2509.14274v1

# Discovering New Theorems via LLMs with In-Context Proof Learning in Lean **Authors**: - Kazumi Kasaura (OMRON SINIC X Corporation) - &Naoto Onda (OMRON SINIC X Corporation) - &Yuta Oriike - &Masaya Taniguchi - RIKEN AIP - &Akiyoshi Sannai (Kyoto University) - &Sho Sonoda - RIKEN AIP ## Abstract Large Language Models have demonstrated significant promise in formal theorem proving. However, previous works mainly focus on solving existing problems. In this paper, we focus on the ability of LLMs to find novel theorems. We propose Conjecturing-Proving Loop pipeline for automatically generating mathematical conjectures and proving them in Lean 4 format. A feature of our approach is that we generate and prove further conjectures with context including previously generated theorems and their proofs, which enables the generation of more difficult proofs by in-context learning of proof strategies without changing parameters of LLMs. We demonstrated that our framework rediscovered theorems with verification, which were published in past mathematical papers and have not yet formalized. Moreover, at least one of these theorems could not be proved by the LLM without in-context learning, even in natural language, which means that in-context learning was effective for neural theorem proving. The source code is available at https://github.com/auto-res/ConjecturingProvingLoop. ## 1 Introduction Large Language Models (LLMs) have demonstrated significant promise in theorem proving. Since LLMs can hallucinate and it is difficult to detect it in natural language, generating formal proofs using LLM and verifying them using an interactive theorem prover (ITP), such as Lean https://lean-lang.org/, has been studied. However, previous works mainly focus on solving existing problems. In this paper, we focus on the ability of LLMs to find novel theorems. We proposed Conjecturing-Proving Loop, a pipeline for automatically generating mathematical conjectures and proving them in Lean 4 format. By separating the conjecturing and proving phases, we avoid the convergence of generated theorems and encourage proving more difficult theorems. A feature of our approach is that we generate and prove further theorems with context including proven theorems and their proofs, which enables the generation of more difficult proofs by in-context learning of proof strategies without training of LLMs. Since the ability of reasoning and generation of Lean code of closed-source LLMs, such as chatGPT, has been improved recently, we use them as both conjecturer and prover. While the disadvantage of using closed-source LLMs is that models cannot be trained freely, in our framework, the proof ability of LLMs can be improved by in-context learning from previous verified proofs. In the experiment, when mathematical notions were given as a seed, we verified whether important properties about them could be rediscovered in our framework. More specifically, we focused a few topological notions which are defined only by notions in Mathlib https://github.com/leanprover-community/mathlib4, Lean’s mathematical library, but not included in Mathlib. We generated theorems about these notions using our framework. As a result, we found that our framework rediscovered important theorems about these notions published in mathematical papers. Moreover, we verified that the in-context learning of proof strategies works: A theorem, which cannot be proved without context by LLMs even in natural language, was proved with context generated by our framework. In summary, the contributions of this paper are as follows. First, we proposed Conjecture and Proof Loop, a pipeline for automatically generating mathematical conjectures and proving them in Lean 4 format. Second, we demonstrated that our framework can rediscover a research-level theorem. Third, we verified that the proof ability of LLMs can be improved by in-context learning, when the verified proofs generated by LLMs itself are given as context. Our work also suggests the potential for automatic expansion of formal mathematics libraries using AI. Formalized mathematics is only a part of mathematics expressed in natural language, and expanding formal libraries, such as Mathlib, is crucial for verifying and automating mathematics. On the other hand, the set of propositions that should be included in the library may not always be obtainable in natural language. Our framework can generate propositions about given notions while learning them. <details> <summary>docs/AIMathematician.png Details</summary>  ### Visual Description ## Diagram Type: Flowchart ### Overview The diagram illustrates a process involving a conjecturer, a prover, and a Lean Server. The conjecturer generates conjectures, which are then sent to the prover for verification. The prover uses the Lean Server to check the validity of the conjectures. ### Components/Axes - **Conjecturer**: Represented by a robot icon with a gear in its head, labeled "Lean Server." - **Prover**: Represented by a robot icon with a gear in its head, labeled "Lean Server." - **Library**: A text box containing a mathematical theorem and proof. - **Context**: Arrows connecting the conjecturer, prover, and library, indicating the flow of information. - **Theorems & Proofs**: A text box containing a mathematical theorem and proof. ### Detailed Analysis or ### Content Details - The conjecturer generates conjectures, which are represented by arrows pointing from the conjecturer to the prover. - The prover uses the Lean Server to verify the conjectures, represented by arrows pointing from the prover to the library. - The library contains a mathematical theorem and proof, which are used to verify the conjectures. - The context arrows indicate the flow of information between the conjecturer, prover, and library. ### Key Observations - The process involves a conjecturer, a prover, and a Lean Server. - The conjecturer generates conjectures, which are sent to the prover for verification. - The prover uses the Lean Server to check the validity of the conjectures. - The library contains a mathematical theorem and proof, which are used to verify the conjectures. ### Interpretation The diagram illustrates a process of conjecturing, proving, and verifying mathematical theorems using a Lean Server. The conjecturer generates conjectures, which are sent to the prover for verification. The prover uses the Lean Server to check the validity of the conjectures. The library contains a mathematical theorem and proof, which are used to verify the conjectures. The process involves a logical flow of information between the conjecturer, prover, and library. </details> Figure 1: Overview of Conjecturing-Proving Loop. Conjecturer generates conjectures using Library as context, Prover attempts to prove them, and proven conjectures and their proofs are stored in Library as theorems. Library also provides context to Prover. Both Conjecturer and Prover processes consist of interactions between LLMs and Lean Server. ## 2 Related Work There are several works that use LLMs for mathematical reasoning, both in natural-language [1] and formal language for ITP [11, 7]. They focused mainly on solving existing problems and used Supervised Fine-Tuning (SFT) and/or Reinforcement Learning with Verified Rewards (RLVR) to improve problem-solving ability of LLMs. To overcome the limited dataset for training, the approach to generate problems to solve by AI is also proposed in some previous works [2, 6, 14]. These works are different from our approach in the following two points: First, our approach is focused on generating and proving meaningful theorems, while these works are focused on training the LLM prover. Second, while these works are based on reinforcement learning, our approach is based on in-context learning, which can be applied to closed-source LLMs. Several studies have reported that including appropriate examples of question-answer pairs in prompts improves the mathematical reasoning ability of LLMs [13, 15, 4]. While these studies used handmade examples or results searched from databases as in-context learning sources, our framework uses outputs from the LLM itself, as in In-Context Reinforcement Learning [9]. The technique to leverage feedback from ITP for formal proof generation was proposed [5, 12, 8] and we also adopted it (See § 3.3). However, what we emphasize here is learning strategies from verified proofs of other propositions. ## 3 Method ### 3.1 Pipeline Overview Figure 1 illustrates our framework. Conjecturing-Proving Loop (CPL) consists of four main parts: conjecturer (LLM agent), prover (LLM agent), Lean server, and library (Lean code data). At first, The library is initialized by the user. 1. The conjecturer generates mathematical conjectures in Lean 4 format based on the library, with accessing Lean server to check syntactical validity and novelty of generated conjectures. 1. For each generated conjectures, the prover generates proofs of the conjectures, with accessing Lean server to verify generated proofs. The library is used as the context also in this step. 1. The verified pairs of conjectures and their proofs are added to the library. We return to the first step. For the details of the second and third steps, see the following subsections. By separating the conjecturing and proving phases, we avoid the convergence of generated theorems and encourage to prove more difficult theorems. The purpose of feeding library to the conjecturer as the context is to prevent the generation of duplicate conjectures and to generate conjectures by analogy from already proven theorems. The purpose of feeding library to the prover as the context is to make already proven theorems available during proof and to learn proof strategies by in-context learning. ### 3.2 Conjecture Loop To generate varied conjectures, for each conjecturing step, we use the following process. 1. The conjecture list is initialized as empty. 1. The conjecturer LLM generates conjectures following the library and the conjecture list. 1. For each generated conjectures, Lean server checks whether the conjecture is syntactically valid and novel. Verified conjectures are added to the conjecture list. 1. We return to the second step until the predefined number of iterations is done. By repeatedly calling the conjecturer and generating many conjectures at one step, we induce the generation of conjectures that deviate from the library, thereby preventing conjectures from converging. The check for novelty of conjectures is done by using exact? command in Lean, which checks whether the conjecture can be proved by existing theorems in the context. Note that this command is done with context importing the whole Mathlib4 (Lean4 standard library) and including the library and the conjecturer list. Thus, the checked novelty means that the conjecture is not already in Mathlib4, the already generated library, and the conjecturer list. The system prompt given to conjecture LLM is as follows: Your are a writer of mathlib4 library. Please generate conjectures of new theorems in Lean 4 format, which do not need to be definitely true, following a given library. Do not generate statements that are already on the list. Do not include proofs, annotations, or imports. The new theorems begin with ’theorem’, not any annotions. They should end with ’:= sorry’. Additionally, please use standard mathematical symbols (e.g., $\forall$ , $\exists$ , $\surd$ ) instead of Unicode escape sequences (e.g., \u2200). In the experiment, the number of iterations is set to $16$ . ### 3.3 Prover Loop For each generated conjectures, the prover try to prove it in following process. 1. The prover LLM produced the proof code of the conjecture. If the LLM judges that the conjecture is not provable, the prover exits the loop as failure. 1. The Lean server verifies the generated proof. If the proof is verified, the prover exits the loop as success. 1. If already enough number of trial is done, the prover exits the loop as failure. Otherwise, the error message is returned to the LLM and we return to the first step. The context is given to both the prover and the Lean server. Thus, not only the prover can learn the proof strategies from the context, but also it can use the theorems in the context as lemmas. The system prompt given to prover LLM is as follows: You are a prover of mathlib4 library. Please prove the last theorem in the given content in Lean 4 format. You should write the Lean4 code which directly follows ":=" in the last theorem. It should begin with ’by’ or represent a term directly. You can use the theorems in the given content as lemmas. Do not use sorry in the proof. If you judge that the theorem is not provable, please return empty string instead of the proof. Do not include any other text. In the experiment, the maximum number of trials is set to $16$ . ## 4 Experiments We demonstrated that the research-level theorems can be rediscovered by our framework and verified that in-context learning worked effectively in our framework. The scripts for these experiments and generated libraries are stored at https://github.com/auto-res/ConjecturingProvingLoop. ### 4.1 Setting In our experiments, we focuse on the minor notions in general topology: semi-openness, alpha-openness, and preopenness. We used the following file including the definitions of these notions in Lean 4 format as the initial library. ⬇ import Mathlib import Aesop namespace Topology variable {X : Type *} [TopologicalSpace X] /– A set is semi - open if it is a subset of the closure of its interior. -/ def SemiOpen (A : Set X) : Prop := A ⊆ closure (interior A) /– A set is alpha - open if it is a subset of the interior of the closure of its interior. -/ def AlphaOpen (A : Set X) : Prop := A ⊆ interior (closure (interior A)) /– A set is preopen if it is a subset of the interior of its closure. -/ def PreOpen (A : Set X) : Prop := A ⊆ interior (closure A) The reason for choosing these notions is that they can be defined using only notions already present in Mathlib, while they themselves are not yet included in Mathlib, and while their mathematical importance has already been recognized and researched, they are not yet well-known enough for LLM to have knowledge of their properties. We focused whether the following theorem could be generated or not: The intersection of two alpha-open sets is alpha-open [10]. This theorem is important because it is the most difficult part to prove in the theorem that alpha-open sets form another topology (Proposition 2 in [10]). We generated theorems for these notions using our framework with $30$ loops. We use chatGPT-4o https://platform.openai.com/docs/models/gpt-4o as the conjecturer and chatGPT-o3 https://platform.openai.com/docs/models/o3 as the prover. ### 4.2 Baseline For comparison, we also generated theorems for them by a simple loop framework that LLM generates theorems and its proofs at once, using chatGPT-3o. At first, The library is initialized by the user. Unlike CPL, in this simple loop baseline, the conjecturer and the prover are not separated, and the single loop is as follows: 1. The LLM generates a statement and its proof in Lean 4 format based on the library, with accessing Lean server to verify it. 1. If the previous step succeeded, the generated pair of statement and proof are stored in Library. We return to the first step. In the experiment, we executed this loop $400$ times. The first step is similar to the prover loop. Its details are as follows: 1. The LLM produced a statement and its proof in Lean. 1. The Lean server verifies the generated content. If it is verified, we exit the loop as success. 1. If already enough number of trial is done, we exit the loop as failure. Otherwise, the error message is returned to the LLM and we return to the first step. The system prompt given to LLM is as follows: Your are a writer of mathlib4 library. Please generate a new theorem with a proof in Lean 4 format, following a given library. Do not return anything else than the Lean 4 code. The generated code should follow the given library. The generated code should contain only the theorem and the proof. Do not generate a theorem that are already on the library. The new theorem should begin with ’theorem’, not any annotions. You can use the theorems in the given library as lemmas in the proof. Do not use sorry in the proof. Additionally, please use standard mathematical symbols (e.g., $\forall$ , $\exists$ , $\surd$ ) instead of Unicode escape sequences (e.g., \u2200). As in the prover loop, the maximum number of trials is set to $16$ . ### 4.3 Results <details> <summary>x1.png Details</summary>  ### Visual Description ## Bar Chart: Proof Length Distribution ### Overview The bar chart displays the distribution of proof lengths in terms of the number of lines. It compares two types of proofs: CPL (Complex Proof) and Simple loop. ### Components/Axes - **X-axis**: Number of lines in proofs, ranging from 0 to 100. - **Y-axis**: Percentage (%), ranging from 0 to 6. - **Legend**: Two categories are represented by different colors: blue for CPL and orange for Simple loop. ### Detailed Analysis or ### Content Details - **CPL (Complex Proof)**: The majority of proofs have between 0 and 20 lines, with a peak at 5 lines. - **Simple loop**: The majority of proofs have between 0 and 20 lines, with a peak at 5 lines. - **Distribution**: Both categories show a similar distribution, with a slight peak at 5 lines. ### Key Observations - **Similarity**: Both CPL and Simple loop proofs have a similar distribution, with a peak at 5 lines. - **No significant outliers**: There are no significant outliers in the data. ### Interpretation The data suggests that both CPL and Simple loop proofs are relatively short, with a peak at 5 lines. This could indicate that these proof types are commonly used in mathematical or logical contexts where brevity is preferred. The similarity in distribution between the two categories suggests that they may be used in similar contexts or that the data is representative of a broader trend. </details> Figure 2: Distribution of proof length of theorems generated by our framework and the simple loop framework, where the size of bins is $10$ . We generated $269$ theorems for these notions, which includes the focused theorem. The generated proof of this theorem is shown in Appendix A. This proof differs from the original proof by Njȧstad, which indicates that LLM found this proof on its own. Other generated theorems which are important properties shown in research papers includes the followings: - The intersection of semi-open set and alpha-open set is semi-open [10]. - The intersection of preopen set and alpha-open set is preopen [3]. On the other hand, these theorems are not included in $359$ theorems generated by the simple loop framework. The reason seems that the simple loop framework tends to generate easily provable theorems. To confirm this, we measured the proof length of the theorems generated by our framework and the simple loop framework. Figure 2 shows the distribution of proof length of the theorems generated by our framework and the simple loop framework. It can be observed that our framework can generate theorems with longer proofs than the simple loop framework. ### 4.4 Effectiveness of Providing Contexts We also verified that feeding the generated library as a context to the prover improves the proof ability. #### 4.4.1 Reproving Generated Theorems First, we attempted to re-prove all generated theorems by the method in § 3.3 in two settings: one where the context includes the library generated before the theorem to be proved was generated, and one where the context includes only the definitions of the notions. As a result, in the setting with contexts, 99% of the theorems (267 theorems) were proved, while in the setting without contexts, only 90% of the theorems (242 theorems) were proved. Thus, the context improves the proof ability of LLMs. #### 4.4.2 Proof Ability for Alpha-Open Intersection Moreover, we attempted to re-prove the focused theorem for $128$ times in each setting. (The number of theorems generated until this theorem was generated is $49$ .) The procedure to evaluate Prover in Lean is as the same as the prover loop, except the system prompt has changed to the following: You are a prover of mathlib4 library. Please prove the last theorem in the given content in Lean 4 format. You should write the Lean4 code which directly follows ":=" in the last theorem. It should begin with ’by’ or represent a term directly. You can use the theorems in the given content as lemmas. Do not use sorry in the proof. If you judge that the theorem is false, please return empty string instead of the proof. Do not include any other text. Note that the condition not to return proof has changed from "not provable" to "false". For comparison, we also asked LLMs (chatGPT-4o and chatGPT-o3) to prove this theorem in the natural language (English) with context including the definitions of the notions $16$ times and checked the responses by hand. In the natural language experiment, we used the following system prompt: Please prove the following theorem. If you judge that the theorem is false, please return "False" instead of the proof. The statement to prove given to LLM is as follows: In a topological space, a set is alpha-open if it is a subset of the interior of the closure of its interior. The intersection of any two alpha-open sets is alpha-open. In both formal and natural language settings, we prompted the LLMs not to return a proof if the given theorem is false. Therefore, the output results fall into three categories: correctly proven, gap-containing proofs, or rejected as incorrect. <details> <summary>x2.png Details</summary>  ### Visual Description ## Bar Chart: Success Rates ### Overview The bar chart displays the success rates of different contexts and conditions. The chart is divided into four categories: w-ctx, wo-cctx, 40-NL, and o3-NL. Each category has a corresponding bar that represents the success rate, with the height of the bar indicating the percentage of successes. ### Components/Axes - **Title**: Success Rates - **X-axis**: Contexts (w-cctx, wo-cctx, 40-NL, o3-NL) - **Y-axis**: Success Rate (0.0 to 1.0) - **Legend**: Three colors represent different success outcomes: Proved (blue), Proof Gap (purple), Rejected (orange) ### Detailed Analysis or ### Content Details - **w-cctx**: The bar is divided into three segments, with the largest segment being orange, indicating the highest success rate. - **wo-cctx**: The bar is divided into three segments, with the largest segment being orange, indicating the highest success rate. - **40-NL**: The bar is divided into three segments, with the largest segment being orange, indicating the highest success rate. - **o3-NL**: The bar is divided into three segments, with the largest segment being orange, indicating the highest success rate. ### Key Observations - All four categories have the same success rate, with the largest segment being orange. - The success rate is consistent across all categories. ### Interpretation The data suggests that the success rate is consistent across all four categories. The largest segment of each bar is orange, indicating the highest success rate. This suggests that the context and condition do not significantly impact the success rate. The blue segment represents the Proved outcome, the purple segment represents the Proof Gap outcome, and the orange segment represents the Rejected outcome. The data does not provide any information about the specific reasons for the success or failure. </details> Figure 3: Success rates of proving the theorem by LLMs. With and without context in Lean (w-ctx, w-ctx), proof in natural language (4o-NL, o3-NL). The result is shown in Figure 3. The result shows that the theorem was proved only in the setting with the generated library. The fact that the majority of judgments in natural language incorrectly identify the theorem as false means that the discovered theorem was not included in chatGPT’s knowledge. An example of proofs with gaps generated by chatGPT-4o is shown in Appendix B. ChatGPT-3o never generated incorrect proof in both Lean and the natural language, but it always judged incorrectly that the theorem is false. The generated proof of this theorem, shown in Appendix A, does not use other generated theorems as lemmas. Thus, the generated library was used for in-context learning of proof strategies rather than as a collection of lemmas for the proof. ## 5 Conclusion and Future Work We presented Conjecturing-Proving Loop, a pipeline for automatically generating mathematical conjectures and proving them in Lean 4 format. We demonstrated that our framework can rediscover a research-level theorem. We also verified the in-context learning of proof strategies works effectively in our framework. The propositions highlighted in this study were easy to predict. Future work should focus on refining the conjecture generation process to produce deeper and more insightful mathematical statements, possibly by incorporating techniques for guiding the LLM towards unexplored areas of mathematical theory. ## Acknowledgements This work was supported by JST Moonshot R&D Program JPMJMS2236, JST BOOST JPMJBY24E2, JST CREST JPMJCR2015, JSPS KAKENHI 24K21316, 24K16077, and Advanced General Intelligence for Science Program (AGIS), the RIKEN TRIP initiative. ## References - [1] DeepSeek-AI, Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Li, Hui Qu, J. L. Cai, Jian Liang, Jianzhong Guo, Jiaqi Ni, Jiashi Li, Jiawei Wang, Jin Chen, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, Junxiao Song, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Lei Xu, Leyi Xia, Liang Zhao, Litong Wang, Liyue Zhang, Meng Li, Miaojun Wang, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Mingming Li, Ning Tian, Panpan Huang, Peiyi Wang, Peng Zhang, Qiancheng Wang, Qihao Zhu, Qinyu Chen, Qiushi Du, R. J. Chen, R. L. Jin, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, Runxin Xu, Ruoyu Zhang, Ruyi Chen, S. S. Li, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shaoqing Wu, Shengfeng Ye, Shengfeng Ye, Shirong Ma, Shiyu Wang, Shuang Zhou, Shuiping Yu, Shunfeng Zhou, Shuting Pan, T. Wang, Tao Yun, Tian Pei, Tianyu Sun, W. L. Xiao, Wangding Zeng, Wanjia Zhao, Wei An, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, X. Q. Li, Xiangyue Jin, Xianzu Wang, Xiao Bi, Xiaodong Liu, Xiaohan Wang, Xiaojin Shen, Xiaokang Chen, Xiaokang Zhang, Xiaosha Chen, Xiaotao Nie, Xiaowen Sun, Xiaoxiang Wang, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xingkai Yu, Xinnan Song, Xinxia Shan, Xinyi Zhou, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, Y. K. Li, Y. Q. Wang, Y. X. Wei, Y. X. Zhu, Yang Zhang, Yanhong Xu, Yanhong Xu, Yanping Huang, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Li, Yaohui Wang, Yi Yu, Yi Zheng, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Ying Tang, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yu Wu, Yuan Ou, Yuchen Zhu, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yukun Zha, Yunfan Xiong, Yunxian Ma, Yuting Yan, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang Zhou, Z. F. Wu, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhen Huang, Zhen Zhang, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhibin Gou, Zhicheng Ma, Zhigang Yan, Zhihong Shao, Zhipeng Xu, Zhiyu Wu, Zhongyu Zhang, Zhuoshu Li, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Ziyi Gao, and Zizheng Pan. DeepSeek-V3 technical report, 2025. - [2] Kefan Dong and Tengyu Ma. STP: Self-play LLM theorem provers with iterative conjecturing and proving, 2025. - [3] Julian Dontchev. Survey on preopen sets. arXiv preprint math/9810177, 1998. - [4] Iddo Drori, Sarah Zhang, Reece Shuttleworth, Leonard Tang, Albert Lu, Elizabeth Ke, Kevin Liu, Linda Chen, Sunny Tran, Newman Cheng, et al. A neural network solves, explains, and generates university math problems by program synthesis and few-shot learning at human level. Proceedings of the National Academy of Sciences, 119(32):e2123433119, 2022. - [5] Emily First, Markus N Rabe, Talia Ringer, and Yuriy Brun. Baldur: Whole-proof generation and repair with large language models. In Proceedings of the 31st ACM Joint European Software Engineering Conference and Symposium on the Foundations of Software Engineering, pages 1229–1241, 2023. - [6] Chengsong Huang, Wenhao Yu, Xiaoyang Wang, Hongming Zhang, Zongxia Li, Ruosen Li, Jiaxin Huang, Haitao Mi, and Dong Yu. R-zero: Self-evolving reasoning LLM from zero data. arXiv preprint arXiv:2508.05004, 2025. - [7] Yong Lin, Shange Tang, Bohan Lyu, Jiayun Wu, Hongzhou Lin, Kaiyu Yang, Jia LI, Mengzhou Xia, Danqi Chen, Sanjeev Arora, and Chi Jin. Goedel-prover: A frontier model for open-source automated theorem proving. In Second Conference on Language Modeling, 2025. - [8] Yong Lin, Shange Tang, Bohan Lyu, Ziran Yang, Jui-Hui Chung, Haoyu Zhao, Lai Jiang, Yihan Geng, Jiawei Ge, Jingruo Sun, et al. Goedel-prover-v2: Scaling formal theorem proving with scaffolded data synthesis and self-correction. arXiv preprint arXiv:2508.03613, 2025. - [9] Amir Moeini, Jiuqi Wang, Jacob Beck, Ethan Blaser, Shimon Whiteson, Rohan Chandra, and Shangtong Zhang. A survey of in-context reinforcement learning. arXiv preprint arXiv:2502.07978, 2025. - [10] Olav Njȧstad. On some classes of nearly open sets. Pacific journal of mathematics, 15(3):961–970, 1965. - [11] Z. Z. Ren, Zhihong Shao, Junxiao Song, Huajian Xin, Haocheng Wang, Wanjia Zhao, Liyue Zhang, Zhe Fu, Qihao Zhu, Dejian Yang, Z. F. Wu, Zhibin Gou, Shirong Ma, Hongxuan Tang, Yuxuan Liu, Wenjun Gao, Daya Guo, and Chong Ruan. DeepSeek-Prover-V2: Advancing formal mathematical reasoning via reinforcement learning for subgoal decomposition, 2025. - [12] Amitayush Thakur, George Tsoukalas, Yeming Wen, Jimmy Xin, and Swarat Chaudhuri. An in-context learning agent for formal theorem-proving. arXiv preprint arXiv:2310.04353, 2023. - [13] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022. - [14] Shaoxiong Zhan, Yanlin Lai, Ziyu Lu, Dahua Lin, Ziqing Yang, and Fei Tang. Mathsmith: Towards extremely hard mathematical reasoning by forging synthetic problems with a reinforced policy. arXiv preprint arXiv:2508.05592, 2025. - [15] Denny Zhou, Nathanael Schärli, Le Hou, Jason Wei, Nathan Scales, Xuezhi Wang, Dale Schuurmans, Claire Cui, Olivier Bousquet, Quoc Le, et al. Least-to-most prompting enables complex reasoning in large language models. arXiv preprint arXiv:2205.10625, 2022. ## Appendix A Generated Proof of Focused Theorem ⬇ theorem intersection_of_alpha_open_sets_is_alpha_open {A B : Set X} (hA : AlphaOpen A) (hB : AlphaOpen B) : AlphaOpen (A ∩ B) := by – We unfold the definition of ‘ AlphaOpen ‘ and start with a point – ‘ x ∈ A ∩ B ‘. intro x hx rcases hx with ⟨ hxA, hxB ⟩ – From the alpha - openness of ‘ A ‘ and ‘ B ‘ we obtain the two interior – memberships below. have hA_int : x ∈ interior (closure (interior A)) := hA hxA have hB_int : x ∈ interior (closure (interior B)) := hB hxB —————————————————————— – Consider the open neighbourhood – ‘ U := interior (closure (interior A)) – ∩ interior (closure (interior B))‘ of ‘ x ‘. —————————————————————— have hU_open : IsOpen (interior (closure (interior A)) ∩ interior (closure (interior B))) := isOpen_interior. inter isOpen_interior have hxU : x ∈ interior (closure (interior A)) ∩ interior (closure (interior B)) := ⟨ hA_int, hB_int ⟩ —————————————————————— – We prove that ‘ U ⊆ closure (interior (A ∩ B))‘. —————————————————————— have hU_subset : (interior (closure (interior A)) ∩ interior (closure (interior B)) : Set X) ⊆ closure (interior (A ∩ B)) := by intro y hyU rcases hyU with ⟨ hyA_int, hyB_int ⟩ – From ‘ hyA_int ‘ and ‘ hyB_int ‘ we obtain the two closure – memberships below. have hyA_cl : y ∈ closure (interior A) := (interior_subset : interior (closure (interior A)) ⊆ closure (interior A)) hyA_int have hyB_cl : y ∈ closure (interior B) := (interior_subset : interior (closure (interior B)) ⊆ closure (interior B)) hyB_int —————————————————————- – To show ‘ y ∈ closure (interior (A ∩ B))‘ – we use the neighbourhood characterisation of the closure. —————————————————————- apply (mem_closure_iff).2 intro V hV hyV ————————————————————– – First step: work inside – ‘ V ₁ := V ∩ interior (closure (interior B))‘. ————————————————————– have hV ₁ _open : IsOpen (V ∩ interior (closure (interior B))) := hV. inter isOpen_interior have hyV ₁ : y ∈ V ∩ interior (closure (interior B)) := ⟨ hyV, hyB_int ⟩ – As ‘ y ∈ closure (interior A)‘, this neighbourhood meets – ‘ interior A ‘. have h_nonempty ₁ : ((V ∩ interior (closure (interior B))) ∩ interior A). Nonempty := (mem_closure_iff).1 hyA_cl _ hV ₁ _open hyV ₁ rcases h_nonempty ₁ with ⟨ y ₁, ⟨⟨ hy ₁ V, hy ₁ B_int ⟩, hy ₁ IntA ⟩⟩ ————————————————————– – Now ‘ y ₁ ∈ V ∩ interior A ‘ ** and ** – ‘ y ₁ ∈ interior (closure (interior B))‘. – Hence ‘ y ₁ ∈ closure (interior B)‘. ————————————————————– have hy ₁ _clB : y ₁ ∈ closure (interior B) := (interior_subset : interior (closure (interior B)) ⊆ closure (interior B)) hy ₁ B_int ————————————————————– – Second step: work inside – ‘ V ₂ := V ∩ interior A ‘. ————————————————————– have hV ₂ _open : IsOpen (V ∩ interior A) := hV. inter isOpen_interior have hy ₁ V ₂ : y ₁ ∈ V ∩ interior A := ⟨ hy ₁ V, hy ₁ IntA ⟩ – Because ‘ y ₁ ∈ closure (interior B)‘, the set ‘ V ₂‘ meets – ‘ interior B ‘. have h_nonempty ₂ : ((V ∩ interior A) ∩ interior B). Nonempty := (mem_closure_iff).1 hy ₁ _clB _ hV ₂ _open hy ₁ V ₂ rcases h_nonempty ₂ with ⟨ z, ⟨⟨ hzV, hzIntA ⟩, hzIntB ⟩⟩ ————————————————————– – The point ‘ z ‘ lies in ‘ V ‘, ‘ interior A ‘, and ‘ interior B ‘; – hence it belongs to ‘ interior (A ∩ B)‘. ————————————————————– have hzIntAB : z ∈ interior (A ∩ B) := by – Use maximality of the interior to pass from – ‘ interior A ∩ interior B ‘ to ‘ interior (A ∩ B)‘. have h_sub : (interior A ∩ interior B : Set X) ⊆ interior (A ∩ B) := interior_maximal (by intro t ht rcases ht with ⟨ htA, htB ⟩ exact And. intro (interior_subset htA) (interior_subset htB)) (isOpen_interior. inter isOpen_interior) exact h_sub ⟨ hzIntA, hzIntB ⟩ – We have found a point in ‘ V ∩ interior (A ∩ B)‘. exact ⟨ z, ⟨ hzV, hzIntAB ⟩⟩ —————————————————————— – Maximality of the interior gives the desired inclusion, – and we conclude with the membership of ‘ x ‘ in ‘ U ‘. —————————————————————— have hU_subset_int : (interior (closure (interior A)) ∩ interior (closure (interior B)) : Set X) ⊆ interior (closure (interior (A ∩ B))) := interior_maximal hU_subset hU_open exact hU_subset_int hxU ## Appendix B Example of Generated Incorrect Proof The following is an example of ‘proof’ of the theorem that the intersection of two alpha-open sets is alpha-open, generated by chatGPT-4o. To prove the given theorem, let’s first define what it means for a set to be alpha-open in a topological space. Let $(X,\tau)$ be a topological space and let $A\subseteq X$ . The set $A$ is called alpha-open if $$ A\subseteq\text{int}(\overline{\text{int}(A)}), $$ where $\text{int}(A)$ denotes the interior of $A$ , and $\overline{\text{int}(A)}$ denotes the closure of $\text{int}(A)$ . We aim to prove that the intersection of any two alpha-open sets is alpha-open. Let $A,B\subseteq X$ be two alpha-open sets; hence, by definition, we have: 1. $A\subseteq\text{int}(\overline{\text{int}(A)})$ 2. $B\subseteq\text{int}(\overline{\text{int}(B)})$ We need to show that $A\cap B$ is also alpha-open, i.e., $$ A\cap B\subseteq\text{int}(\overline{\text{int}(A\cap B)}). $$ Firstly, observe that $\text{int}(A\cap B)\subseteq\text{int}(A)\cap\text{int}(B)$ . Both closure and interior operations are monotonic, so: $$ \overline{\text{int}(A\cap B)}\subseteq\overline{\text{int}(A)\cap\text{int}(B)}\subseteq\overline{\text{int}(A)}\cap\overline{\text{int}(B)}. $$ Taking interiors again yields: $$ \text{int}(\overline{\text{int}(A\cap B)})\subseteq\text{int}(\overline{\text{int}(A)}\cap\overline{\text{int}(B)})\subseteq\text{int}(\overline{\text{int}(A)})\cap\text{int}(\overline{\text{int}(B)}). $$ By the assumptions that $A\subseteq\text{int}(\overline{\text{int}(A)})$ and $B\subseteq\text{int}(\overline{\text{int}(B)})$ , it follows that $$ A\cap B\subseteq\text{int}(\overline{\text{int}(A)})\cap\text{int}(\overline{\text{int}(B)})\supset\text{int}(\overline{\text{int}(A\cap B)}). $$ Therefore, $$ A\cap B\subseteq\text{int}(\overline{\text{int}(A\cap B)}). $$ This shows that $A\cap B$ is indeed alpha-open. Therefore, the intersection of any two alpha-open sets is alpha-open. Hence, the theorem is true. This proof clearly has a gap because, when deriving the last relation from the second-to-last relations, the direction of a subset relation is reversed.