+- MKLab (https://mklab.gr)
+-- Forum: [INDEX] (https://mklab.gr/forumdisplay.php?fid=1)
+--- Forum: ARTFICIAL INTELLIGENCE (AI) (https://mklab.gr/forumdisplay.php?fid=5)
+---- Forum: ARTICLES (https://mklab.gr/forumdisplay.php?fid=33)
+---- Thread: China’s TongGeometry Is Redefining AI Reasoning (/showthread.php?tid=2179)
China’s TongGeometry Is Redefining AI Reasoning - mklabgr - 09-25-2026
TongGeometry is an AI system designed to tackle a particularly difficult form of mathematical reasoning: Olympiad-level Euclidean geometry. Unlike conventional AI systems that mainly solve problems supplied by humans, TongGeometry has a dual role: it can propose new geometry problems and prove them. The system combines a symbolic geometry engine with guided tree search and fine-tuned language models, allowing it to explore possible constructions and proof paths systematically. The underlying research was published in Nature Machine Intelligence in January 2026.
A striking feature is the scale of its exploration. Using statistics from 196 human Olympiad problems, TongGeometry generated a repository containing approximately 6.7 billion geometry theorems requiring auxiliary constructions, of which about 4.1 billion displayed geometric symmetry. The system does not simply generate random configurations: existing competition problems provide statistical guidance toward configurations that are difficult, elegant and suitable for Olympiad mathematics. Three AI-generated problems subsequently received real-world recognition, including selection for a Chinese national-team qualifying competition and shortlisting for a US mathematical competition.
On problem solving, TongGeometry uses a neuro-symbolic architecture. One learned model helps identify promising auxiliary constructions, while another estimates how much reasoning remains on a particular search path. The symbolic engine then performs deductive reasoning through geometric relations rather than depending entirely on numerical computation. For example, angle-chasing can be represented through logical relationships such as $ \angle ABC=\angle CDE $, from which additional angle equalities can be derived. On the IMO-AG-30 benchmark, TongGeometry solved all 30 geometry problems within about 38 minutes using 32 CPU cores and a single GPU, although the authors explicitly caution that this benchmark result should not be interpreted as meaning that the system generally surpasses human IMO contestants at geometry.
The deeper significance is therefore not simply that AI can solve difficult geometry problems. TongGeometry attempts to automate part of the creative mathematical process: exploring a structured mathematical space, discovering relationships, identifying elegant configurations and turning them into problems that humans consider competition-worthy. Its use of symmetry and canonical representations also reduces duplicated search paths, making the exploration substantially more efficient. For mathematics education, the system has already been used to generate problems that are reviewed and refined by experienced Olympiad coaches, suggesting a possible future role for AI as a problem-generation and teaching assistant, rather than merely an answer generator.
Key takeaways
- TongGeometry is both a solver and a problem generator, rather than merely a system answering predefined questions.
- It combines large language models + symbolic reasoning + guided tree search.
- It explored around $6.7$ billion geometry theorems involving auxiliary constructions.
- About $4.1$ billion of these exhibited geometric symmetry.
- It solved all 30 problems in the IMO-AG-30 geometry benchmark within about 38 minutes on relatively accessible hardware.
- Its reasoning can be expressed through formal geometric relations such as $ \angle ABC=\angle CDE $ and derived angle-chasing rules.
- Several AI-generated problems were considered for actual high-level mathematical competitions.
- The system rediscovered some classical geometric lemmas, suggesting that systematic search can converge toward mathematically meaningful structures.
- A major innovation is its treatment of symmetry and canonical representations, which reduces redundant search.
- For teachers, an important application is AI-assisted generation of challenging geometry problems, subsequently checked and refined by humans.
- The main limitation is domain specificity: success in structured Euclidean geometry does not by itself establish general mathematical intelligence or AGI.
- The most interesting shift is from AI as a “student” solving human-created problems toward AI as a “coach” helping discover and formulate new mathematics.
ARTICLE