README.md

---
language:
- en
license: mit
pipeline_tag: text-generation
library_name: transformers
tags:
- rebalance
- steering-vector
- reasoning
- llm
- iclr-2026
---

<h1 align="center">ReBalance Steering Vectors</h1>

<p align="center">
  Steering vectors for <strong>Efficient Reasoning with Balanced Thinking</strong> (ICLR 2026)
</p>

<p align="center">
  <a href="https://huggingface.co/papers/2603.12372"><img src="https://img.shields.io/badge/Paper-Hugging_Face-b31b1b.svg" alt="Paper (Hugging Face)"></a>
  <a href="https://openreview.net/forum?id=cJseWJJ5IM"><img src="https://img.shields.io/badge/Paper-Open_Review-8D1B12.svg" alt="Paper (Open Review)"></a>
  <a href="https://github.com/yu-lin-li/ReBalance"><img src="https://img.shields.io/badge/Code-GitHub-black.svg" alt="Code (GitHub)"></a>
  <a href="https://rebalance-ai.github.io"><img src="https://img.shields.io/badge/Project-Page-2EA44F.svg" alt="Project Page"></a>
  <a href="https://github.com/yu-lin-li/ReBalance/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-MIT-yellow.svg" alt="License: MIT"></a>
</p>

## Description
This repository contains the steering vectors presented in the paper **"Efficient Reasoning with Balanced Thinking"** (ICLR 2026).

**Authors**: Yulin Li, Tengyao Tu, Li Ding, Junjie Wang, Huiling Zhen, Yixin Chen, Yong Li, Zhuotao Tian.

ReBalance is a training-free framework that achieves efficient reasoning by identifying and mitigating "overthinking" (redundant computational steps) and "underthinking" (insufficient exploration) in Large Reasoning Models (LRMs). It leverages confidence as a dynamic indicator to modulate reasoning trajectories using the steering vectors provided here.

## Overview
This repository provides **steering vectors only** for ReBalance.

It does **not** include:
- base model weights,
- inference/training code,
- datasets.

Use this repository together with the official codebase:
- Code: https://github.com/yu-lin-li/ReBalance
- Inference script: `transformer_inference_steer_dp.py`

## Available Steering Vectors
| Base model | Vector path | Recommended `--steer_layer` |
|---|---|---|
| DeepSeek-R1-Distill-Qwen-1.5B | `vectors/DeepSeek-R1-Distill-Qwen-1.5B/steer_vector_layer19_conf_mixed.pt` | `19` |
| DeepSeek-R1-Distill-Qwen-7B | `vectors/DeepSeek-R1-Distill-Qwen-7B/steer_vector_layer22_conf_mixed.pt` | `22` |
| QwQ-32B | `vectors/QwQ-32B/steer_vector_layer58_conf_mixed.pt` | `58` |

## Local Directory Layout (after download)
```text
ReBalance/
├── transformer_inference_steer_dp.py
└── vectors/
    ├── DeepSeek-R1-Distill-Qwen-1.5B/
    │   └── steer_vector_layer19_conf_mixed.pt
    ├── DeepSeek-R1-Distill-Qwen-7B/
    │   └── steer_vector_layer22_conf_mixed.pt
    └── QwQ-32B/
        └── steer_vector_layer58_conf_mixed.pt
```

## Download
### Option 1: Clone the full model repository
```bash
git lfs install
git clone https://huggingface.co/Yulin-Li/ReBalance
```

Then copy the `vectors/` folder into your local `ReBalance/` root directory.

### Option 2: Download only the vectors with `huggingface_hub`
```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Yulin-Li/ReBalance",
    repo_type="model",
    allow_patterns="vectors/*",
    local_dir="."
)
```

## Quick Usage with ReBalance
```bash
python transformer_inference_steer_dp.py \
  --model_name_or_path 'DeepSeek-R1-Distill-Qwen-1.5B' \
  --dataset_dir "./Data/" \
  --output_path "./outputs" \
  --dataset "Math_AIME2024" \
  --max_generated_tokens 16000 \
  --num_gpus 8 \
  --steer_vector_path ./vectors/DeepSeek-R1-Distill-Qwen-1.5B/steer_vector_layer19_conf_mixed.pt \
  --steer_layer 19 \
  --steer_coef -1
```

## Paper, Code, and Project Links
- Paper (Hugging Face): https://huggingface.co/papers/2603.12372
- Paper (Open Review): https://openreview.net/forum?id=cJseWJJ5IM
- Code: https://github.com/yu-lin-li/ReBalance
- Project page: https://rebalance-ai.github.io

## Intended Use
- Research and reproducibility for ReBalance.
- Experiments on reasoning efficiency and accuracy trade-offs.
- Comparative studies on overthinking mitigation.

## Citation
If you find ReBalance useful in your research, please cite our paper:

```bibtex
@article{li2026efficient,
  title={Efficient Reasoning with Balanced Thinking},
  author={Li, Yulin and Tu, Tengyao and Ding, Li and Wang, Junjie and Zhen, Huiling and Chen, Yixin and Li, Yong and Tian, Zhuotao},
  booktitle={Proceedings of the 14th International Conference on Learning Representations},
  year={2026}
}
```