BiNSGPS: Geometry Problem Solving via Bidirectional Neuro-Symbolic Interaction
BiNSGPS uses bidirectional neuro-symbolic feedback, achieving 90.5% on Geometry3K and 90.1% on PGPS9K.
Key Findings
Methodology
BiNSGPS combines Multimodal Representations Alignment, a Symbolic Solver, and an MLLM Adviser. PGDPNet parses visual primitives while the MLLM parses text, producing representation set L. The solver performs hypergraph expansion under theorem library R. Conflicts trigger correction or deletion in L; deductive deadlocks trigger auxiliary-hypothesis generation H, followed by symbolic verification.
Key Results
- Completion accuracy reaches 90.5% on Geometry3K (3,001 problems), exceeding GPT-5.2’s 77.9%, and 90.1% on PGPS9K (9,022 problems). In Choice mode, BiNSGPS reaches 95.2%, outperforming the next-best method by 10.7 and 10.9 percentage points on the two datasets.
- The reported Step-wise Logical Coherence is 96%, indicating that the output is not merely answer-correct but substantially aligned with valid derivations. Hypergraph nodes, theorem hyperedges, and parent pointers preserve an auditable proof trace.
- The feedback loop handles two failure states: representation conflicts invoke correction, while theorem deadlocks invoke auxiliary hypotheses. PGDPNet reports over 99% accuracy for element and positional-relation recognition; however, the paper excerpt does not provide a complete ablation table or per-mode gains.
Significance
The work addresses a persistent tension in geometry AI: symbolic systems are rigorous but brittle and rule-bound, whereas neural systems are flexible but hallucination-prone. BiNSGPS turns the solver into an active diagnostic partner that can send evidence back to perception and formalization. Its broader significance is architectural: an MLLM supplies multimodal interpretation, strategy, and geometric intuition, while symbolic execution supplies constraints, verification, and proof traceability. This offers a practical route toward more trustworthy multimodal reasoning agents.
Technical Contribution
The principal contribution is an agentic tool-calling architecture replacing the conventional neural-to-symbolic pipeline with a closed loop. The solver represents facts as hypergraph nodes and theorems as hyperedges, and emits diagnostics when contradictions or deadlocks occur. In rectification mode, the Adviser is constrained to modify or delete existing representations; in hypothesis mode, it generates H under a Rationale-Conclusion protocol. Successful traces are reduced to a minimal solution graph and verbalized by the MLLM.
Novelty
Compared with predominantly one-way systems such as AlphaGeometry, AlphaGeometry2, AutoGPS, and GeoDRL, BiNSGPS makes the symbolic engine an active conversational component. Its fundamental novelty is the explicit separation of feedback into representation rectification and auxiliary-hypothesis generation, targeting perceptual errors and incomplete theorem coverage respectively. The contribution is therefore interaction design and self-correction, rather than simply scaling the neural model.
Limitations
- Performance depends on PGDPNet’s perception quality, MLLM parsing, and theorem-library coverage. Ambiguous diagrams, out-of-distribution configurations, or difficult auxiliary constructions may cause repeated feedback or failure.
- The excerpt reports aggregate accuracy and 96% coherence but lacks full ablations, iteration counts, latency, hardware cost, and failure distributions, limiting precise attribution of gains to individual components.
- After the iteration limit, the system falls back to an MLLM-generated Analyze-Answer-Proof response, which may not retain the strict verifiability of symbolic proof.
Future Work
Future work should quantify how feedback rounds, theorem-library size, and hypothesis quality affect accuracy and cost, with reproducible ablations. Extensions to solid geometry, open-ended proofs, and learned auxiliary constructions are natural. Formal checking of generated hypotheses, confidence-ranked search, and replacing fallback answers with proof-producing verifiers could further reduce hallucination while preserving adaptability.
AI Executive Summary
Geometry problem solving tests whether an AI system can jointly understand diagrams, language, and formal deduction. Symbolic methods are precise but collapse when initial formalization is wrong and cannot easily exceed their theorem libraries. Neural and multimodal large language models are adaptable, yet frequently produce plausible-looking invalid reasoning. Systems such as AlphaGeometry, AutoGPS, and GeoDRL improve the interface between the two paradigms, but generally retain one-way information flow.
BiNSGPS introduces Bidirectional Neuro-Symbolic Interaction. PGDPNet first extracts points, lines, circles, and positional relations; an MLLM Adviser aligns these visual facts with textual constraints and forms representation set L. The Symbolic Solver expands a hypergraph using theorem library R. If it detects contradictory angle assignments or other conflicts, the Adviser repairs or deletes representations. If expansion reaches a deadlock, the Adviser proposes auxiliary hypotheses H, which the solver must validate. A minimal proof subgraph is then verbalized into a readable solution.
On Geometry3K and PGPS9K, Completion accuracy is 90.5% and 90.1%, while Choice accuracy reaches 95.2%, exceeding the next-best method by 10.7 and 10.9 points. Geometry3K performance surpasses GPT-5.2’s 77.9%, and step-wise logical coherence reaches 96%. The approach still depends on PGDPNet and predefined rules, and the excerpt lacks complete ablations and cost analysis. Its central lesson is that reliable multimodal reasoning may require not only neural models that call formal tools, but formal tools that can call neural models back.
Deep Analysis
Background
Geometry reasoning has progressed from specialized neural networks to symbolic engines and hybrid systems. PGPSNet, LANS, and G-LLaVA emphasize visual or learned intuition; InterGPS and E-GPS formalize geometry for rigorous deduction; AlphaGeometry and AlphaGeometry2 combine language models with symbolic engines. Yet neural systems hallucinate, symbolic systems are sensitive to parsing errors, and most neuro-symbolic frameworks still pass information forward only once.
Core Problem
Given diagram D and text T, a solver must recognize structure, align constraints, select theorems, and produce both an answer and a valid proof. Difficulties include perceptual mistakes, ambiguous references, missing implicit facts, and theorem-library incompleteness. In a one-way pipeline, early errors become irreversible and a symbolic deadlock provides no mechanism for neural intuition to repair or extend the reasoning state.
Innovation
- ��Bidirectional feedback lets the solver report conflicts and deadlocks. •Representation rectification lets the Adviser modify or delete suspicious elements of L, anchored by PGDPNet facts B and P. •Auxiliary hypotheses H extend reasoning when R is insufficient. •Proof extraction back-traces a minimal solution graph, while the MLLM only verbalizes it. This differs from prior one-way systems by making verification an active source of corrective information.
Methodology
- ��Input: diagram D and text T. •Standardization: a small model labels otherwise unlabeled points, yielding D′. •Dual extraction: PGDPNet parses points, lines, circles, and relations such as PointLiesOn; few-shot MLLM prompting parses textual assertions. •Integration: the MLLM produces L while prioritizing high-fidelity visual facts. •Completion and solving: implicit properties are added, then a hypergraph is expanded with theorem library R; every conclusion stores parent nodes and the applied rule. •Feedback: conflicts invoke correction, deadlocks invoke H. •Output: recursive back-tracing yields a minimal proof, which the MLLM translates into readable text; iteration-limit cases use Analyze-Answer-Proof fallback.
Experiments
Evaluation uses Geometry3K with 3,001 geometry problem-image pairs and PGPS9K with 9,022 pairs; each question has four choices. Choice mode selects an option, whereas Completion mode requires an independently derived numerical answer. Human evaluation measures Step-wise Logical Coherence. Comparisons include neural, symbolic, MLLM, and neuro-symbolic systems; the excerpt explicitly reports GPT-5.2 and PGDPNet figures but does not reproduce a complete ablation protocol or all baseline values.
Results
BiNSGPS reaches 95.2% in Choice mode, leading the next-best approach by 10.7 points on Geometry3K and 10.9 on PGPS9K. Completion accuracy is 90.5% and 90.1%, with Geometry3K substantially above GPT-5.2’s 77.9%. PGDPNet’s recognition accuracy exceeds 99%, and step-wise logical coherence is 96%. The pattern supports combining specialized perception, symbolic verification, and MLLM-driven feedback rather than relying on a pure neural solver.
Applications
The framework could support geometry tutoring, competition-problem grading, explainable mathematical QA, and textbook-proof generation. Deployment requires a reliable diagram parser, a curriculum-specific theorem library, and symbolic validation of generated hypotheses. Its tool-calling design may also transfer to physics diagrams, engineering sketches, and scientific chart reasoning, provided that domain representations and rules are formalized.
Limitations & Outlook
The approach assumes stable recognition of basic visual facts and adequate theorem coverage. Blurred diagrams, occlusion, complex auxiliary lines, and out-of-distribution configurations may cause failure or repeated feedback. Iterative calls increase latency and inference cost, while fallback MLLM answers are not guaranteed to be formally proved. The excerpt does not report average iterations, hardware, full ablations, or statistical significance; these are important for reproducibility and deployment planning.
Plain Language Accessible to non-experts
Imagine a geometry-solving factory. The first worker inspects a picture and identifies points, lines, and circles. The second reads the written instructions and turns them into clear statements. The third worker is a strict quality inspector who checks every proposed step against the factory rulebook.
In an ordinary factory, information moves only forward. If the first worker mistakes a chord for a tangent, later workers may confidently build the entire product around that mistake. BiNSGPS adds a return conveyor belt. When the inspector detects two incompatible statements, the problem is sent back for correction. If the rulebook contains no useful next step, the second worker may suggest an extra fact, but the inspector must test it before accepting it.
Once the answer is found, the factory keeps only the steps actually needed and turns them into a clean explanation. This division of labor is powerful: flexible workers interpret pictures and suggest ideas, while the inspector prevents unsupported claims from becoming final answers. The paper’s results show that this cooperation can be both accurate and logically disciplined.
ELI14 Explained like you're 14
Think of solving a geometry problem like clearing a tricky game level. You have three teammates: one reads the map, one understands the mission text, and one acts like a referee checking whether every move follows the rules. The map-reader is PGDPNet, the mission strategist is the MLLM Adviser, and the referee is the Symbolic Solver.
In a normal team, messages move only forward. If someone mistakes a chord for a tangent at the beginning, everyone may keep building on the wrong idea. BiNSGPS lets the referee talk back: “These facts clash—please inspect that step again!” If the rulebook has no next move, the Adviser can suggest a new geometric fact, but the referee checks it before it enters the game.
The system records the reasoning like a game quest tree. After finding the answer, it keeps the shortest useful path and rewrites it as a readable proof. On Geometry3K it solved 90.5% of completion questions, and on PGPS9K it solved 90.1%; its step-by-step logic coherence was 96%. That is impressive! But extremely messy diagrams can still confuse it, and repeated checking costs extra time.
The big idea is simple: let creative teammates suggest moves, but let a strict referee approve them. Could this work beyond geometry? Maybe in physics or engineering—but those games would need their own rulebooks and reliable map-readers!
Glossary
Bidirectional Neuro-Symbolic Interaction
A two-way exchange between a neural model and a symbolic reasoner. Unlike a one-way pipeline, the reasoner can send diagnostic evidence back to the neural component.
This is the defining mechanism of BiNSGPS.
MLLM Adviser
A multimodal large language model that interprets diagrams and text, coordinates tools, repairs representations, and proposes hypotheses. It supplies flexible strategy rather than final unchecked deduction.
It orchestrates PGDPNet and the Symbolic Solver.
Symbolic Solver
A rule-based engine that derives conclusions from formal geometric facts. In this paper it uses hypergraph expansion and records theorem dependencies.
It validates reasoning and diagnoses failure.
PGDPNet
A specialized neural diagram parser for geometric elements and positional relations. The paper reports recognition accuracy above 99%.
It provides high-fidelity visual anchors for alignment.
Hypergraph Expansion
A reasoning representation in which facts are nodes and theorems connect multiple premises to a conclusion. Applying rules adds new nodes while preserving provenance.
This is the Symbolic Solver’s deduction mechanism.
Auxiliary Hypothesis
A supplementary formal fact proposed when the existing theorem library cannot continue the proof. The hypothesis is not trusted automatically; the solver checks it for consistency.
It is used to overcome deductive deadlocks.
Open Questions Unanswered questions from this research
- 1 The independent benefit of feedback over one-way systems remains unclear because the excerpt lacks complete ablations, iteration statistics, and categorized failures.
- 2 How to guarantee that auxiliary hypotheses are complete, low-hallucination, and formally provable remains open; larger open-ended benchmarks and proof checkers are needed.
- 3 Robustness to blurred, occluded, or three-dimensional diagrams, and transfer across languages and educational curricula, is not sufficiently evaluated.
Applications
Immediate Applications
Interactive geometry tutoring
Educational platforms could use BiNSGPS to read a problem diagram and text, then provide verified step-by-step hints rather than only an answer. Deployment requires a curriculum-aligned theorem library, PGDPNet-style parsing, and an interface that exposes corrected reasoning.
Automated contest grading
Schools and problem banks could evaluate both multiple-choice and numerical solutions, using symbolic traces to identify inconsistent steps. Human review should remain available for ambiguous diagrams, unsupported hypotheses, and fallback outputs.
Long-term Vision
Verifiable scientific reasoning agents
The same feedback architecture could connect foundation models with physics, engineering, or 3D-geometry solvers. Major obstacles include cross-domain formal languages, reliable theorem discovery, computational expense, and guarantees that generated hypotheses remain physically or mathematically valid.
Abstract
Geometry problem solving poses distinct challenges in artificial intelligence. Existing approaches typically fall into two paradigms: symbolic methods, which exhibit limited adaptability, and neural methods, which are prone to hallucinations. Recent neuro-symbolic hybrids predominantly rely on a unidirectional pipeline where neural outputs are fed into solvers without feedback, making system brittle to early-stage errors. To break this unidirectional bottleneck, we propose BiNSGPS, a framework that establishes Bidirectional Neuro-Symbolic Interaction (BiNS) between a MLLM Adviser and a Symbolic Solver. MLLM Adviser actively incorporates feedback from the symbolic solver to dynamically rectify inconsistent formal representations or propose auxiliary hypotheses, resolving symbolic conflicts and facilitating complex deductions.