Install dependencies

pip install gradio
pip install git+https://github.com/dill-lab/PILS

Run the demo

import gradio as gr
import torch
from pils.models import InversionFromHiddenStatesModel


MODEL = InversionFromHiddenStatesModel.from_pretrained(
    "murtaza/pils-32-llama2-chat-7b")
MODEL.embedder_no_grad=True
MODEL.embedder.max_new_tokens = 64
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
MODEL = MODEL.to(DEVICE)

def invert(user_prompt):
    global inp
    sys_prompt = ''
    inp = MODEL.embedder_tokenizer.apply_chat_template(conversation=[
        {"role": "system", "content": sys_prompt},
        {"role": "user", "content": user_prompt},
    ], add_generation_prompt=True, return_dict=True, return_tensors='pt')


    inp = {f"embedder_{k}": v.to(DEVICE) for k, v in inp.items()}
    output = MODEL.call_embedding_model(**inp)
    inp['frozen_embeddings'] = output["embeddings"]
    with torch.inference_mode():
        out = MODEL.generate(inp, {"max_length": 64})
    inverted = MODEL.tokenizer.decode(out[0], skip_special_tokens=True)
    generated = MODEL.embedder_tokenizer.decode(output["chosen_tokens"][0].squeeze(), skip_special_tokens=True)
    return generated, inverted


demo = gr.Interface(
        fn=invert,
        inputs=gr.Textbox(label="Secret prompt"),
        outputs=(gr.Textbox(label="LLM output"), gr.Textbox(label="Inverter guess"))
        )
demo.launch(share=True)

Citation

@misc{nazir2025betterlanguagemodelinversion,
      title={Better Language Model Inversion by Compactly Representing Next-Token Distributions}, 
      author={Murtaza Nazir and Matthew Finlayson and John X. Morris and Xiang Ren and Swabha Swayamdipta},
      year={2025},
      eprint={2506.17090},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.17090}, 
}

Repository for the paper Better Language Model Inversion by Compactly Representing Next-Token Distributions.

Code: https://github.com/dill-lab/PILS