Geoint-R1: Multimodal Geometric Reasoning

• Geoint-R1 is a multimodal framework that integrates textual descriptions and visual diagrams to create formally verified geometric proofs using Lean4.
• It employs a two-stage training pipeline—supervised fine-tuning followed by reinforcement learning—to optimize auxiliary construction and proof verification.
• Interactive TikZ visualization synchronizes evolving proofs with diagram updates, enhancing clarity in geometric reasoning and auxiliary element generation.

Searching arXiv for the Geoint-R1 paper and closely related work for context. Geoint-R1 is a multimodal reasoning framework for formal geometric reasoning from textual descriptions and visual diagrams. It was introduced to generate formally verifiable geometric solutions by combining auxiliary elements construction, formal reasoning represented via Lean4, and interactive visualization. In the reported experiments, the framework extends a standard multimodal LLM, specifically Qwen2.5-VL-7B, and is evaluated on the Geoint benchmark, a corpus of 1,885 rigorously annotated geometry problems spanning plane, spatial, and solid geometry, including 429 problems requiring auxiliary lines and 1,456 without them (Wei et al., 5 Aug 2025).

1. Problem setting and motivating gaps

Geoint-R1 is motivated by two difficulties that prior multimodal geometry systems did not adequately resolve. The first is auxiliary-element identification and construction. Human solvers routinely add auxiliary lines or points, such as drawing a parallel or dropping a perpendicular, to unlock a proof. The framework is explicitly designed around the observation that existing vision-LLMs rarely propose such constructions, leaving many problems unsolved (Wei et al., 5 Aug 2025).

The second difficulty is formal verifiability and visual alignment. A natural-language proof outline may be correct in spirit while still lacking the rigor of a proof assistant. The paper also identifies a missing tight coupling between the diagram and the formal proof steps. Geoint-R1 therefore targets solutions that can be checked by Lean4 while simultaneously producing a diagram, via TikZ, that highlights each construction (Wei et al., 5 Aug 2025).

This formulation places Geoint-R1 at the intersection of multimodal reasoning, symbolic verification, and diagram-grounded proof generation. A central implication of the design is that geometric reasoning is treated not merely as answer prediction, but as a structured proof-generation task in which auxiliary constructions are first-class objects.

2. Architecture and learning pipeline

The system architecture augments a standard multimodal LLM with three major components: a vision encoder with multimodal projector, a two-stage training pipeline, and an interactive visualization layer (Wei et al., 5 Aug 2025).

Component	Role
Vision encoder + multimodal projector	Fuses the diagram image $I_i$ and text prompt $T_i$ into a joint representation
Two-stage training pipeline	Learns reasoning steps, Lean4 auxiliary construction code, and final answers
Interactive visualization	Updates a TikZ-rendered diagram as constructions are emitted

In the multimodal front end, the input diagram image $I_i$ is tokenized by the frozen vision encoder, while the problem statement $T_i$ is tokenized by the language backbone. A cross-modal projector fuses the two streams into a joint representation (Wei et al., 5 Aug 2025).

Training proceeds in two stages. In supervised fine-tuning, the model is trained on the Geoint dataset to emit a triplet $(P_i, C_i, A_i)$, where $P_i$ is a sequence of reasoning steps in natural language and Lean4 comments, $C_i$ is the Lean4 code for auxiliary constructions, and $A_i$ is the final numeric or textual answer. In reinforcement learning, model roll-outs are scored by a learned verifier that checks answer correctness, completeness and logical validity of the Lean4 proof, accuracy of the auxiliary-line code, and strict format compliance (Wei et al., 5 Aug 2025).

This training design makes auxiliary construction and proof verification explicit optimization targets rather than incidental by-products of answer supervision.

3. Lean4 formalization and dynamic auxiliary constructions

A defining feature of Geoint-R1 is its representation of geometric objects and relations in Lean4. The paper describes a representative style in which points, lines, circles, incidence, and parallelism are encoded within Lean4, with the full dataset using the Lean4 mathlib geometry library (Wei et al., 5 Aug 2025).

For example, the formalization of parallelism is represented by a cross-product equality constraint: $$(x_B - x_A)(y_D - y_C) \;=\; (y_B - y_A)(x_D - x_C).$$

A typical theorem statement shown in the system description is a midline-parallel result: $T_i$9

At the core of the framework is dynamic auxiliary element construction. When the current proof state demands a new construction, such as a parallel through a given point, the model emits Lean4 code to declare the object and prove the relevant relation. The paper illustrates this with the case of establishing a line through $E$ parallel to $T_i$0, expressed through Lean4 declarations such as def ℓ_par and the lemma aux_parallel (Wei et al., 5 Aug 2025).

The workflow is described in four steps:

1. The model identifies the need for an auxiliary object.
2. It declares the new geometric object in Lean4.
3. It states and proves the corresponding lemma, such as parallelism.
4. It reuses the new object in subsequent proof steps.

This mechanism distinguishes Geoint-R1 from systems that only verbalize a proof plan. Here, auxiliary constructions become executable formal artifacts that can be checked and reused.

4. Visualization layer and benchmark design

Geoint-R1 includes an interactive visualization component built around TikZ rendering. As the model emits each construction $T_i$1, a TikZ renderer updates the diagram and highlights newly added lines or points. The interface allows a user to step through the proof, and clicking a Lean4 lemma or tactic highlights the corresponding region of the diagram (Wei et al., 5 Aug 2025).

This visual synchronization is designed to align the evolving formal proof state with the evolving geometric figure. The result is not only proof generation, but proof inspection.

The associated Geoint benchmark was introduced to support systematic evaluation of formal multimodal geometric reasoning. It contains 1,885 geometry question-answer pairs and covers nine knowledge types, including Triangle, Quadrilateral, Circle, and Solid Geometry. Each entry includes the text prompt $T_i$2, the original diagram in image and TikZ form, verified Lean4 construction code $T_i$3, a reference proof $T_i$4, and a reference answer $T_i$5 (Wei et al., 5 Aug 2025).

The annotation process is explicitly expert-verified. Diagrams are converted to TikZ and re-rendered; human annotators ensure fidelity; DeepSeek-R1 translates TikZ to a Lean4 skeleton; and three math experts manually complete and verify each proof. When problems are multi-part, they are decomposed so that each benchmark entry contains a single question with all necessary prior context (Wei et al., 5 Aug 2025).

This benchmark design emphasizes formal correctness, diagram fidelity, and explicit treatment of auxiliary constructions, which are often absent from earlier multimodal math datasets.

5. Experimental methodology and empirical results

The training and evaluation protocol is structured around both answer quality and proof quality. During supervised fine-tuning, the vision encoder is frozen and the LLM is fine-tuned with cross-entropy on the tokens of $T_i$6. During reinforcement learning, the model generates $T_i$7 candidates per problem, rejects trivial or unsolvable outputs, computes a curriculum score $T_i$8, and optimizes expected verification reward via GRPO (Wei et al., 5 Aug 2025).

The evaluation includes open-source multimodal baselines such as Phi-3.5-V-4B, LLaVA-7B, InternLM-XComposer, Yi-VL, InternVL3, and Qwen-VL; closed-source baselines including GPT-4o and Gemini-1.5-pro/flash; and math-specific systems such as MMR1-Math-7B and Math-LLaVA-13B (Wei et al., 5 Aug 2025).

Two principal metric families are reported. For answer tasks, evaluation uses exact or equivalent answer match. For proof tasks, evaluation uses a five-dimensional process score with weights of 30% for logical validity, 20% for completeness, 20% for correctness, 20% for auxiliary construction, and 10% for clarity (Wei et al., 5 Aug 2025).

The main quantitative results reported are as follows:

Evaluation	Result
Overall accuracy	64.72%
Answer score	57.01%
Proof score	72.43%

On auxiliary-line problems specifically, Geoint-R1 reaches 68.63% on answer-based evaluation and 68.40% on proof-based evaluation, while most compared models remain below 40% on these cases (Wei et al., 5 Aug 2025). The paper also reports that Geoint-R1 significantly surpasses existing multimodal and math-specific reasoning models, particularly on challenging problems requiring explicit auxiliary element constructions (Wei et al., 5 Aug 2025).

The ablation results further indicate that removing the verification reward, RL fine-tuning, or curriculum learning causes an approximately 8–12 percentage point drop. The paper interprets this as evidence that all modules are essential, especially auxiliary verification (Wei et al., 5 Aug 2025).

6. Interpretation, limits, and related systems

Geoint-R1 is specifically a framework for formal geometric reasoning, not a geospatial-reasoning model. It should not be confused with Geo-R1, which is a reasoning-centric post-training framework for vision-LLMs in geospatial reasoning and cross-view pairing tasks (Xu et al., 29 Sep 2025). The similarity in naming masks a difference in domain: Geoint-R1 addresses theorem proving from diagrams and text, whereas Geo-R1 addresses geospatial reasoning over panorama and satellite imagery (Xu et al., 29 Sep 2025).

Within its own domain, Geoint-R1 addresses a common misconception in multimodal mathematical reasoning: the primary bottleneck is not only reading the diagram, but also deciding when and how to introduce auxiliary constructions and how to bind those constructions to a formally checkable proof state. The framework makes both issues explicit through Lean4 code generation and verifier-driven optimization (Wei et al., 5 Aug 2025).

The reported error analysis identifies remaining gaps in strict theorem application and variable tracking. The paper points to future work in deeper formal-theory integration and stronger symbolic capacity. It also suggests extensions to automated lemma discovery, richer 3D geometry formalisms in Lean4, and tighter human-AI collaboration through the interactive interface (Wei et al., 5 Aug 2025).

A plausible implication is that Geoint-R1 represents a shift from answer-centric multimodal math systems toward proof-centric systems in which diagram interpretation, auxiliary construction, symbolic formalization, and user-facing visualization are treated as a single integrated problem.