AutoVSR: Automatic Visual-to-Symbolic Reasoning for Symbolic Expression Generation from Circuit Schematic

Deriving symbolic expressions (such as transfer functions) from circuit schematics is a labor-intensive task that requires both accurate visual interpretation of circuit diagrams and precise algebraic derivation. Existing end-to-end Vision Language Models (VLMs) often struggle with the multi-step reasoning required for these tasks, leading to error accumulation. This paper asks: can we improve the accuracy and reliability of visual-to-symbolic reasoning by decoupling visual perception from symbolic computation?

AutoVSR introduces a two-stage framework that separates the interpretation of circuit schematics from the mathematical derivation of circuit behavior.

1. Visual-to-Textual Topology Generation: Instead of asking a VLM to generate symbolic expressions directly, AutoVSR forces the model to generate an 'Executable Intermediate Representation' (IR). This process is guided by dynamic context prompting and a verification-based feedback loop that checks for syntactic, structural, and physical consistency (e.g., ensuring the circuit has a ground node and is solvable).
2. IR-Driven Symbolic Reasoning: Once a validated IR is produced, the system uses a planning agent equipped with a symbolic tool library (such as Lcapy or SymPy). The VLM acts as a high-level planner that decomposes the analysis goal into discrete steps, while the symbolic tools handle the actual algebraic manipulation, ensuring that the final expressions are mathematically sound.

AutoVSR significantly outperforms existing end-to-end VLM approaches and specialized baseline methods on symbolic expression generation tasks, achieving accuracy improvements ranging from 30% to nearly 60%. By delegating the heavy lifting of algebraic derivation to deterministic symbolic solvers, the framework avoids the common pitfalls of hallucinated math often found in pure VLM-based reasoning. Furthermore, AutoVSR demonstrates superior computational efficiency and lower inference costs compared to state-of-the-art closed-source models.