Hidden in Plain Sight: Visual-to-Symbolic Analytical Solution Inference from Field Visualizations

Abstract: Recovering analytical solutions of physical fields from visual observations is a fundamental yet underexplored capability for AI-assisted scientific reasoning. We study visual-to-symbolic analytical solution inference (ViSA) for two-dimensional linear steady-state fields: given field visualizations (and first-order derivatives) plus minimal auxiliary metadata, the model must output a single executable SymPy expression with fully instantiated numeric constants. We introduce ViSA-R2 and align it with a self-verifying, solution-centric chain-of-thought pipeline that follows a physicist-like pathway: structural pattern recognition solution-family (ansatz) hypothesis parameter derivation consistency verification. We also release ViSA-Bench, a VLM-ready synthetic benchmark covering 30 linear steady-state scenarios with verifiable analytical/symbolic annotations, and evaluate predictions by numerical accuracy, expression-structure similarity, and character-level accuracy. Using an 8B open-weight Qwen3-VL backbone, ViSA-R2 outperforms strong open-source baselines and the evaluated closed-source frontier VLMs under a standardized protocol.

• The paper introduces the ViSA task to recover executable symbolic expressions from 2D steady-state field visualizations and their derivatives.
• It leverages high-bandwidth visual cues and a structured chain-of-thought rationale in a fine-tuned VLM to improve symbolic structure recovery.
• Empirical results on VISA-Bench demonstrate enhanced structure similarity and reduced numerical error compared to traditional methods.

Visual-to-Symbolic Analytical Solution Inference from Field Visualizations: An Expert Review

Problem Formulation and Motivation

The paper "Hidden in Plain Sight: Visual-to-Symbolic Analytical Solution Inference from Field Visualizations" (2604.08863) introduces and precisely formulates the task of visual-to-symbolic analytical solution inference (ViSA) for two-dimensional linear steady-state fields. The objective is to recover executable symbolic expressions, including all numeric parameters, directly from field visualizations and their first-order derivatives, relying on minimal auxiliary metadata. This setting targets the core of scientific reasoning: abstracting the underlying closed-form governing law from experimental or simulated field data presented visually.

Unlike conventional symbolic regression (SR) and program synthesis approaches—where models operate on sampled data points or tabular representations—the ViSA task is inherently multimodal and requires extracting rich geometric and topological signatures salient in images, such as symmetry, singularity structure, oscillation, and decay patterns. The core hypothesis is that these patterns represent a high-bandwidth channel for solution inference, often inaccessible to models operating on numerical data alone. Accordingly, the ViSA challenge constitutes a rigorous testbed for assessing the induction and abstraction capabilities of vision-LLMs (VLMs) in the context of scientific discovery.

Dataset Construction and Benchmark Design

A substantial contribution of this work is the construction of VISA-Bench, a synthetic but physically plausible dataset consisting of 30 parametric scenarios covering key classes of linear steady-state partial differential equations (PDEs) and associated boundary value problems. Each scenario is parameterized to yield 500 diverse field realizations, with the total corpus exceeding 15,000 instances, of which 1,500 are curated for reasoning-aligned training and 150 held out for evaluation.

Field scenarios include archetypes such as point-source potentials, radial/axially-symmetric oscillatory modes, polynomial and exponential families, and special function solutions (e.g., Bessel, Airy). For each instance, the dataset comprises three core inputs:

• A scalar field image (heatmap of $u(x, y)$)
• A composite gradient image (half-panel $\partial u/\partial x$, half-panel $\partial u/\partial y$)
• Lightweight numerical metadata (domain boundaries, normalization, basic statistics)

Each instance is annotated with the ground-truth SymPy expression and a chain-of-thought (CoT) style rationale, synthesizing a stepwise expert reasoning trajectory from visual observations to symbolic closure.

Evaluation leverages a tripartite metric system: character-level accuracy (edit distance on SymPy string), structure similarity (using sympified skeletons with numeric placeholders), and numerical accuracy (bounded relative $L_2$ error between predicted and ground-truth field values under the predicted expression). This holistic metric design enables fine-grained attribution of model failure modes: structural, syntactic, or coefficient-level errors.

ViSA-R2: Model Architecture and Reasoning Alignment

ViSA-R2 is built by fine-tuning an 8B parameter Qwen3-VL VLM, leveraging its multi-image conditioning and autoregressive text generation capabilities. The model is supervised not only on the final symbolic output but also on an intermediate chain-of-thought rationale, programmatically synthesized for each training instance via a staged pipeline:

1. Visual feature extraction (observation of symmetry, extrema, oscillation, etc.)
2. Numerical evidence verification (quantitative corroboration of observed cues)
3. Ground-truth feature mapping (mapping visual/numeric cues to theoretical templates)
4. Feature matching, evidence traceability, and parameter estimation via multiple independent modes (colorbar reading, extremum localization, decay/growth rates, etc.)
5. Assembly of a natural-language rationale, culminating in the symbolic solution

This alignment profoundly shapes model behavior, enforcing explicit, stepwise causal inference rather than direct pattern-indexing or overfitting to superficially correlative visual cues. During inference, ViSA-R2 produces both rationale and symbolic expression, with post-processing to ensure parsability and adherence to SymPy syntax.

Empirical Results and Analysis

ViSA-R2 yields state-of-the-art performance on VISA-Bench, outperforming both established open-source LLM/VLM baselines (e.g., Claude-Haiku-4-5, GPT-5.2, Grok-4-1) and several closed-source commercial VLMs under a standardized test protocol.

Key findings include:

• Structure recovery is substantively improved by visual conditioning: VLMs leveraging field/gradient images demonstrate markedly higher structure similarity compared to LLMs exposed only to tabular numerical samples. For GPT-5.2, structure similarity jumps from 0.323 (LLM-only) to 0.768 (VLM).
• Fine-tuning with gold CoT rationales is essential: Without CoT-aligned supervision, the model degenerates into repetitive, incoherent, or structurally nonsensical reasoning, often failing to output valid symbolic expressions.
• Programmatic parameter refinement further reduces numerical error: While structure recovery is the bottleneck, post-hoc coefficient fitting (e.g., via L-BFGS-B optimization) on the model's predicted symbolic scaffold yields considerable gains in numerical accuracy, supporting a tandem approach: structure via vision and reasoning, coefficients via targeted numeric minimization.

The ablations underline that mere prompting or test-time CoT injection in absence of aligned training produces weak results. The evidence supports the claim that semantically-structured visual patterns constitute an effective inductive bias for analytical solution inference.

Implications and Future Directions

This work establishes a new, challenging paradigm for scientific machine learning: end-to-end analytical solution inference from visual field observations. In contrast to modalities limited to description, extraction, or basic regression, the ViSA-R2 approach demonstrates the feasibility of symbolic abstraction—recovering concise, parametric laws that are amenable to mathematical analysis and verification.

Practical Implications:

• Accelerated discovery in physical sciences, automating a workflow analogous to a theoretical physicist's: observe data → hypothesize solution template → fit parameters → verify/iterate.
• New capabilities for scientific automation systems to process field data (simulations, experiments) and propose candidate governing laws.
• Augmented analysis for domains such as microscopy, remote sensing, and engineering diagnostics, where visualizations are the primary evidence.

Theoretical Implications:

• Provides evidence for the utility of vision-language abstraction in capturing geometrically/physically salient solution features inaccessible to sequence-based models.
• Validates the chain-of-thought alignment regime for bridging multi-step reasoning with symbolic output spaces.
• Suggests that future progress in machine scientific reasoning will hinge on integrating high-bandwidth visual inductive signals with explicit multi-stage symbolic reasoning.

Future Directions:

• Extension to nonlinear, time-dependent, and partial-observability regimes, introducing more realistic and challenging scientific conditions.
• Enlarging operator families and source conditions, including mixed-type PDEs and complex boundaries.
• Scalable, automated synthesis of high-quality reasoning traces, potentially via self-distillation, reward modeling, or curriculum construction.

Conclusion

The ViSA-R2 framework and VISA-Bench benchmark represent a substantive advance toward vision-grounded symbolic reasoning in AI. The results substantiate that, with structured reasoning alignment and careful task design, VLMs can directly infer parametric closed-form physics solutions from rich field visualizations. This capability unlocks new frontiers in automated scientific reasoning, bridging the gap between human-expert inductive insight and data-driven machine inference (2604.08863).