CUE

CUE is a user-simulator framework that can utilize any LLM to generate messages as realistic users. Based on a user's dialogue history, CUE produces a persona manual: a short list of behavioral commands ("keep requests terse", "ask about price before committing") that can be handed to the system prompt of any LLM so it role-plays that user. Underneath the hood, a chat session from user is encoded into a single 1024-d CUE embedding, and a decoder turns that embedding back into text commands. We call this user-conditioned CUE.

You can alternatively use CUE to sample a persona manual representing a novel user without any existing user dialogue.

This Hugging Face model is a general checkpoint from joint-training on this dataset described in our paper:

CUEing User Simulators: Calibrated User Embeddings for Multi-Turn Benchmarking

Usage

The repo carries its own code, so no install is required:

from transformers import AutoModel

model = AutoModel.from_pretrained("handshake-ai-research/cue", trust_remote_code=True)

session = [[
    {"role": "user", "content": "hey can you tighten this paragraph"},
    {"role": "assistant", "content": "Sure — here is a shorter version."},
    {"role": "user", "content": "shorter, and less formal"},
]]

embedding = model.encode(session)                    # (1, 1024) CUE embedding
manual = model.generate_manual(sessions=session)[0]  # {"commands": [...], "examples": [...]}

For the CLI, the parity tests, or export tooling, install the package instead:

pip install "git+https://github.com/Handshake-AI-Research/CUE-user-simulator"

Sampling synthetic users

sampler.pt (647 MB) in this repo is a latent diffusion prior over CUE embeddings, so you can draw users that no transcript describes:

out = model.sample_user(n=4, seed=0)   # {"embeddings": (4, 1024), "manuals": [...]}

Session preprocessing

session_preprocess defaults to full (turns as-is) for all inputs. Pass strip_document yourself for document-heavy inputs such as document editing, where an assistant turn carries a whole draft that would otherwise dominate the encoded session:

embedding = model.encode(session, session_preprocess="strip_document")

Example retrieval (optional)

config.example_pool_id points at handshake-ai-research/cue-example-pool, which lets the decoder ground commands in real turns from similar users. It needs the retrieval extra (pip install "cue-hf[retrieval]", i.e. datasets plus optional faiss):

manual = model.generate_manual(sessions=session, example_retrieval=True)[0]

Architecture

Both backbones were finetuned during joint training, so model.safetensors (3.03 GB, 757 tensors) holds every weight. The base repos above are used only for their architecture configs and tokenizers.

Citation

@article{kantharuban2026cue,
  title={CUEing User Simulators: Calibrated User Embeddings for Multi-Turn Benchmarking},
  author={Kantharuban, Anjali and Mueller, Jonas},
  journal={arXiv preprint arXiv:2610.02460},
  year={2026}
}