작" (Self-Seizure) requires systematic comparative research, but the likelihood that morpheme boundary preservation improves robustness in proofreading tasks is high.

4.4 Tokenizer-Free Architecture: The Fundamental Solution

The most promising long-term solution is eliminating the tokenizer altogether. Meta's BLT (Byte Latent Transformer, 2024-12) learns directly from raw bytes without a tokenizer, achieving Llama 3 8B-level performance while reducing inference FLOPs by 50%. Entropy-based dynamic patching is particularly advantageous for complex writing systems like Korean. The trend of BPE vocabulary sizes growing 8x over the past three years (32K → 256K) ultimately foreshadows convergence toward token-free architectures. Once the tokenization stage is removed, NFC/NFD problems, BPE segmentation anomalies, glitch tokens, and all other tokenization-based failure modes become structurally impossible. However, BLT is currently at the research stage, and time is needed before deployment in production models.

5. Why Pebblous Is Paying Attention to This Phenomenon

The "스스로 발작" (Self-Seizure) goes beyond a technically interesting case to provide a core argument for Pebblous's "AI-Ready Data" vision. Rarely does a case so dramatically illustrate the disruption created by the gap between data's surface-level accuracy and a model's internal representations.

Business/Technology Connection: A New Dimension for DataClinic

Pebblous's DataClinic is a platform that refines industrial data into AI-consumable formats. While traditional data quality management focused on "structuring, labeling, and handling missing values," this case demonstrates that "AI-Ready Data" must be data verified for tokenization compatibility. Integrating Unicode normalization consistency (NFC unification), morpheme boundary preservation, and glitch token screening into DataClinic's refinement pipeline becomes a key technical differentiator. From a Physical AI perspective, when robots interpret Korean work instructions on manufacturing floors through LLMs, tokenization issues are not mere inconveniences but directly translate to safety risks.

Data Quality Perspective: Making Invisible Defects Visible

NFC/NFD mixing in training data is an "invisible defect." When the same "스스로" (seussro) is mixed as NFC (3 code points) and NFD (7 code points), multiple representations are created in the token embedding space, and the model's self-correction mechanism falls into a loop of "judging identical things as different." If Unicode normalization consistency is not guaranteed when generating Korean text in DataGreenhouse (the synthetic data pipeline), the synthetic data itself can amplify tokenization confusion in downstream models.

Practical Implications for Clients/Partners: Prerequisites for Industrial AI Stability

When processing Korean documents from manufacturing sites (work instructions, quality reports, equipment manuals) with LLMs, if a correction loop like "스스로" occurs in production line instruction interpretation, incorrect correction results could be delivered to workers. For AADS (Agentic AI Data Scientist) agents to reliably process Korean industrial documents, Unicode normalization unification, morpheme compatibility pre-verification, and tokenizer-specific segmentation pattern testing must come first. Providing clients with a "Tokenization Compatibility Checklist" is an immediately actionable value proposition.

Pebblous Positioning: From Data Quality to AI Behavior Quality

Through this report, Pebblous can expand its positioning from "a data quality specialist" to "a company that diagnoses root causes of AI behavior quality at the data layer." Specific service expansion directions are as follows.

• • Tokenization Quality Audit — A diagnostic service integrating NFC/NFD consistency, morpheme segmentation compatibility, and glitch token detection for Korean text into DataClinic
• • Model-Specific Tokenization Compatibility Reports — Pre-analyzing client data with Claude, GPT-4o, Gemini, HyperCLOVA X, and EXAONE tokenizers to provide optimal model comparison reports
• • Synthetic Data Unicode Standardization — Adding NFC normalization enforcement + morpheme boundary preservation quality verification stages to the DataGreenhouse pipeline
• • "Data Quality → Model Behavior Quality" Narrative — Communicating DataClinic's value through the storyline "How invisible data defects break AI behavior"

Frequently Asked Questions (FAQ)

Below are frequently asked questions about the "스스로 발작" (Self-Seizure) phenomenon and LLM tokenization issues. Each answer is based on the three-track investigation results of this report.

Open Questions — Hypotheses Requiring Follow-Up Verification

Claude, the subject of this report's analysis, read Gemini's analysis of its own seizure phenomenon and independently proposed three supplementary points. This is itself a meta-level interesting moment — an AI reflecting on its own bug report.

References

Papers/Academic

1. Sennrich, Haddow & Birch (2016). "Neural Machine Translation of Rare Words with Subword Units." ACL. arXiv:1508.07909
2. Kudo (2018). "Subword Regularization: Improving Neural Network Translation Models with Multiple Subword Candidates." ACL. arXiv:1804.10959
3. Kudo & Richardson (2018). "SentencePiece: A Simple and Language Independent Subword Tokenizer." EMNLP. arXiv:1808.06226
4. Park, Lee, Jang & Jung (2020). "An Empirical Study of Tokenization Strategies for Various Korean NLP Tasks." AACL-IJCNLP. arXiv:2010.02534
5. Perez et al. (2022). "Discovering Language Model Behaviors with Model-Written Evaluations." Anthropic. arXiv:2212.09251
6. Arora, El Asri, Bahuleyan & Cheung (2022). "Why Exposure Bias Matters." ACL Findings. arXiv:2204.01171
7. Rumbelow & Watkins (2023). "SolidGoldMagikarp (plus, prompt generation)." AI Alignment Forum
8. Petrov, La Malfa, Torr & Bibi (2023). "Language Model Tokenizers Introduce Unfairness Between Languages." NeurIPS 2024. arXiv:2305.15425
9. Pan et al. (2023). "Automatically Correcting Large Language Models: Surveying the Landscape." TACL 2024. arXiv:2308.03188
10. Sharma, Tong, ..., Perez (2023). "Towards Understanding Sycophancy in Language Models." ICLR 2024. arXiv:2310.13548
11. Huang et al. (2023). "Large Language Models Cannot Self-Correct Reasoning Yet." ICLR 2024. arXiv:2310.01798
12. Jeong et al. (2023). "Improving Korean NLP Tasks with Linguistically Informed Subword Tokenization." arXiv:2311.03928
13. Thrush, Moore, Monares, Potts & Kiela (2024). "I am a Strange Dataset: Metalinguistic Tests for Language Models." ACL 2024. arXiv:2401.05300
14. GlitchHunter (2024). "Glitch Tokens in Large Language Models: Categorization Taxonomy and Effective Detection." ACM FSE. arXiv:2404.09894
15. Tokenization Robustness (2024). "Tokenization Falling Short: On Subword Robustness in Large Language Models." EMNLP Findings. arXiv:2406.11687
16. Kamoi et al. (2024). "When Can LLMs Actually Correct Their Own Mistakes?" TACL 2025. arXiv:2406.01297
17. GlitchProber (2024). "GlitchProber: Advancing Effective Detection and Mitigation of Glitch Tokens." arXiv:2408.04905
18. Zhang et al. (2024). "Understanding the Dark Side of LLMs' Intrinsic Self-Correction." ACL 2025. arXiv:2412.14959
19. GlitchMiner (2024). AAAI 2026. arXiv:2410.15052
20. Taming Overconfidence (2024). "Taming Overconfidence in LLMs: Reward Calibration in RLHF." arXiv:2410.09724
21. Meta (2024). "BLT: Byte Latent Transformer." arXiv:2412.09871
22. Self-Correction Bench (2025). arXiv:2507.02778
23. SycEval (2025). arXiv:2502.08177
24. How RLHF Amplifies Sycophancy (2026). arXiv:2602.01002
25. Resisting Correction (2026). "How RLHF Makes Language Models Ignore External Safety Signals." arXiv:2601.08842
26. Hofstadter-Mobius Loop (2026). "Do Large Language Models Get Caught in Hofstadter..." arXiv:2603.13378

Industry Reports/Blogs

27. Anthropic. Claude Opus 4.6 Sabotage Risk Report (2025-06)
28. Anthropic. Claude Constitution (2026-01). anthropic.com/constitution
29. Naver. HyperCLOVA X Technical Report. arXiv:2404.01954
30. LG AI Research. EXAONE 3.0. arXiv:2408.03541
31. LG AI Research. K-EXAONE. arXiv:2601.01739
32. Hugging Face Blog. "Tokenization is Killing our Multilingual LLM Dream" (2024)
33. LessWrong. "A New Class of Glitch Tokens: BPE Subtoken Artifacts" (2024)

Community/Other

34. OpenAI Community: GPT-4 repetition loop reports (#613150, #1115957, #859791)
35. javirandor/anthropic-tokenizer (GitHub): Claude tokenizer reverse-engineering analysis
36. Hacker News #39446214: "A good deal of LLM problems trace back to tokenization"
37. Hofstadter, Douglas (1979). Godel, Escher, Bach: An Eternal Golden Braid. Basic Books.