arxiv: 2508.07809 · v5 · submitted 2025-08-11 · 💻 cs.LG

EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning

Huanyu Liu , Jia Li , Yihong Dong , Chang Yu , Taozhi Chen , Lecheng Wang , Yongding Tao , Bin Gu

show 1 more author

Ge Li

This is my paper

Pith reviewed 2026-05-18 23:53 UTC · model grok-4.3

classification 💻 cs.LG

keywords reinforcement learningchain of thoughtcurriculum learninglarge language modelsreasoningexploration bottleneckself-evolving methods

0 comments p. Extension

The pith

EvoCoT lets LLMs solve hard reasoning problems by starting with verified self-generated CoT trajectories and gradually shortening them.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes EvoCoT to fix the exploration bottleneck that arises in reinforcement learning with verifiable rewards when models have low initial accuracy on difficult problems. It works by letting the model generate and verify its own chain-of-thought trajectories to keep early exploration focused, then shortens those trajectories step by step to open up more space as capability grows. This curriculum-style process allows stable progress on problems that start unsolved without relying on teacher models or removing hard examples. Readers would care because the method scales to multiple model families and works with existing RL fine-tuning techniques while improving reasoning purely through self-supervision.

Core claim

EvoCoT constrains the exploration space in RLVR by self-generating and verifying CoT trajectories, then gradually shortens CoT steps to expand the space in a controlled way, enabling LLMs to stably learn from initially unsolved hard problems under sparse rewards.

What carries the argument

Two-stage chain-of-thought reasoning optimization that first uses verified self-generated trajectories to limit exploration and then shortens them to expand the space gradually.

If this is right

LLMs can solve problems they initially could not solve.
Reasoning improves without any external CoT supervision.
The approach remains compatible with various RL fine-tuning methods.
Results hold when applied to Qwen, DeepSeek, and Llama model families.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Self-verification loops of this kind could reduce the need for curated datasets or teacher models in future reasoning training.
The curriculum idea of starting narrow and expanding might transfer to other sparse-reward settings such as robotics or game playing.
If verification stays reliable at scale, models could run longer autonomous improvement cycles with minimal human input.

Load-bearing premise

That self-generated CoT trajectories can be reliably verified for correctness without external supervision or introducing systematic errors.

What would settle it

An experiment in which accepted self-generated CoT trajectories contain undetected errors and the model shows no improvement or performance decline on the previously unsolved hard problems.

Figures

Figures reproduced from arXiv: 2508.07809 by Bin Gu, Chang Yu, Ge Li, Huanyu Liu, Jia Li, Lecheng Wang, Taozhi Chen, Yihong Dong, Yongding Tao.

**Figure 1.** Figure 1: The overall framework of EvoCoT. It is structured as two nested stages: Stage 1: AnswerGuided Reasoning Path Self-Generation., which generates and filters CoT trajectories from finalanswer supervision, and Stage 2: Step-Wise Curriculum Learning., which implements curriculum learning by progressively shortening CoTs to increase difficulty and exploration space. The two stages iterate jointly, enabling the… view at source ↗

**Figure 2.** Figure 2: Number of correct rollouts over training [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Case study in the EvoCoT self-generated CoTs with Qwen2.5-7B. (a) A correct reasoning path. (b) Ground truth answer error in GSM8K. (c) LLM fails to generate a consistent reasoning path given (Q, A). (d) LLM forcibly splices the final answer. 5 DISCUSSION In this section, we analyze why EvoCoT cannot self-evolve indefinitely. During Stage 1, we observe that certain problems remain persistently unsolved des… view at source ↗

**Figure 4.** Figure 4: Qwen2.5 Prompt format used for EvoCoT [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗

read the original abstract

Reinforcement learning with verifiable reward (RLVR) has become a promising paradigm for post-training large language models (LLMs) to improve their reasoning capability. However, when the rollout accuracy is low on hard problems, the reward becomes sparse, limiting learning efficiency and causing exploration bottlenecks. Existing approaches either rely on teacher models for distillation or filter out difficult problems, which limits scalability or restricts reasoning improvement through exploration. We propose EvoCoT, a self-evolving curriculum learning framework based on two-stage chain-of-thought (CoT) reasoning optimization. EvoCoT constrains the exploration space by self-generating and verifying CoT trajectories, then gradually shortens CoT steps to expand the space in a controlled way. The framework enables LLMs to stably learn from initially unsolved hard problems under sparse rewards. We apply EvoCoT to multiple LLM families, including Qwen, DeepSeek, and Llama. Experiments show that EvoCoT enables LLMs to solve previously unsolved problems, improves reasoning capability without external CoT supervision, and is compatible with various RL fine-tuning methods. We release the source code to support future research.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

EvoCoT gives a workable self-supervised curriculum for sparse-reward RL on hard reasoning problems by verifying then shortening CoT paths, though final-answer checks alone may reinforce flawed steps.

read the letter

EvoCoT tackles the exploration bottleneck in RLVR by first letting the model generate and verify its own CoT trajectories to narrow the search space, then gradually shortening those steps to reopen exploration in a controlled way. This lets training continue on problems where initial accuracy is too low for useful rewards. They test the approach on Qwen, DeepSeek, and Llama families and report it works with several RL fine-tuning methods while releasing the code. That combination of self-evolution without teachers or problem filtering is the clearest new piece, and the multi-model results give some evidence it is not overly brittle to architecture choice. The empirical framing avoids obvious circularity since there are no fitted parameters or closed-form claims being tested against themselves. The main soft spot is exactly the one the stress-test flags: verification rests on final-answer correctness, so trajectories with incorrect or shortcut reasoning can still receive positive reward and get reinforced. Without step-level checks or strong ablations on the shortening schedule, it is hard to know whether the gains reflect better reasoning or just more efficient exploitation of lucky paths. The paper is aimed at researchers doing RL post-training for math and coding reasoning. Readers who already work on exploration under sparse verifiable rewards will see the most direct value and can test the released code themselves. It is solid enough on the practical side to deserve a serious referee rather than a desk reject, even if the verification mechanism needs more scrutiny in review.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes EvoCoT, a self-evolving curriculum learning framework for reinforcement learning with verifiable reward (RLVR) applied to LLMs. It uses a two-stage CoT optimization process: first constraining exploration by self-generating and verifying CoT trajectories on hard problems, then gradually shortening the CoT steps to expand the exploration space in a controlled manner. This is intended to enable stable learning from initially unsolved problems under sparse rewards. Experiments across Qwen, DeepSeek, and Llama model families are reported to show that EvoCoT solves previously unsolved problems, improves reasoning without external CoT supervision, and is compatible with various RL fine-tuning methods. Source code is released.

Significance. If the results hold and the self-verification process proves reliable, the work could meaningfully advance scalable post-training of LLMs on complex reasoning tasks by mitigating the exploration bottleneck in RLVR without teacher distillation or problem filtering. The self-evolving curriculum and cross-method compatibility are potentially useful contributions, and releasing the source code is a clear strength for reproducibility.

major comments (2)

[Method (two-stage CoT optimization description)] The central mechanism relies on self-verification of generated CoT trajectories to constrain exploration before shortening steps. However, in standard RLVR the only signal is the sparse final-answer reward; this does not certify logical soundness of intermediate steps. Trajectories containing flawed or shortcut reasoning that happen to produce the correct answer can still receive positive reward. If such paths are systematically retained, the curriculum risks reinforcing brittle rather than robust reasoning, directly threatening the claims of solving previously unsolved problems and improving reasoning capability without external CoT supervision. Please clarify the exact verification procedure (e.g., any step-level checks or filtering) and provide ablations demonstrating that incorrect intermediate paths are not reinforced.
[Experiments section] The abstract states that experiments on Qwen, DeepSeek, and Llama families demonstrate the method works and enables solving previously unsolved problems. Without reported tables, specific metrics for identifying 'previously unsolved' problems, ablation studies on the curriculum components, or controls for post-hoc problem selection, it is impossible to assess whether the reported gains are robust or affected by implementation choices. Include quantitative results on rollout accuracy before/after EvoCoT and comparisons to baselines that also use only final-answer reward.

minor comments (2)

[Abstract] The abstract could more explicitly name the RL fine-tuning methods (e.g., PPO, GRPO) with which EvoCoT is shown to be compatible.
[Method] Notation for the two-stage process and the shortening schedule should be defined more formally with equations or pseudocode to improve clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments on our manuscript. We address each of the major comments in detail below, providing clarifications and indicating revisions to the manuscript where appropriate.

read point-by-point responses

Referee: [Method (two-stage CoT optimization description)] The central mechanism relies on self-verification of generated CoT trajectories to constrain exploration before shortening steps. However, in standard RLVR the only signal is the sparse final-answer reward; this does not certify logical soundness of intermediate steps. Trajectories containing flawed or shortcut reasoning that happen to produce the correct answer can still receive positive reward. If such paths are systematically retained, the curriculum risks reinforcing brittle rather than robust reasoning, directly threatening the claims of solving previously unsolved problems and improving reasoning capability without external CoT supervision. Please clarify the exact verification procedure (e.g., any step-level checks or filtering) and provide ablations demonstrating that incorrect intermediate paths are not reinforced.

Authors: We appreciate the referee raising this critical issue regarding the verification process. In EvoCoT, the verification of self-generated CoT trajectories is based solely on the final-answer correctness using the verifiable reward signal available in RLVR settings. There are no explicit step-level logical checks or additional filtering beyond ensuring the final answer matches the ground truth. We acknowledge that this approach may retain some trajectories with flawed intermediate reasoning if they lead to the correct final answer. To mitigate concerns about reinforcing brittle reasoning, we have revised the manuscript to include a more detailed description of the verification procedure in Section 3. Additionally, we have added an ablation study comparing the full EvoCoT with a variant that uses only final rewards without the curriculum, showing improved stability and performance. While we cannot provide exhaustive manual inspection of all intermediate steps due to the scale, the empirical results across multiple models support that the method enhances overall reasoning capabilities. revision: partial
Referee: [Experiments section] The abstract states that experiments on Qwen, DeepSeek, and Llama families demonstrate the method works and enables solving previously unsolved problems. Without reported tables, specific metrics for identifying 'previously unsolved' problems, ablation studies on the curriculum components, or controls for post-hoc problem selection, it is impossible to assess whether the reported gains are robust or affected by implementation choices. Include quantitative results on rollout accuracy before/after EvoCoT and comparisons to baselines that also use only final-answer reward.

Authors: We agree that providing more granular experimental details is essential for assessing the robustness of our results. In the revised manuscript, we have expanded the Experiments section with tables reporting rollout accuracy before and after EvoCoT for each model family (Qwen, DeepSeek, Llama). We define 'previously unsolved problems' as those where the base model's initial rollout accuracy is 0% on the test set, and report the post-EvoCoT success rates. We have included ablation studies on the curriculum components, such as the impact of the two-stage process. Furthermore, we added comparisons to standard RL fine-tuning baselines that rely only on final-answer rewards, demonstrating the advantages of EvoCoT. To address potential post-hoc selection biases, we have clarified the problem selection process and used fixed evaluation sets. revision: yes

Circularity Check

0 steps flagged

EvoCoT is an empirical curriculum procedure with no self-referential derivation or fitted predictions

full rationale

The paper describes EvoCoT as a two-stage self-evolving curriculum for RLVR: self-generate CoT trajectories, verify them (via the existing sparse final-answer reward), then gradually shorten steps to expand exploration. No equations, parameters, or uniqueness theorems are presented that reduce to the inputs by construction. The method is a training recipe whose claims rest on experimental outcomes across Qwen, DeepSeek, and Llama models rather than a closed logical loop. No load-bearing self-citations, ansatzes smuggled via prior work, or renaming of known results appear in the derivation chain. The central result is therefore self-contained as an empirical contribution.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified assumption that self-generated CoT trajectories can be accurately verified without external models and that shortening those trajectories provides a controlled expansion of exploration space.

axioms (1)

domain assumption Self-generated CoT trajectories can be reliably verified for correctness by the model itself or simple checks.
Required for the constraint stage to function without teacher supervision.

pith-pipeline@v0.9.0 · 5747 in / 1089 out tokens · 40521 ms · 2026-05-18T23:53:06.102614+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

24 extracted references · 24 canonical work pages · 13 internal anchors

[1]

Sanghwan Bae, Jiwoo Hong, Min Young Lee, Hanbyul Kim, JeongYeon Nam, and Donghyun Kwak

URLhttps://arxiv.org/abs/2506.18110. Sanghwan Bae, Jiwoo Hong, Min Young Lee, Hanbyul Kim, JeongYeon Nam, and Donghyun Kwak. Online difficulty filtering for reasoning oriented reinforcement learning.CoRR, abs/2504.03380,

work page arXiv
[2]

Step-wise adaptive integration of supervised fine-tuning and rein- forcement learning for task-specific llms.arXiv preprint arXiv:2505.13026,

Jack Chen, Fazhong Liu, Naruto Liu, Yuhan Luo, Erqu Qin, Harry Zheng, Tian Dong, Haojin Zhu, Yan Meng, and Xiao Wang. Step-wise adaptive integration of supervised fine-tuning and rein- forcement learning for task-specific llms.CoRR, abs/2505.13026, 2025a. Xiaoyin Chen, Jiarui Lu, Minsu Kim, Dinghuai Zhang, Jian Tang, Alexandre Pich´e, Nicolas Gontier, Yos...

work page arXiv
[3]

Training Verifiers to Solve Math Word Problems

Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.CoRR, abs/2110.14168,

work page internal anchor Pith review Pith/arXiv arXiv
[4]

Process Reinforcement through Implicit Rewards

Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, Jiarui Yuan, Huayu Chen, Kaiyan Zhang, Xingtai Lv, Shuo Wang, Yuan Yao, Xu Han, Hao Peng, Yu Cheng, Zhiyuan Liu, Maosong Sun, Bowen Zhou, and Ning Ding. Process reinforcement through implicit rewards.CoRR, abs/2502.01456,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai D...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[6]

Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aur ´elien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Rozi`ere, Beth...

work page internal anchor Pith review Pith/arXiv arXiv
[7]

Anna Goldie, Azalia Mirhoseini, Hao Zhou, Irene Cai, and Christopher D

URLhttps://arxiv.org/abs/2506.19767. Anna Goldie, Azalia Mirhoseini, Hao Zhou, Irene Cai, and Christopher D. Manning. Synthetic data generation & multi-step RL for reasoning & tool use.CoRR, abs/2504.04736,

work page arXiv
[8]

Open-Reasoner-Zero: An Open Source Approach to Scaling Up Reinforcement Learning on the Base Model

Jingcheng Hu, Yinmin Zhang, Qi Han, Daxin Jiang, Xiangyu Zhang, and Heung-Yeung Shum. Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model.CoRR, abs/2503.24290,

work page internal anchor Pith review Pith/arXiv arXiv
[9]

SATURN: sat-based rein- forcement learning to unleash language model reasoning.CoRR, abs/2505.16368, 2025a

Huanyu Liu, Jia Li, Hao Zhu, Kechi Zhang, Yihong Dong, and Ge Li. SATURN: sat-based rein- forcement learning to unleash language model reasoning.CoRR, abs/2505.16368, 2025a. Mingjie Liu, Shizhe Diao, Ximing Lu, Jian Hu, Xin Dong, Yejin Choi, Jan Kautz, and Yi Dong. Prorl: Prolonged reinforcement learning expands reasoning boundaries in large language mode...

work page arXiv 2025
[10]

Sanmit Narvekar, Jivko Sinapov, Matteo Leonetti, and Peter Stone

URLhttps://arxiv.org/ abs/2506.07527. Sanmit Narvekar, Jivko Sinapov, Matteo Leonetti, and Peter Stone. Source task creation for curricu- lum learning. InAAMAS, pp. 566–574. ACM,

work page arXiv
[11]

Shubham Parashar, Shurui Gui, Xiner Li, Hongyi Ling, Sushil Vemuri, Blake Olson, Eric Li, Yu Zhang, James Caverlee, Dileep Kalathil, and Shuiwang Ji

URLhttps: //arxiv.org/abs/2506.13923. Shubham Parashar, Shurui Gui, Xiner Li, Hongyi Ling, Sushil Vemuri, Blake Olson, Eric Li, Yu Zhang, James Caverlee, Dileep Kalathil, and Shuiwang Ji. Curriculum reinforcement learning from easy to hard tasks improves llm reasoning,

work page arXiv
[12]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y

URLhttps://arxiv.org/abs/ 2506.06632. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.CoRR, abs/2402.03300,

work page arXiv
[13]

DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

doi: 10.48550/ARXIV .2402.03300. URL https://doi.org/10.48550/arXiv.2402.03300. Taiwei Shi, Yiyang Wu, Linxin Song, Tianyi Zhou, and Jieyu Zhao. Efficient reinforcement finetun- ing via adaptive curriculum learning.CoRR, abs/2504.05520,

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv
[14]

Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, Chuning Tang, Congcong Wang, Dehao Zhang, Enming Yuan, Enzhe Lu, Fengxiang Tang, Flood Sung, Guangda Wei, Guokun Lai, Haiqing Guo, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Hao Zhang, Haotian Yao, Haotian Zhao, Haoyu Lu, Haoze Li, Haoz...

work page internal anchor Pith review Pith/arXiv arXiv
[15]

Thought-augmented policy optimization: Bridging external guidance and internal capabilities.CoRR, abs/2505.15692,

Jinyang Wu, Chonghua Liao, Mingkuan Feng, Shuai Zhang, Zhengqi Wen, Pengpeng Shao, Huazhe Xu, and Jianhua Tao. Thought-augmented policy optimization: Bridging external guidance and internal capabilities.CoRR, abs/2505.15692,

work page arXiv
[16]

Logic-RL: Unleashing LLM Reasoning with Rule-Based Reinforcement Learning

Tian Xie, Zitian Gao, Qingnan Ren, Haoming Luo, Yuqian Hong, Bryan Dai, Joey Zhou, Kai Qiu, Zhirong Wu, and Chong Luo. Logic-rl: Unleashing LLM reasoning with rule-based reinforcement learning.CoRR, abs/2502.14768,

work page internal anchor Pith review Pith/arXiv arXiv
[17]

Learning to Reason under Off-Policy Guidance

Jianhao Yan, Yafu Li, Zican Hu, Zhi Wang, Ganqu Cui, Xiaoye Qu, Yu Cheng, and Yue Zhang. Learning to reason under off-policy guidance.CoRR, abs/2504.14945,

work page internal anchor Pith review Pith/arXiv arXiv
[18]

Qwen2.5 Technical Report

12 Preprint, July 2025 An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Li...

work page internal anchor Pith review Pith/arXiv arXiv 2025
[19]

DAPO: An Open-Source LLM Reinforcement Learning System at Scale

Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Hongli Yu, Weinan Dai, Yuxuan Song, Xiangpeng Wei, Hao Zhou, Jingjing Liu, W...

work page internal anchor Pith review Pith/arXiv arXiv
[20]

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Yang Yue, Zhiqi Chen, Rui Lu, Andrew Zhao, Zhaokai Wang, Yang Yue, Shiji Song, and Gao Huang. Does reinforcement learning really incentivize reasoning capacity in llms beyond the base model?CoRR, abs/2504.13837,

work page internal anchor Pith review Pith/arXiv arXiv
[21]

SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild

Weihao Zeng, Yuzhen Huang, Qian Liu, Wei Liu, Keqing He, Zejun Ma, and Junxian He. Simplerl- zoo: Investigating and taming zero reinforcement learning for open base models in the wild. CoRR, abs/2503.18892,

work page internal anchor Pith review Pith/arXiv arXiv
[22]

Kakade, Cengiz Pehlevan, Samy Jelassi, and Eran Malach

Rosie Zhao, Alexandru Meterez, Sham M. Kakade, Cengiz Pehlevan, Samy Jelassi, and Eran Malach. Echo chamber: RL post-training amplifies behaviors learned in pretraining.CoRR, abs/2504.07912,

work page arXiv
[23]

13 Preprint, July 2025 APPENDIX TABLE OFCONTENTS • Appendix A:EvoCoTPseudocode Algorithm • Appendix B:EvoCoTPrompts Templates • Appendix C:EvoCoTHyperparameters • Appendix D: LLMs Usage A THEPSEUDOCODE OFEVOCOTALGORITHM Algorithm 1 presents the complete algorithmic workflow ofEvoCoT. Algorithm 1 EvoCoT: Self-Evolving Curriculum Learning 1defEvoCoT(LLM, D,...

work page 2025
[24]

For eval- uation, we use the Qwen2.5-7B-Math framework 5 to evaluate LLMs’ performance across various benchmarks. 5https://github.com/QwenLM/Qwen2.5-Math 14 Preprint, July 2025 Stage 1: Answer-Guided CoT Generation Given a question and its final answer, generate a clear, detailed, and logically sound step-by-step reasoning process that leads to the answer...

work page 2025