Unnatural Languages Are Not Bugs but Features for LLMs
Fuente:
arXiv
Enregistré dans:
| Auteurs principaux: | , , , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Publié: |
2025
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866910981748162560 |
|---|---|
| author | Duan, Keyu Zhao, Yiran Feng, Zhili Ni, Jinjie Pang, Tianyu Liu, Qian Cai, Tianle Dou, Longxu Kawaguchi, Kenji Goyal, Anirudh Kolter, J. Zico Shieh, Michael Qizhe |
| author_facet | Duan, Keyu Zhao, Yiran Feng, Zhili Ni, Jinjie Pang, Tianyu Liu, Qian Cai, Tianle Dou, Longxu Kawaguchi, Kenji Goyal, Anirudh Kolter, J. Zico Shieh, Michael Qizhe |
| contents | Large Language Models (LLMs) have been observed to process non-human-readable text sequences, such as jailbreak prompts, often viewed as a bug for aligned LLMs. In this work, we present a systematic investigation challenging this perception, demonstrating that unnatural languages - strings that appear incomprehensible to humans but maintain semantic meanings for LLMs - contain latent features usable by models. Notably, unnatural languages possess latent features that can be generalized across different models and tasks during inference. Furthermore, models fine-tuned on unnatural versions of instruction datasets perform on-par with those trained on natural language, achieving 49.71 win rates in Length-controlled AlpacaEval 2.0 in average across various base models. In addition, through comprehensive analysis, we demonstrate that LLMs process unnatural languages by filtering noise and inferring contextual meaning from filtered words. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2503_01926 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Unnatural Languages Are Not Bugs but Features for LLMs Duan, Keyu Zhao, Yiran Feng, Zhili Ni, Jinjie Pang, Tianyu Liu, Qian Cai, Tianle Dou, Longxu Kawaguchi, Kenji Goyal, Anirudh Kolter, J. Zico Shieh, Michael Qizhe Computation and Language Artificial Intelligence Large Language Models (LLMs) have been observed to process non-human-readable text sequences, such as jailbreak prompts, often viewed as a bug for aligned LLMs. In this work, we present a systematic investigation challenging this perception, demonstrating that unnatural languages - strings that appear incomprehensible to humans but maintain semantic meanings for LLMs - contain latent features usable by models. Notably, unnatural languages possess latent features that can be generalized across different models and tasks during inference. Furthermore, models fine-tuned on unnatural versions of instruction datasets perform on-par with those trained on natural language, achieving 49.71 win rates in Length-controlled AlpacaEval 2.0 in average across various base models. In addition, through comprehensive analysis, we demonstrate that LLMs process unnatural languages by filtering noise and inferring contextual meaning from filtered words. |
| title | Unnatural Languages Are Not Bugs but Features for LLMs |
| topic | Computation and Language Artificial Intelligence |
| url | https://arxiv.org/abs/2503.01926 |