From Symbolic Perception to Logical Deduction: A Fr... | AI Research

From Symbolic Perception to Logical Deduction: A Framework for Guiding Language Models in Geometric Reasoning
This research addresses the challenge of solving complex plane geometry problems, which require both visual understanding and logical reasoning. While Large Multimodal Models (LMMs) can process images and text, they are often computationally expensive and prone to opaque, "black-box" reasoning. This paper introduces a framework that enables a standard Large Language Model (LLM) to solve geometry problems by pairing it with specialized modules that handle visual parsing and formal logical deduction, resulting in performance that rivals state-of-the-art models while providing clearer, human-like explanations. To see google in practice, PowerPoint is DEAD walks through a concrete example.

Bridging Vision and Logic

The framework functions through a three-stage pipeline. First, a "Geometric Vision Parser" converts raw diagram images into structured, symbolic data. It uses detection models to identify geometric primitives—such as points, lines, and circles—and assigns labels and symbols to them. Second, a "Symbolic Solver" acts as a logical engine. It takes the parsed information and applies geometric theorems to derive new facts. By using strict rules to filter these deductions, the system avoids the "combinatorial explosion" that occurs when a computer tries to test every possible geometric relationship, focusing only on those relevant to the problem.

Guiding the Language Model

Once the visual data is parsed and the logical facts are verified, the system feeds this information into an LLM. By providing the model with a structured set of verified geometric truths alongside the original problem text, the framework "grounds" the LLM. This prevents the model from relying on unreliable visual interpretations or "brute-force" coordinate calculations. Instead, the LLM is guided along a logically sound path, which significantly reduces hallucinations and produces a step-by-step solution that is easier for humans to read and verify. The google story also surfaces in Google AI Introduces EnvHarness for Adaptive..., adding another angle.

A New Standard for Evaluation

To test the effectiveness of this approach, the researchers identified that existing geometry benchmarks were becoming outdated and potentially contaminated by data present in the models' original training sets. To solve this, they curated a new, high-difficulty benchmark based on the 2025 Chinese Zhongkao (middle school) examinations. These problems are specifically designed to test deep deductive skills and multi-step reasoning. Because these questions are from recent, public examinations, they provide a reliable, "un-memorized" testbed to evaluate how well AI models truly reason through novel geometric challenges.

Performance and Interpretability

Experiments show that this modular approach allows a standard LLM to achieve performance comparable to advanced models like Gemini 2.5 Pro. Beyond just getting the right answer, the framework excels in transparency. Because the reasoning is grounded in formal geometric theorems rather than just statistical pattern matching, the output is more interpretable. This shift away from "brute-force" math toward deductive, theorem-based reasoning marks a significant step toward making AI-driven mathematical problem-solving more reliable and user-friendly. The google story also surfaces in OpenAI’s Opaque Reasoning Technique Raises Alarm..., adding another angle. as detailed in the full paper on Arxiv