Geometric problem-solving (GPS) has been a long-standing challenge in the fields of formal mathematics and artificial intelligence. To address the limitations of unidirectional approaches, we developed a neuro-symbolic system that integrates forward and backward reasoning. The neural component employs a gating-enhanced attention network to select candidate theorems, guiding the heuristic search and pruning irrelevant branches. The symbolic component is a bidirectional solver built on FormalGeo, which performs rigorous geometric relational reasoning and algebraic computation. The neural component predicts the theorems based on the current problem state, while the symbolic component applies these theorems and updates the problem state. These two parts interact iteratively until the problem is solved. The solving process is organized as a graph structure where facts and goals serve as nodes and theorems as edges, thereby generating a human-readable solution. The proposed neuro-symbolic system achieved an 89.63% problem-solving success rate (PSSR) on the FormalGeo7K dataset, surpassing the previous best result.

Geometric problem-solving has always been a great challenge in the field of deductive reasoning and artificial intelligence. Symmetry is a defining characteristic of geometric shapes and properties. Consequently, the application of symmetry principles to geometric reasoning arises as a natural choice. To address the efficiency degradation and limited generalization, we propose FGeo-ISRL, a neural-symbolic inverse search framework whose core is the synergistic integration of a task-fine-tuned large language model and Monte Carlo Tree Search. Under the formal framework of FormalGeo, geometric theorems are iteratively applied starting from the given conditions and the target conclusion, in order to infer the necessary supporting premises. The large language model simultaneously serves as a policy network and a value network, guiding theorem application decisions and evaluating intermediate proof states, whereas the Monte Carlo Tree Search performs structured exploration over the state space, both training for policy refinement and inference for online search. The reinforcement learning agent is trained with a hybrid reward scheme, combining immediate feedback from the value difference and a sparse success reward. Experiments demonstrate the effectiveness and correctness of FGeo-ISRL. It not only achieves a Single-Step Theorem Accuracy of 90.2% and a Geometric Problem-Solving Accuracy of 83.14%, but also ensures that every step of the proof process remains readable, verifiable, and traceable.

Euclidean plane geometry problem solving is a challenging benchmark for artificial intelligence because it requires complex diagram understanding, symbolic deduction, and multi-step reasoning. Constructing effective datasets for this task requires geometric instances that are realizable, non-degenerate, structurally diverse, and paired with human-like proofs. However, existing random or template-based generation pipelines often produce redundant, singular, or infeasible candidates, causing substantial computation to be spent before useful reasoning trajectories can be extracted. To address these limitations, we present FGeo-GCG, a hybrid geometric data synthesis framework built on the FormalGeo-V2 deductive engine. It formulates Geometric Configuration Generation as an incremental linear construction process that decomposes global constraint satisfaction into local construction steps, thereby pruning invalid branches during the generation process. To improve reliability and efficiency, FGeo-GCG combines two validation stages: a safe stochastic Jacobian-rank filter estimates whether local candidate constraints contribute independent algebraic restrictions, and progressive geometric validation checks whether the resulting partial construction remains realizable and non-degenerate. By encoding incidence-, metric-, and symmetry-related dependencies within unified constraint graphs, the framework also connects geometric data synthesis with structural symmetry analysis. Validated constraint graphs are then converted into problem instances through forward deduction, goal decomposition, and multi-dimensional complexity filtering, producing proof targets without manual annotation. Experiments show that the full validation pipeline reduces the failure rate for highly constrained instances. The resulting FGeo-GCG dataset contains more than 50,000 formally validated plane geometric configurations and provides engine-derived reasoning traces and targets for future training and evaluation of neuro-symbolic geometry problem-solving systems.