| --- |
| license: mit |
| --- |
| |
| # KTAE: A Model-Free Algorithm to Key-Tokens Advantage Estimation in Mathematical Reasoning |
|
|
| <div align="center"> |
| <br> |
| <a>Wei Sun</a>, |
| <a>Wen Yang</a>, |
| <a>Pu Jian</a>, |
|
|
| <a>Qianlong Du</a>, |
| <a>Fuwei Cui</a>, |
| <a>Shuo Ren</a>, |
| <a>Jiajun Zhang</a> |
| <br> Institute of Automation Chinese Academy of Sciences <br> |
| |
|   <a href='https://arxiv.org/abs/2505.16826'><img src='https://img.shields.io/badge/Paper-Arxiv-red'></a> <a href='https://github.com/xiaolizh1/KTAE'><img src='https://img.shields.io/badge/Project-Github-red'></a> |
| </div> |
| |
| ## π Overview |
| Recent advances have demonstrated that integrating reinforcement learning with rule-based rewards can significantly enhance the reasoning capabilities of large language models (LLMs), even without supervised fine-tuning (SFT). However, prevalent reinforcement learning algorithms such as GRPO and its variants like DAPO, suffer from a coarse granularity issue when computing the advantage. Specifically, they compute rollout-level advantages that assign identical values to every token within a sequence, failing to capture token-specific contributions. To address this limitation, we propose Key-token Advantage Estimation ($\textit{KTAE}$)βa novel algorithm that estimates fine-grained, token-level advantages without introducing additional models. KTAE leverages the correctness of sampled rollouts and applies statistical analysis to quantify the importance of individual tokens within a sequence to the final outcome. This quantified token-level importance is then combined with the rollout-level advantage to obtain a more fine-grained token-level advantage estimation. Empirical results show that models trained with GRPO+KTAE and DAPO+KTAE outperform baseline methods across five mathematical reasoning benchmarks. Notably, they achieve higher accuracy with shorter responses and even surpass R1-Distill-Qwen-1.5B using the same base model. |
|
|
| <p align="center"> |
| <img src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/-dygcr2iG28gI6jRh0KG1.png width="100%" height="100%"> |
| </p> |
| |
|
|
| In summary, the KTAE algorithm offers several advantages: |
|
|
| + KTAE provides more fine-grained advantage information without introducing extra models, resulting in lower training costs. |
|
|
| + KTAE directly computes the importance differences between tokens using statistical analysis methods, offering strong interpretability. |
|
|
| + KTAE's key-token value is computed based on the correctness of the final answer and retains the original rollout-level advantage, making it less susceptible to reward hacking. |
|
|
| + KTAE can make the model pay more attention to key tokens and reduce the learning of irrelevant tokens, which can effectively reduce the response length. |
|
|
| ## π₯ Update |
| - [21/05/2025]π₯Key-token Advantage Estimation is coming! |
|
|
| ## π Contents |
|
|
| - [Models](#Available_Models) |
| - [Setup](#Setup) |
| - [Preparation](#Preparation) |
| - [Train](#Train) |
| - [Inference](#Inference) |
| - [Experiments](#Experiments) |
| - [Citation](#citation) |
|
|
| ## π§ Available Models |
|
|
| | Model Size | DAPO+KTAE | GRPO+KTAE | |
| |------------|--------------|--------------| |
| | 1.5B | <a href="https://huggingface.co/SunW7777/DAPO_KTAE_1.5B"><img src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg" alt="HF" width="20"/> DAPO_KTAE_1.5B</a> | <a href="https://huggingface.co/SunW7777/GRPO_KTAE_1.5B"><img src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg" alt="HF" width="20"/> GRPO_KTAE_1.5B</a> | |
| | 7B | <a href="https://huggingface.co/SunW7777/DAPO_KTAE-7B"><img src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg" alt="HF" width="20"/> DAPO_KTAE-7B</a> | <a href="https://huggingface.co/SunW7777/GRPO_KTAE-7B"><img src="https://huggingface.co/front/assets/huggingface_logo-noborder.svg" alt="HF" width="20"/> GRPO_KTAE-7B</a> | |
|
|
|
|
|
|
| ## π· Setup |
|
|
| Please follow the instructions below to install the required packages. |
|
|
|
|
| 1. Clone this repository |
|
|
| ```bash |
| https://github.com/xiaolizh1/KTAE.git |
| ``` |
|
|
| 2. Install Package |
|
|
| ```bash |
| conda create -n KTAE python=3.10 -y |
| conda activate KTAE |
| cd KTAE |
| pip install -r requirements.txt |
| ``` |
|
|
| ## π Train |
|
|
| Our training is mostly performed on [Verl](https://github.com/volcengine/verl) code base and makes some changes. |
|
|
| ## π GRPO+KTAE |
|
|
| ```bash |
| bash examples/grpo_trainer/run_qwen2.5_7b.sh #train 7b model |
| bash examples/grpo_trainer/run_qwen2.5_math_1.5b.sh #train 1.5b model |
| ``` |
|
|
| ## π DAPO+KTAE |
|
|
| ```bash |
| bash recipe/dapo/run_dapo_qwen2.5_7b.sh #train 7b model |
| bash recipe/dapo/run_dapo_qwen2.5_1.5b.sh #train 1.5b model |
| ``` |
|
|
| ## π Merge Model |
|
|
| ```bash |
| cd scripts |
| bash merge_model.sh #merge checkpoint |
| ``` |
|
|
| ## β
Evaluation |
|
|
| Our evaluate code is base on [Dr.GRPO](https://github.com/sail-sg/understand-r1-zero) |
|
|
| ```bash |
| cd eval |
| bash run_eval.sh |
| ``` |
|
|
| ## π Experiments |
|
|
| We provide some results in this section. More detailed results can be found in our paper. |
| <div align=center> |
| <img width="90%" src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/exOLI0iPBFljL6x2ZOIFe.jpeg> |
| </div> |
|
|
| ### Main Results |
|
|
| + Method validation result. |
| <div align=center> |
| <img width="90%" src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/kFJiRcr47hGylp29x9pqx.png> |
| </div> |
|
|
| + Comparison with baselines on Accuracy. |
| <div align=center> |
| <img width="90%" src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/N5tKMS6w12ir0geF1oIgz.jpeg> |
| </div> |
|
|
| + Comparison with baselines on Efficiency. |
| <div align=center> |
| <img width="90%" src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/nfypL2d3jS1GuH7mM_v9y.jpeg> |
| </div> |
|
|
|
|
| ### π More Analysis |
| + Ablation analysis. |
| <div align=center> |
| <img width="80%" src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/Qphwk_uwp_uIJOTp4RAMw.png> |
| </div> |
|
|
| + Visualization example. |
| <div align=center> |
| <img width="80%" src=https://cdn-uploads.huggingface.co/production/uploads/654a05493ee6a84ff2fd3fc1/5aj7uS9uohvvLDoCyFSr0.png> |
| </div> |
|
|
|
|
| ## π Citation |
|
|
| If you find this repo useful for your research, please consider citing the paper |
|
|
| ``` |
| @misc{sun2025ktaemodelfreealgorithmkeytokens, |
| title={KTAE: A Model-Free Algorithm to Key-Tokens Advantage Estimation in Mathematical Reasoning}, |
| author={Wei Sun and Wen Yang and Pu Jian and Qianlong Du and Fuwei Cui and Shuo Ren and Jiajun Zhang}, |
| year={2025}, |
| eprint={2505.16826}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.AI}, |
| url={https://arxiv.org/abs/2505.16826}, |
| } |
| ``` |
|
|
| ## π Acknowledgement |
|
|
| We would like to thank the following repos for their great work: |
|
|
| + [Verl](https://github.com/volcengine/verl) for providing the training framework |
| + [Vllm](https://github.com/vllm-project/vllm) for the efficient inference engine with high throughput |
| + [transformers](https://github.com/huggingface/transformers) for providing the model-base and fune-tuning framework |
|
|
|
|
| ## π License |
|
|
| This project is released under the Apache 2.0 license. Parts of this project contain code and models from other sources, which are subject to their respective licenses. |
|
|
|
|
|
|