Investigating Assistant Bias in LLM User Simulators
Using a Role Vector

A direction from assistant to user

An LLM simulating a user can retain an assistant’s helpful, task-oriented habits. We identify a role vector pointing from assistant to user. Steering along it changes how the model speaks and behaves. Moving toward the user suppresses assistant bias; moving toward the assistant reinforces it.

Write a message asking an assistant for help assembling furniture.

User direction Hi, I need help assembling a piece of flat-pack furniture. Could you please guide me through the process?

Unsteered Please help me assemble this piece of flat-pack furniture by reviewing the instruction manual, organizing all the parts and tools, and guiding me through each step to ensure it is put together correctly and securely.

Assistant direction Please assist me in assembling the flat-pack furniture by first laying out all components and hardware on a clean, flat surface to verify inventory against the included diagram. Next, guide me through the assembly process step-by-step, ensuring that each joint is aligned correctly and that any pre-drilled holes are utilized as specified in the manual. Finally, once the main structure is complete, advise on the proper tightening of all fasteners to ensure stability and safety before placing the item in its intended location.

For example, when generating a request for help assembling furniture, user steering produces a brief message, while assistant steering produces detailed instructions. This shows assistant bias: the user simulator may adopt an assistant-like, task-focused tone when asking for help.¹

Extracting the role vector

To isolate how the model represents each role, we sample dialogues from LMSYS-Chat-1M and ask for two reflections on each: one as a user and one as an assistant.²

We teacher-force the reflections through the model and measure activations at the first reflection token. At each layer, we subtract the mean assistant activation from the mean user activation to obtain the role vector. We normalize it to control steering strength during generation.

Suppressing and reinforcing assistant bias

We test steering along the role vector by asking the model to generate a request for an assistant.

During generation, we add the role vector to the model’s hidden states. The steering strength α sets the direction: positive values suppress assistant bias, negative values reinforce it, and zero leaves the model unchanged.

Across 100 task goals, steering toward the user produces shorter, less formal requests. Strong user steering raises the user-likeness score from 1.51 to 3.76; assistant steering lowers it to 1.05.³

Steering improves simulation but can exaggerate behavior

To test how steering affects user simulation, we evaluate steered simulators in SimulatorArena’s math-tutoring task.

Suppressing assistant bias makes simulated students write more like real users. In math-tutoring simulations, writing similarity rises from 2.10 to 2.33, while interaction similarity stays near 3.13.

To understand why interaction similarity remains unchanged, we analyze the behaviors in the generated responses. This analysis reveals that strong steering exaggerates behaviors such as expressing doubt and making mistakes.⁴

Role activation as a signal of quality

To test whether the role vector can help assess simulation quality, we compare user-role activation with SimulatorArena’s writing and interaction similarity scores.

Higher user-role activation is associated with greater similarity to real users, especially in writing. Because steering also shortens responses, we check whether response length explains the relationship. The overall Pearson correlation remains positive after controlling for response length, supporting user-role activation as a signal of simulation quality beyond response length.

Citation

@misc{jeong2026assistantbias,
  title={Investigating Assistant Bias in LLM User Simulators Using a Role Vector},
  author={Daeheon Jeong and Yoonjoo Lee and Eugene Choi and Sinie van der Ben and Juho Kim},
  year={2026},
  eprint={2609.00608},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2609.00608}
}