Red Hot Cyber
Cybersecurity is about sharing. Recognize the risk, combat it, share your experiences, and encourage others to do better than you.
Redazione RHC : 30 November 2025 08:49
The Chinese company DeepSeek has introduced a new specialized model for solving mathematical problems , DeepSeekMath-V2. This large-scale language model, specifically designed for theorem proving and Olympiad problems, is unique in that it not only produces answers but also verifies the correctness of its own reasoning.
DeepSeekMath-V2 essentially answers an age-old question in artificial intelligence: how to ensure that a model has arrived at the correct solution fairly, rather than guessing the outcome or finding a shortcut that is incorrect . Most modern models are trained to arrive at the correct final answer more frequently, for which they are rewarded with a reward system similar to reinforcement learning.
But in mathematics, this isn’t enough: in many problems, the answer itself isn’t as important as a rigorous and transparent proof . The authors explicitly state that an exact final result doesn’t guarantee the correctness of the reasoning, and for theorems, there’s no predefined “ correct number ” to verify.
DeepSeekMath-V2 is based on the experimental DeepSeek-V3.2-Exp-Base . The team trains a separate verification model that evaluates mathematical proofs, checking for logical gaps and errors step by step, and then uses this verifier as a ” judge ” for the main proof generator model.
The generator receives a reward not only for the correct final answer , but also for its reasoning’s ability to pass a rigorous validation test. If the test fails, the model is rewarded for attempting to independently identify the weaknesses of its solution and rewrite the proof so that it passes the validation test.
To prevent the system from crashing as the generator becomes smarter than the verifier, developers scale the computational resources for the verifier separately. The verifier learns from increasingly complex and difficult-to-verify examples generated by the model itself as its capabilities increase. This closed loop of ” generating, verifying, and improving the verifier ” helps bridge the skill gap between the two parts of the system and preserves its ability to self-correct.
The results are impressive. In a GitHub post, the team claims that DeepSeekMath-V2 won gold at the 2025 International Mathematical Olympiad and the 2024 Chinese Mathematical Olympiad, and at the 2024 Putnam Mathematical Competition , the model scored 118 out of 120 points using scalable computation in the solution phase.
In the specialized benchmark IMO-ProofBench, developed by the Google DeepMind team for its Gemini DeepThink model, DeepSeekMath-V2, according to an independent technical analysis, outperforms DeepThink in basic tests.
Informal score reports published by researchers and enthusiasts provide more specific data: DeepSeekMath-V2 scores around 99% in the basic part of IMO-ProofBench and 61.9% in the advanced part. The report itself claims that this result surpasses the performance of the GPT-5 and Gemini models on this set of tasks, although this is not an official ranking, but rather a comparison of individual tests.
Another important point for the community: DeepSeekMath-V2 is being touted as the first open-source mathematical AI to achieve gold-level performance on IMO-level problems. This news has already been reported on specialized forums, where links to the paper and the model weights are being posted.
The model is available on GitHub and Hugging Face . The repository is hosted on Apache 2.0 , and the models themselves are licensed under a separate license that governs their use, including commercial use. The launch and open source nature of DeepSeekMath-V2 are further announced in specialized blogs and social media posts, emphasizing that the weights can be freely downloaded and run on your own hardware, subject to the terms of the model license.
For now, DeepSeekMath-V2 remains a highly specialized but highly illustrative example of how artificial intelligence is shifting its focus from “guessing the correct answer” to controlling the model’s thinking process . And the enthusiastic response to this new product from developers, researchers, and competitive math enthusiasts demonstrates that the race is now on not only for general intelligence, but also for the quality and verifiability of reasoning.
Redazione