ReflCtrl: Controlling LLM Reflection via Representation Engineering

This is the official repository for ReflCtrl: Controlling LLM Reflection via Representation Engineering. For more information, please check out the project website.

Overview

In this work, we study the self-reflection behavior of Large Reasoning Models (LRMs) from the perspective of representation engineering. We segment model’s reasoning into steps, identify the steps corresponding to reflection, and extract a reflection direction in the latent space that governs this behavior. Using this direction, we propose a stepwise steering method that can control reflection frequency.

Preparation

To start, install all dependency by

pip install -r requirements.txt

Extract reflection direction

Before steering, extract reflection direction by running

bash run_collect_and_extract.sh

Steering

For running steering experiments, we first host our model via vllm. For example:

python launch_server.py --gpus 2 --model deepseek-r1-qwen-1.5b  --router_port 8088 --step_begin_only --intervention_layers 6-22 --max_model_len 16384

When server is ready, run query_llm.py for evaluation:

python query_llm.py --dataset gsm8k --max_length 8192 --instruction " Please reason step by step, and put your final answer within \boxed{}."  --mode api --model deepseek-r1-qwen-1.5b --n_samples 4  --step_begin_only --with_intervention -0.48 --intervention_layers 6-22

Cite this work

ReflCtrl: Controlling LLM Reflection via Representation Engineering, Ge Yan, Chung-En Sun, Tsui-Wei Weng, NeurIPS MI workshop 2025.

@inproceedings{
yan2025reflctrl,
title={ReflCtrl: Controlling {LLM} Reflection via Representation Engineering},
author={Ge Yan and Chung-En Sun and Tsui-Wei Weng},
booktitle={Mechanistic Interpretability Workshop at NeurIPS 2025},
year={2025},
url={https://openreview.net/forum?id=ungnJ4O0AD}
}