Research map / Vision-language and multimodal / Multimodal LLMs

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

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AutoVSR, a VLM-based pipeline that converts circuit schematics into a verified executable intermediate representation (netlist or signal-flow graph) and solves symbolic expressions with a tool-augmented plan-and-execute agent, substantially outperforming end-to-end VLMs and CircuitSense on CircuitSense-derived benchmarks.

Problem

Generating correct symbolic expressions (transfer functions, transient responses) directly from circuit schematic images is error-prone due to visual-to-symbolic construction errors and multi-step symbolic derivation errors.

Approach

Decouple visual perception from symbolic reasoning: VLM builds an Executable IR (netlist or SFG JSON) guided by task routing, component-rule retrieval, and multi-stage verification feedback; a planning agent then calls deterministic symbolic tools (Lcapy, Mason's rule) for derivation.

Key results

• On transfer function generation, AutoVSR with Gemini-3-Flash reaches 82.85% overall vs 23.40% for the end-to-end baseline (+59.45 points).
• On transient response generation, AutoVSR with Gemini-3-Flash reaches 85.84% vs 2.85% end-to-end.
• AutoVSR with lightweight Gemini-3-Flash achieves 89.3% accuracy at 4,563 tokens vs 52.4% for Gemini-3-Pro at 13,412 tokens on a 100-case subset.

Surprising findings

• Removing the symbolic tools library drops GLM-4.6V-Flash accuracy from 49.85% to 4.21%, indicating tools are the dominant contributor.
• Lightweight open-weight GLM-4.6V-Flash with AutoVSR outperforms much larger end-to-end frontier models such as GPT-5 and Gemini-3-Pro.

Limitations

• Benchmark is filtered and annotations corrected by the authors, so comparability with original CircuitSense is limited.
• Relies on API-based closed models; no training, so gains depend on backbone and fixed pipeline design.

Experimental setup

Datasets: CircuitSense

Benchmarks: CircuitSense transfer function CircuitSense transient response

Metrics: accuracy symbolic equivalence accuracy token usage runtime

Baselines: Gemini-3-Flash GPT-5-mini Qwen3-VL-Plus Claude-Haiku-4.5 Qwen3-VL-32B-Instruct Llama-4-17B-128E GLM-4.6V-Flash CircuitSense Gemini-3-Pro GPT-5 Claude-Sonnet-4.5

Builds on: Chain-of-Thought Mason's gain formula Lcapy SymPy Mixture of Experts

Closest papers by meaning

• CircuitSense: A Hierarchical Circuit System Benchmark Bridging Visual Comprehension and Symbolic Reasoning in Engineering Design Process (ICLR 2026 · similarity 0.89)
• CircuitFusion: Multimodal Circuit Representation Learning for Agile Chip Design (ICLR 2025 · similarity 0.85)
• Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment (ICLR 2025 · similarity 0.84)
• BetterV: Controlled Verilog Generation with Discriminative Guidance (ICML 2024 · similarity 0.83)
• SymRTLO: Enhancing RTL Code Optimization with LLMs and Neuron-Inspired Symbolic Reasoning (NeurIPS 2025 · similarity 0.83)
• LaMAGIC: Language-Model-based Topology Generation for Analog Integrated Circuits (ICML 2024 · similarity 0.83)