新增: DeepSeek 全部 33 篇官方论文(原文 PDF 走 Git LFS + 中文精读笔记与索引)
This commit is contained in:
ylhan
@@ -0,0 +1,336 @@
|
||||
TITLE: DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning
|
||||
ARXIV: 2511.22570
|
||||
AUTHORS: Zhihong Shao , Yuxiang Luo , Chengda Lu , Z.Z. Ren , Jiewen Hu , Tian Ye , Zhibin Gou , Shirong Ma , Xiaokang Zhang
|
||||
DATE: [Submitted on 27 Nov 2025]
|
||||
|
||||
=== SECTION HEADINGS ===
|
||||
## Report GitHub Issue
|
||||
## DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning
|
||||
## Abstract
|
||||
## 1 Introduction
|
||||
## 2 Method
|
||||
## 2.1 Proof Verification
|
||||
## 2.1.1 Training a Verifier to Identify Issues and Score Proofs
|
||||
## Curating Cold Start RL Data
|
||||
## RL Objective.
|
||||
## 2.1.2 Introducing Meta-Verification to Review Proof Analyses
|
||||
## Meta-Verifier Training Process
|
||||
## 2.2 Proof Generation
|
||||
## 2.2.1 Training a Generator for Theorem Proving
|
||||
## 2.2.2 Enhancing Reasoning via Self-Verification
|
||||
## 2.3 Synergy Between Proof Verification and Generation
|
||||
## 3 Experiments
|
||||
## 3.1 Training Settings
|
||||
## 3.2 Evaluation Benchmarks
|
||||
## 3.3 Evaluation Results
|
||||
## 3.3.1 One-Shot Generation
|
||||
## 3.3.2 Sequential Refinement with Self-Verification
|
||||
## 3.3.3 High-Compute Search
|
||||
## 4 Related Work
|
||||
## 5 Conclusion
|
||||
|
||||
=== BODY (head) ===
|
||||
DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning
|
||||
## Report GitHub Issue
|
||||
×
|
||||
Title:
|
||||
Content selection saved. Describe the issue below:
|
||||
Description:
|
||||
Submit without GitHub
|
||||
Submit in GitHub
|
||||
arXiv is now an independent nonprofit!
|
||||
Learn more
|
||||
×
|
||||
Back to arXiv
|
||||
Why HTML?
|
||||
Report Issue
|
||||
Back to Abstract
|
||||
Download PDF
|
||||
Abstract
|
||||
1 Introduction
|
||||
2 Method
|
||||
2.1 Proof Verification
|
||||
2.1.1 Training a Verifier to Identify Issues and Score Proofs
|
||||
Curating Cold Start RL Data
|
||||
RL Objective.
|
||||
2.1.2 Introducing Meta-Verification to Review Proof Analyses
|
||||
Meta-Verifier Training Process
|
||||
2.2 Proof Generation
|
||||
2.2.1 Training a Generator for Theorem Proving
|
||||
2.2.2 Enhancing Reasoning via Self-Verification
|
||||
2.3 Synergy Between Proof Verification and Generation
|
||||
3 Experiments
|
||||
3.1 Training Settings
|
||||
3.2 Evaluation Benchmarks
|
||||
3.3 Evaluation Results
|
||||
3.3.1 One-Shot Generation
|
||||
3.3.2 Sequential Refinement with Self-Verification
|
||||
3.3.3 High-Compute Search
|
||||
4 Related Work
|
||||
5 Conclusion
|
||||
References
|
||||
A Prompt Templates
|
||||
A.1 Proof Generation Prompt
|
||||
A.2 Proof Verification Prompt
|
||||
A.3 Meta-Verification Prompt
|
||||
A.4 Proof Refinement Prompt
|
||||
License: arXiv.org perpetual non-exclusive license
|
||||
arXiv:2511.22570v1 [cs.AI] 27 Nov 2025
|
||||
\reportnumber
|
||||
001
|
||||
## DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning
|
||||
Zhihong Shao*, Yuxiang Luo*, Chengda Lu* † , Z.Z. Ren*
|
||||
Jiewen Hu, Tian Ye, Zhibin Gou, Shirong Ma, Xiaokang Zhang
|
||||
DeepSeek-AI
|
||||
zhihongshao@deepseek.com
|
||||
https://github.com/deepseek-ai/DeepSeek-Math-V2
|
||||
Corresponding author: Core contributors † Work done during internship at DeepSeek-AI.
|
||||
## Abstract
|
||||
Large language models have made significant progress in mathematical reasoning, which serves as an important testbed for AI and could impact scientific research if further advanced.
|
||||
By scaling reasoning with reinforcement learning that rewards correct final answers, LLMs have improved from poor performance to saturating quantitative reasoning competitions like AIME and HMMT in one year.
|
||||
However, this approach faces fundamental limitations.
|
||||
Pursuing higher final answer accuracy doesn’t address a key issue: correct answers don’t guarantee correct reasoning.
|
||||
Moreover, many mathematical tasks like theorem proving require rigorous step-by-step derivation rather than numerical answers, making final answer rewards inapplicable.
|
||||
To push the limits of deep reasoning, we believe it is necessary to verify the comprehensiveness and rigor of mathematical reasoning.
|
||||
Self-verification is particularly important for scaling test-time compute, especially for open problems without known solutions.
|
||||
Towards self-verifiable mathematical reasoning, we investigate how to train an accurate and faithful LLM-based verifier for theorem proving.
|
||||
We then train a proof generator using the verifier as the reward model, and incentivize the generator to identify and resolve as many issues as possible in their own proofs before finalizing them.
|
||||
To maintain the generation-verification gap as the generator becomes stronger, we propose to scale verification compute to automatically label new hard-to-verify proofs, creating training data to further improve the verifier.
|
||||
Our resulting model, DeepSeekMath-V2, demonstrates strong theorem-proving capabilities, achieving gold-level scores on IMO 2025 and CMO 2024 and a near-perfect 118/120 on Putnam 2024 with scaled test-time compute.
|
||||
While much work remains, these results suggest that self-verifiable mathematical reasoning is a feasible research direction that may help develop more capable mathematical AI systems.
|
||||
## 1 Introduction
|
||||
The conventional approach to reinforcement learning (RL) for mathematical reasoning involves rewarding large language models (LLMs) based on whether their predicted final answers to quantitative reasoning problems match ground-truth answers ( Guo et al., 2025 ) .
|
||||
This methodology suffices to allow frontier LLMs to saturate mathematical competitions that primarily evaluate final answers, such as AIME and HMMT.
|
||||
However, this reward mechanism has two fundamental limitations.
|
||||
First, it serves as an unreliable proxy for reasoning correctness – a model can arrive at the correct answer through flawed logic or fortunate errors.
|
||||
Second, it is inapplicable to theorem proving tasks, where problems may not require producing numerical final answers and rigorous derivation is the primary objective.
|
||||
Consequently, LLMs trained on quantitative reasoning problems with such final answer reward still frequently produce mathematically invalid or logically inconsistent natural-language proofs.
|
||||
Moreover, this training approach does not naturally develop the models’ ability to verify proof validity – they exhibit high false-positive rates, often claiming incorrect proofs are valid even when they contain obvious logical flaws.
|
||||
The lack of a generation-verification gap in natural-language theorem proving hinders further improvement.
|
||||
To address this, we propose developing proof verification capabilities in LLMs.
|
||||
Our approach is motivated by several key observations:
|
||||
•
|
||||
Humans can identify issues in proofs even without reference solutions – a crucial ability when tackling open problems.
|
||||
•
|
||||
A proof is more likely to be valid when no issues can be identified despite scaled verification efforts.
|
||||
•
|
||||
The efforts required to identify valid issues can serve as a proxy for proof quality, which can be exploited to optimize proof generation.
|
||||
We believe that LLMs can be trained to identify proof issues without reference solutions.
|
||||
Such a verifier would enable an iterative improvement cycle: (1) using verification feedback to optimize proof generation, (2) scaling verification compute to auto-label hard-to-verify new proofs, thereby creating the training data to improve the verifier itself, and (3) using this enhanced verifier to further optimize proof generation.
|
||||
Moreover, a reliable proof verifier enables us to teach proof generators to evaluate proofs as the verifier does.
|
||||
This allows a proof generator to iteratively refine its proofs until it can no longer identify or resolve any issues.
|
||||
In essence, we make the model explicitly aware of its reward function and enable it to maximize this reward through deliberate reasoning rather than blind trial-and-error.
|
||||
Built on DeepSeek-V3.2-Exp-Base ( DeepSeek-AI, 2025 ) , we developed DeepSeekMath-V2 , a large language model optimized for natural-language theorem proving that demonstrates self-verifiable mathematical reasoning.
|
||||
Our model can assess and iteratively improve its own proofs, achieving gold-level performance in premier high-school mathematics competitions including IMO 2025 and CMO 2024.
|
||||
On the Putnam 2024 undergraduate competition, it scored 118/120, exceeding the highest score of 90 1 1
|
||||
1
|
||||
https://kskedlaya.org/putnam-archive/putnam2024stats.html obtained by human participants.
|
||||
## 2 Method
|
||||
## 2.1 Proof Verification
|
||||
## 2.1.1 Training a Verifier to Identify Issues and Score Proofs
|
||||
We developed high-level rubrics ℐ v \mathcal{I}_{v} for proof evaluation (see Appendix A.2 ) with the goal of training a verifier to evaluate proofs according to these rubrics, mirroring mathematical experts’ assessment process.
|
||||
Specifically, given a problem X X and a proof Y Y , the verifier π φ ( ⋅ | X , Y , ℐ v ) \pi_{\varphi}(\cdot|{}X,Y,\mathcal{I}_{v}) is designed to produce a proof analysis that first summarizes identified issues (if any) and then assigns a score based on three levels:
|
||||
1 for complete and rigorous proofs with all logical steps clearly justified;
|
||||
0.5 for proofs with sound overall logic but minor errors or omitted details;
|
||||
and 0 for fundamentally flawed proofs containing fatal logical errors or critical gaps.
|
||||
## Curating Cold Start RL Data
|
||||
We constructed our initial training data through the following process:
|
||||
1.
|
||||
We crawled problems from Art of Problem Solving (AoPS) contests 2 2
|
||||
2
|
||||
https://artofproblemsolving.com/community/c13_contest_collections , prioritizing math olympiads, team selection tests, and post-2010 problems explicitly requiring proofs, totaling 17,503 problems.
|
||||
This problem set is denoted as 𝒟 p \mathcal{D}_{p} .
|
||||
2.
|
||||
We generated candidate proofs using a variant of DeepSeek-V3.2-Exp-Thinking. As this model was not optimized for theorem proving and tended to produce concise but error-prone outputs, we prompted it to iteratively refine its proofs over multiple rounds to improve comprehensiveness and rigor.
|
||||
3.
|
||||
We randomly sampled proofs across diverse problem types (e.g., algebra and number theory) and had mathematical experts score each proof according to the evaluation rubrics described above.
|
||||
This process yielded an initial RL dataset 𝒟 v = { ( X i , Y i , s i ) } \mathcal{D}_{v}=\{(X_{i},Y_{i},s_{i})\} , where each item consists of a problem X i X_{i} , a proof Y i Y_{i} , and an overall proof score s i ∈ { 0 , 0.5 , 1 } s_{i}\in\{0,0.5,1\} .
|
||||
## RL Objective.
|
||||
Building on a version of DeepSeek-V3.2-Exp-SFT which was supervised fine-tuned on reasoning data related to mathematics and code, we trained the model with reinforcement learning to produce proof analyses using two reward components:
|
||||
•
|
||||
Format reward R format R_{\text{format}} : An indicator function that enforces the model to generate both a summary of identified issues and a proof score, by checking whether the final response contains the key phrase “Here is my evaluation of the solution:” as well as a score within \boxed{} following “Based on my evaluation, the final overall score should be:”.
|
||||
•
|
||||
Score reward R score R_{\text{score}} : Rewards based on proximity between predicted score s i ′ s_{i}^{\prime} and annotated score s i s_{i} :
|
||||
R score ( s i ′ , s i ) = 1 − | s i ′ − s i | R_{\text{score}}(s_{i}^{\prime},s_{i})=1-|s_{i}^{\prime}-s_{i}|
|
||||
(1)
|
||||
The RL objective for training the verifier is:
|
||||
max π φ 𝔼 ( X i , Y i , s i ) ∼ 𝒟 v , ( V i ′ , s i ′ ) ∼ π φ ( ⋅ | X i , Y i ) [ R format ( V i ′ ) ⋅ R score ( s i ′ , s i ) ] \max_{\pi_{\varphi}}\mathbb{E}_{(X_{i},Y_{i},s_{i})\sim\mathcal{D}_{v},(V_{i}^{\prime},s_{i}^{\prime})\sim\pi_{\varphi}(\cdot|X_{i},Y_{i})}\left[R_{\text{format}}(V_{i}^{\prime})\cdot R_{\text{score}}(s_{i}^{\prime},s_{i})\right]
|
||||
(2)
|
||||
where V i ′ V_{i}^{\prime} denotes the verifier’s final response and s i ′ s_{i}^{\prime} is the proof score extracted from it.
|
||||
## 2.1.2 Introducing Meta-Verification to Review Proof Analyses
|
||||
The approach described in Section 2.1.1 trains proof verification through RL to align predicted proof scores with expert annotations, but provides no direct supervision on the identified issues themselves.
|
||||
This creates a critical vulnerability: when evaluating flawed proofs (where s i < 1 s_{i}<1 ) during training, the verifier can receive full reward by predicting the correct scores while hallucinating non-existent issues, undermining its trustworthiness.
|
||||
To address this problem, we introduce meta-verification : a secondary evaluation process that assesses whether issues identified by the verifier indeed exist and whether these issues logically justify the predicted proof score according to the evaluation rubrics ℐ v \mathcal{I}_{v} .
|
||||
The complete meta-verification rubrics ℐ m v \mathcal{I}_{mv} are detailed in Appendix A.3 .
|
||||
We trained a dedicated meta-verifier using RL to perform this evaluation.
|
||||
By incorporating the meta-verifier’s feedback into verifier training, we can improve the faithfulness of the verifier’s issue identification.
|
||||
## Meta-Verifier Training Process
|
||||
1.
|
||||
We obtained an initial verifier π φ \pi_{\varphi} following Section 2.1.1 .
|
||||
2.
|
||||
Mathematical experts scored the quality of verifier responses according to ℐ m v \mathcal{I}_{mv} , creating dataset 𝒟 m v = { ( X i , Y i , V i , m s i ) } \mathcal{D}_{mv}=\{(X_{i},Y_{i},V_{i},ms_{i})\} , where V i V_{i} is the analysis of proof Y i Y_{i} and m s i ∈ { 0 , 0.5 , 1 } ms_{i}\in\{0,0.5,1\} is the expert-annotated quality score.
|
||||
3.
|
||||
We trained a meta-verifier π η ( ⋅ | X , Y , V , ℐ m v ) \pi_{\eta}(\cdot|X,Y,V,\mathcal{I}_{mv}) to analyze the verifier’s proof analysis V V .
|
||||
The meta-verifier produces a summary of issues found in the analysis itself, followed by a quality score measuring how accurate and justified the verifier’s analysis is.
|
||||
The RL objective follows the same structure as the verifier training, with format and score rewards.
|
||||
Using the trained meta-verifier π η \pi_{\eta} , we enhanced the verifier training by integrating meta-verification feedback into the reward function:
|
||||
R V = R format ⋅ R score ⋅ R meta R_{V}=R_{\text{format}}\cdot R_{\text{score}}\cdot R_{\text{meta}}
|
||||
(3)
|
||||
where R meta R_{\text{meta}} is the quality score from the meta-verifier.
|
||||
We trained the enhanced verifier on both the verification dataset 𝒟 v \mathcal{D}_{v} and the meta-verification dataset 𝒟 m v \mathcal{D}_{mv} , using the same reward mechanism on 𝒟 m v \mathcal{D}_{mv} as used for training the meta-verifier.
|
||||
The resulting model can perform both proof verification and meta-verification tasks.
|
||||
On a validation split of 𝒟 v \mathcal{D}_{v} , the average quality score of the verifier’s proof analyses – as evaluated by the meta-verifier – improved from 0.85 to 0.96, while maintaining the same accuracy in proof score prediction.
|
||||
## 2.2 Proof Generation
|
||||
## 2.2.1 Training a Generator for Theorem Proving
|
||||
With verifier π φ \pi_{\varphi} serving as a generative reward model, we train a proof generator π θ ( ⋅ | X ) \pi_{\theta}(\cdot|{}X) with the RL objective:
|
||||
max π θ 𝔼 X i ∼ 𝒟 p , Y i ∼ π θ ( ⋅ | X i ) [ R Y ] \max_{\pi_{\theta}}\mathbb{E}_{X_{i}\sim\mathcal{D}_{p},Y_{i}\sim\pi_{\theta}(\cdot|X_{i})}[R_{Y}]
|
||||
(4)
|
||||
where R Y R_{Y} is the proof score produced by π φ ( ⋅ | X i , Y i , ℐ v ) \pi_{\varphi}(\cdot|{}X_{i},Y_{i},\mathcal{I}_{v}) .
|
||||
## 2.2.2 Enhancing Reasoning via Self-Verification
|
||||
When a proof generator fails to produce a completely correct proof in one shot – common for challenging problems from competitions like IMO and CMO – iterative verification and refinement can improve results.
|
||||
This involves analyzing the proof with an external verifier and prompting the generator to address identified issues.
|
||||
However, we observed a critical limitation: when prompted to both generate and analyze its own proof in one shot, the generator tends to claim correctness even when the external verifier easily identify flaws.
|
||||
In other words, while the generator can refine proofs based on external feedback, it fails to evaluate its own work with the same rigor as the dedicated verifier.
|
||||
This observation motivated us to endow the proof generator with genuine verification capabilities.
|
||||
During training, we prompt the generator π θ \pi_{\theta} to produce a proof Y Y followed by a self-analysis Z Z that follows the same format and rubrics ℐ v \mathcal{I}_{v} as the verifier (see Appendix A.1 ).
|
||||
We denote the proof score predicted in the self-analysis as s ′ s^{\prime} .
|
||||
To ensure faithful self-evaluation, we use the verifier π φ \pi_{\varphi} to assess both components: the proof Y Y receives score R Y = s R_{Y}=s , and the self-analysis Z Z receives a meta-verification score R meta ( Z ) = m s R_{\text{meta}}(Z)=ms .
|
||||
The reward function combines these assessments:
|
||||
R \displaystyle R
|
||||
= R format ( Y , Z ) ⋅ ( α ⋅ R Y + β ⋅ R Z ) \displaystyle=R_{\text{format}}(Y,Z)\cdot(\alpha\cdot R_{Y}+\beta\cdot R_{Z})
|
||||
(5)
|
||||
R Z \displaystyle R_{Z}
|
||||
= R score ( s ′ , s ) ⋅ R meta ( Z ) \displaystyle=R_{\text{score}}(s^{\prime},s)\cdot R_{\text{meta}}(Z)
|
||||
(6)
|
||||
where R format ( Y , Z ) R_{\text{format}}(Y,Z) verifies that both the proof and self-analysis follow the specified format, R score ( s ′ , s ) R_{\text{score}}(s^{\prime},s) rewards accurate self-assessment.
|
||||
We set α = 0.76 \alpha=0.76 and β = 0.24 \beta=0.24 .
|
||||
This reward structure creates the following incentives:
|
||||
•
|
||||
Faithful acknowledgment of errors is rewarded over false claims of correctness.
|
||||
•
|
||||
The highest rewards come from producing correct proofs and accurately recognizing their rigor.
|
||||
•
|
||||
A good strategy to obtain high rewards for the proof generator is to identify and resolve as many issues as possible before finalizing the response.
|
||||
## 2.3 Synergy Between Proof Verification and Generation
|
||||
The proof verifier and generator create a synergistic cycle:
|
||||
the verifier improves the generator, and as the generator improves, it produces new proofs that challenge the verifier’s current capabilities.
|
||||
These challenging cases – where the verifier may fail to identify issues in a single attempt – become valuable training data for enhancing the verifier itself.
|
||||
To retrain and improve the verifier, we need labeled correctness data for newly generated proofs.
|
||||
Manual annotation, while straightforward, becomes increasingly time-consuming as problems grow harder and errors become more subtle.
|
||||
To boost annotation efficiency, we generated multiple verifier analyses per proof to surface potential issues for human review.
|
||||
From this AI-assisted annotation process, we recognized two facts that make it feasible to push the level of automation a step further:
|
||||
1.
|
||||
Scaling verifier samples increases the probability of catching real issues in flawed proofs.
|
||||
2.
|
||||
Reviewing the verifier’s identified issues is exactly meta-verification , which is easier than identifying issues from scratch. Meta-verification is also more sample-efficient for LLMs to master.
|
||||
Building on these observations, we developed the following automated labeling process:
|
||||
1.
|
||||
For each proof, generate n n independent verification analyses
|
||||
2.
|
||||
For analyses reporting issues (scores 0 or 0.5), generate m m meta-verification assessments to validate the identified problems.
|
||||
An analysis is deemed valid if the majority of meta-assessments confirm its findings
|
||||
3.
|
||||
For each proof, we examine analyses that assign the lowest score.
|
||||
If at least k k such analyses are deemed valid, the proof is labeled with that lowest score.
|
||||
If no legitimate issues are identified across all verification attempts, the proof is labeled with 1.
|
||||
Otherwise, the proof is discarded or routed to human experts for labeling
|
||||
In our last two training iterations, this fully automated pipeline replaced human annotation entirely.
|
||||
Quality checks confirmed that the automated labels aligned well with expert judgments.
|
||||
## 3 Experiments
|
||||
Figure 1: Average proof scores on CNML-level problems by category and model, as evaluated by our verifier.
|
||||
## 3.1 Training Settings
|
||||
We employed Group Relative Policy Optimization (GRPO) ( Shao et al., 2024 ) for reinforcement learning, iteratively optimizing proof verification and generation capabilities as described in Section 2 .
|
||||
In each iteration, we first optimized proof verification. The proof generator was then initialized from the verifier checkpoint and optimized for proof generation.
|
||||
Starting from the second iteration, the proof verifier was initialized with a checkpoint that consolidated both verification and generation capabilities from the previous iteration through rejection fine-tuning.
|
||||
## 3.2 Evaluation Benchmarks
|
||||
We evaluate our final proof generator on the following theorem proving benchmarks:
|
||||
In-House CNML-Level Problems 91 theorem-proving problems spanning algebra (13), geometry (24), number theory (19), combinatorics (24), and inequality (11), comparable in difficulty to problems from Chinese National High School Mathematics League (CNML)
|
||||
Competition Problems
|
||||
•
|
||||
IMO 2025 (6 problems): The International Mathematical Olympiad, the premier global mathematics competition for pre-university students
|
||||
•
|
||||
CMO 2024 (6 problems): The China Mathematical Olympiad, China’s national championship
|
||||
•
|
||||
Putnam 2024 (12 problems): The William Lowell Putnam Competition, the preeminent mathematics competition for undergraduate students in North America
|
||||
•
|
||||
ISL 2024 (31 problems): The IMO Shortlist, a collection of problems proposed by participating countries and considered by the Problem Selection Committee for potential inclusion in IMO 2024
|
||||
•
|
||||
IMO-ProofBench (60 problems): Developed by the DeepMind team behind DeepThink IMO-Gold ( Luong and Lockhart, 2025 ) , this benchmark ( Luong et al., 2025 ) is divided into a basic set (30 problems, pre-IMO to IMO-Medium difficulty) and an advanced set (30 challenging problems simulating complete IMO examinations, up to IMO-Hard level)
|
||||
## 3.3 Evaluation Results
|
||||
## 3.3.1 One-Shot Generation
|
||||
We first evaluate the model’s ability to generate correct proofs without iterative refinement.
|
||||
On the in-house problems, we generated 8 proof samples per problem for each evaluated model.
|
||||
Proof correctness was measured by majority voting across 8 verification analyses produced by our final verifier.
|
||||
As shown in Figure 1 , across all categories of CNML-level problems – algebra, geometry, number theory, combinatorics, and inequality – DeepSeekMath-V2 consistently outperforms GPT-5-Thinking-High ( OpenAI, 2025 ) and Gemini 2.5-Pro ( DeepMind, 2025 ) , demonstrating superior theorem-proving ability across domains.
|
||||
## 3.3.2 Sequential Refinement with Self-Verification
|
||||
Figure 2: Proof quality improvements as the maximum sequential iterations varies from 1 (no refinement) to 8 (initial generation plus up to 7 refinements based on self verification).
|
||||
For challenging problems from competitions like IMO and CMO, models often cannot generate comprehensive and rigorous proofs in a single attempt within the 128K token limit.
|
||||
When this occurs, our proof generator recognizes its proof is invalid through self-verification but lacks the context length to resolve all identified issues in a single attempt.
|
||||
To explore how extended context and self-verification can improve proof quality, we evaluate sequential refinement with self-verification.
|
||||
This approach first generates a proof with self-analysis, then iteratively re-prompts the generator with its previous output (see Appendix A.4 for the refinement prompt), allowing it to address identified issues.
|
||||
The process continues until the generator assigns itself a perfect score or reaches the maximum number of sequential attempts.
|
||||
Figure 2 demonstrates proof quality improvement through sequential refinement on IMO Shortlist 2024 problems.
|
||||
For each problem, we launched 32 independent refinement threads.
|
||||
Proof correctness was measured by majority voting across 32 verification analyses from our final verifier.
|
||||
We report two metrics in Figure 2 : (1) Pass@1 – the average score of the final proof from each thread, and (2) Best@32 – the score of the best proof per problem, selected by self-assigned scores across all threads.
|
||||
The self-selected best proofs achieve significantly higher verification scores than the thread average, demonstrating our generator’s ability to accurately assess proof quality.
|
||||
Furthermore, Pass@1 improves substantially as maximum sequential attempts increase, showing that self-verification effectively guides iterative improvement.
|
||||
These results confirm that our generator can reliably differentiate between high-quality and flawed proofs, and leverage this self-awareness to systematically improve its mathematical reasoning.
|
||||
Figure 3: Expert evaluation results on the Basic and Advanced subsets of IMO-ProofBench. All results are sourced from Luong et al. (2025) , with the exception of DeepSeekMath-V2, which was evaluated by our experts following the grading guidelines.
|
||||
## 3.3.3 High-Compute Search
|
||||
Contest
|
||||
Problems
|
||||
Points
|
||||
IMO 2025
|
||||
P1 , P2 , P3 , P4 , P5
|
||||
83.3%
|
||||
CMO 2024
|
||||
P1 , P2 , P4 , P5 , P6
|
||||
73.8%
|
||||
Putnam 2024
|
||||
A1 ∼ \sim B4 , B5 , B6
|
||||
98.3%
|
||||
Table 1: Problems in gray are fully solved , while underlined problems received partial credit .
|
||||
To solve the most challenging problems, we scaled both verification and generation compute – using extensive verification to identify subtle issues and parallel generation to explore diverse proof strategies.
|
||||
Our approach maintains a pool of candidate proofs for each problem, initialized with 64 proof samples with 64 verification analyses generated for each.
|
||||
In each refinement iteration, we select the 64 highest-scoring proofs based on average verification scores and pair each with 8 randomly selected analyses, prioritizing those identifying issues (scores 0 or 0.5).
|
||||
Each proof-analysis pair is used to generate one refined proof, which then updates the candidate pool.
|
||||
This process continues for up to 16 iterations or until a proof successfully passes all 64 verification attempts, indicating high confidence in correctness.
|
||||
All experiments used a single model, our final proof generator, which performs both proof generation and verification.
|
||||
To validate our results, mathematical experts assessed the highest-scoring proofs.
|
||||
As shown in Table 1 , our approach solved 5 of 6 problems from IMO 2025 and 4 problems plus partial credit on another from CMO 2024, achieving gold medal performance in both pinnacle high-school competitions.
|
||||
On Putnam 2024, the preeminent undergraduate mathematics competition, our model solved 11 of 12 problems completely and the remaining problem with minor errors, scoring 118/120 and surpassing the highest human score of 90.
|
||||
Figure 3 shows the results on IMO-ProofBench.
|
||||
Our approach outperforms DeepMind’s DeepThink (IMO Gold) on the basic set and remains competitive on the advanced set, while substantially outperforming all other baselines.
|
||||
We observe that the hardest IMO-level problems remain challenging for our model.
|
||||
Notably, for problems not fully solved, our generator typically identifies the genuine issues in its proofs, while fully solved problems pass all 64 verification attempts.
|
||||
This demonstrates that we can successfully train LLM-based verifiers to assess proofs previously considered difficult to verify automatically.
|
||||
By scaling test-time compute under verifier guidance, our model solves problems that require hours of effort from human competitors.
|
||||
## 4 Related Work
|
||||
Reasoning models ( OpenAI, 2024 ; Guo et al., 2025 ) have saturated quantitative reasoning benchmarks like AIME and HMMT within one year.
|
||||
This rapid advancement is partly attributed to the well-defined evaluation criterion: if we care only about final answers, then quantitative reasoning is easy to verify.
|
||||
However, this final answer metric is inapplicable to theorem proving, which often requires no numerical answers but demands rigorous step-by-step derivation.
|
||||
Informal mathematical proofs have long been considered hard to verify automatically, lacking reliable approaches to assess proof correctness.
|
||||
Recent developments suggest this barrier may be surmountable.
|
||||
Models like Gemini-2.5 Pro already demonstrate a certain level of self-verification capabilities, which can refine their own solutions to improve quality ( Huang and Yang, 2025 ) .
|
||||
More significantly, DeepMind’s internal DeepThink variant ( Luong and Lockhart, 2025 ) achieved gold medal performance at IMO 2025 using pure natural language reasoning – serving as an existence proof that LLM-based verification of complex proofs is achievable.
|
||||
Recent research has begun exploring whether reasoning models can evaluate proofs, both with and without reference solutions ( Dekoninck et al., 2025 ; Luong et al., 2025 ) , showing promising results.
|
||||
In this work, we open source DeepSeekMath-V2 and our training methodology as concrete steps toward self-verifiable mathematical reasoning, showing how models can learn to verify and improve their own proofs.
|
||||
Proof assistants like Lean ( de Moura et al., 2015 ) and Isabelle ( Paulson, 1994 ) offer a reliable approach to verify proofs – proofs must be written in formal language, but once compiled, correctness is guaranteed.
|
||||
AlphaProof ( AlphaProof and teams, 2024 ; Trinh et al., 2024 ; Chervonyi et al., 2025 ) , a system specialized for formal proof search, achieved silver-level performance at IMO 2024 but required intensive computation.
|
||||
While using informal reasoning to guide formal proof generation has been explored extensively ( Jiang et al., 2023 ) , recent reasoning models have dramatically improved informal reasoning quality, making this guidance far more effective.
|
||||
Systems like DeepSeek-Prover-V2 ( Ren et al., 2025 ) and Seed-Prover ( Chen et al., 2025 ) can now produce substantially more valid formal proofs within the same computational budget, with Seed-Prover solving 5 of 6 problems at IMO 2025.
|
||||
Notably, these results were achieved without specifically optimizing the informal reasoning components for theorem proving tasks.
|
||||
We believe advancing natural language theorem proving will significantly benefit formal reasoning.
|
||||
We hope to contribute toward truly reliable mathematical reasoning systems that leverage both informal insights and formal guarantees to advance mathematical research.
|
||||
## 5 Conclusion
|
||||
We presented DeepSeekMath-V2, a model capable of both generating and verifying mathematical proofs.
|
||||
By training models to identify issues in their own reasoning and incentivizing them to address these issues before finalizing outputs, we move beyond the limitations of final-answer-based rewards toward self-verifiable mathematical reasoning.
|
||||
Our iterative training process – alternating between improving verification capabilities and using these to enhance generation – creates a sustainable cycle where each component drives the other forward.
|
||||
Our key technical contributions include: (1) training an accurate and faithful LLM-based verifier for mathematical proofs, (2) using meta-verification to largely reduce hallucinated issues and ensure verification quality, (3) incentivizing the proof generator to maximize proof quality through self-verification, and (4) scaling verification compute to automatically label increasingly hard-to-verify proofs to improve the verifier without human annotation.
|
||||
DeepSeekMath-V2 demonstrates strong performance on competition mathematics.
|
||||
With scaled test-time compute, it achieved gold-medal scores in high-school competitions including IMO 2025 and CMO 2024, and a near-perfect score on the undergraduate Putnam 2024 competition.
|
||||
This work establishes that LLMs can develop meaningful self-evaluation abilities for complex reasoning tasks.
|
||||
While significant challenges remain, we hope this research direction contributes to the goal of creating self-verifiable AI systems that can solve research-level mathematics.
|
||||
|
||||
=== BODY (tail) ===
|
||||
Reference in New Issue
Block a user