AlienLM mascot
ICML 2026 | API-boundary privacy

Say it in Alien.

AlienLM rewrites your prompt into a private language before it ever reaches a black-box LLM. Same answers, zero plaintext on the wire.

Read the paper → Try the live demo
You type · stays on your device
“▍”
↓  A L I E N I Z E D  ↓
What the API receives
▍
81%+
of plaintext-oracle performance retained on every backbone
4
LLM backbones across LLaMA, Qwen & Gemma families
7
benchmarks: knowledge, reasoning, truthfulness & QA
<0.22%
token recovery across all observer scenarios (O1–O3)
The core idea

One shared token space, two dictionaries.

Your text and its alien twin share the exact same token IDs — only the dictionary that spells them out is swapped. Encode with yours, decode with the alien one, and the surface text turns unreadable while the model still understands every token. The provider processes gibberish; your client quietly reads it back.

AlienLM pipeline: the user's text is alienized by the on-device translator, sent to the API provider and the alienized LLM, then recovered locally.
The on-device translator alienizes your prompt, the black-box provider and its alienized LLM only ever handle alien text, and the reply is recovered to plaintext locally.
How it works

Three moving parts.

Built to deploy with nothing but API-only fine-tuning and public tokenizer information.

1

Build an alien vocabulary

A vocabulary-scale token bijection makes surface strings unreadable while keeping token relationships useful for adaptation.

2

Adapt with AAT

Alien Adaptation Training teaches the target model to operate directly on alienized inputs using standard fine-tuning APIs.

3

Translate locally

The client encodes natural text into alien before the API call, and decodes the alien response right after it.

Live, in your browser

Run the real tokenizer.

Pick an alienization level ρ and watch a prompt cross the API boundary with Llama 3's real tokenizers. Tokens that keep their original surface stay plain; alienized tokens light up. Everything runs client-side; no text ever leaves your browser.

ALIENIZATION LEVEL (ρ)
BACKBONE Llama 3 8B
⇄
0 chars
TRY:
Pick a model pair, then run a translation.

WHAT THE API RECEIVES

UNREADABLE

Run a translation to see the alienized or recovered text.

alienized token kept original

RECOVERED ON YOUR DEVICE

Awaiting a run…

—

SOURCE TOKEN IDS · 0

—

ALIEN-ENCODED IDS · 0

—

Main results

Keeps the smarts, drops the plaintext.

Averages across MMLU, ARC, HellaSwag, WinoGrande, TruthfulQA and GSM8K, plus recovery robustness under three observer scenarios.

Utility across backbones

AlienLM keeps >81% of plaintext-oracle performance while beating random-bijection and character-level baselines.

Backbone Oracle avg. AlienLM avg. Ratio
LLaMA 3 8B64.7752.9281.70%
Qwen 2.5 7B65.6057.3387.40%
Qwen 2.5 14B73.4962.0584.43%
Gemma 2 9B69.2056.6381.83%
Resources

Dig in.

Paper for concepts, Hugging Face for checkpoints, GitHub for tokenizer, training and recovery code. The poster, the figure cast and a BibTeX entry are below.

Poster

ICML 2026 · Poster Session 7 · Hall A #2214
AlienLM ICML 2026 poster
Open PDF → Download (5.7 MB)

Figure kit

The cast of the paper's figures, as transparent PNGs. Reuse them in slides or posts, with a credit to AlienLM (Kim & Kang, ICML 2026).

User illustration
User Writes the plaintext prompt. PNG ↓
Translator illustration
Translator Alienizes and recovers text on the client. PNG ↓
API provider illustration
API provider Sees only alien text. PNG ↓
Alienized LLM illustration
Alienized LLM Fine-tuned to read and write Alien. PNG ↓
Alien dictionary illustration
Alien dictionary The vocabulary-scale bijection. PNG ↓
Alien instruction data illustration
Alien instruction data AAT training examples. PNG ↓

The logo on the API provider's shirt is OpenAI's trademark and is not part of that permission.

Cite

@inproceedings{kim2026alienlm,
  title     = {AlienLM: Alienization of Language for API-Boundary Privacy in Black-Box LLMs},
  author    = {Kim, Jaehee and Kang, Pilsung},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning},
  year      = {2026}
}