Agentic Symbolic Search: Approximating Solutions of PDEs Beyond Hand-crafted Expressions, Meshes, and Neural Networks

arXiv:2606.20467 · cs.LG, cs.NA, math.NA, physics.comp-ph · Submitted 2026-06-18 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Agentic Symbolic Search".

Jane: Here's the gist of the paper. Agentic Symbolic Search (ASYS) proposes a prior-guided framework that translates PDE theory and constraints into testable differentiable symbolic programs to characterize PDE solutions beyond traditional methods.

Tom: Core Methodology The method operates through two nested loops with strictly decoupled optimization objectives, where the outer loop is scientifically a hypothesis generator over an ansatz space and the inner loop fits continuous parameters by quasi-Newton optimization.

Jane: Each representation is a differentiable program composed of coordinate transformations, analytic building blocks, piecewise branches, and trainable parameters theta written so that automatic differentiation can compute ∂u theta /∂theta for gradient-based fitting.

Tom: The search space is discrete at the level of program structure and continuous at the level of parameter values.

Paper summary: Tom: So to wrap up the paper Agentic Symbolic Search: Approximating Solutions of PDEs Beyond Hand-crafted Expressions, Meshes, and Neural Networks, we see a framework where an agent builds symbolic programs from theory and constraints, which are then refined by evolutionary search.

Jane: The authors claim that this approach moves beyond just generating numerical data or opaque neural network weights to produce explicit mathematical structures like scaling laws or interface geometries that mathematicians can actually inspect.

Lu: It’s about automating the structural search process, making it an automated form of inductive bias injection rather than just blind symbolic regression, and they showed success in recovering known analytical forms naturally.

Meng: The practical takeaway for deployment right now is that this is a powerful tool for deep understanding on difficult single problems, but it’s not yet a ready-made solver for massive families of initial conditions or the most complex cases without more optimization flexibility.

Lalam: The paper shows we can get nine-parameter candidates with relatively low error in some blow-up problems, and they clearly identified that the next step is incorporating observed data as a third scoring signal to extend the method toward inverse problems.

Conclusion: Tom: So, we've been looking at this paper called Agentic Symbolic Search, by some authors who seem pretty deep into physics and machine learning.

Jane: Yeah, it’s about this framework where an agent builds mathematical programs from the rules of a partial differential equation and its constraints. It tries to find solutions that aren't just numbers but actual structures you can see.

Lu: Exactly. The core idea is using a search process—a nested loop—to generate these symbolic programs, which are then optimized by continuous parameters to fit the physics rules directly, not just some black box network weights.

Meng: From my side, it’s interesting that they focus so much on this structural characterization for individual difficult problems instead of trying to build something that solves millions of different initial conditions all at once.

Lalam: What I see is a method that learns how to represent the solution in terms of building blocks like coordinate transformations and analytic functions, which is a big step away from just fitting arbitrary points.

Tom: So it’s not about finding *a* solution, it’s about finding an interpretable *way* to write down the solution. That makes sense for understanding why something behaves the way it does.

Jane: Right. It's moving toward solutions that are actually understandable by a human who knows math, rather than just a computer outputting a number or some set of weights.

Lu: The results they show, like getting those nine-parameter contraction laws for problems like Keller-Segel blow-up, prove that this structural approach can find these hidden mathematical relationships when they exist.

Meng: But the paper does flag that the search process itself is still a bit fragile; if the agent doesn't propose a good starting structure, it just wanders into non-convex optimization trouble.

Lalam: And they admit that L-BFGS, their inner optimizer, struggles with those tricky singular spots because the landscape gets really weird there.

Tom: So, it’s powerful for finding these specific instances of solutions that look mathematically clean and structured. But can this process scale up to solve a whole class of problems reliably?

Jane: That’s the big question. They show promise on single hard cases, but we still need to see if this structural search becomes a reliable way to tackle massive families of initial conditions.

National University of Singapore

cs.LG, cs.NA, math.NA, physics.comp-ph

Submitted: 2026-06-18

Updated: 2026-11-10

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 86/100

The gist: Agentic Symbolic Search (ASYS) proposes a prior-guided framework that translates PDE theory and constraints into testable differentiable symbolic programs to characterize PDE solutions beyond…

Key concepts

Core Methodology
The method uses two nested loops: an outer loop generates hypotheses about the solution structure (ansatz space), and an inner loop fine-tunes continuous parameters using optimization. Each representation is a differentiable program allowing automatic differentiation for gradient-based fitting.
Four-Dimensional Score Vector
Evaluation relies on four scores: physics residual, initial value match, boundary consistency, and kinematic consistency. This score vector guides the search by prioritizing representations that satisfy multiple physical constraints simultaneously.
Structural Characterization
ASYS aims to produce interpretable representations of PDE solutions. Instead of just a numerical answer, it generates formulas—like geometric interfaces or contraction laws—that reveal the underlying mathematical dynamics, such as blow-up behavior.
Evolutionary Ensemble of Agents (EvE)
The outer search mechanism uses EvE to maintain a population of high-scoring candidates and a strategy population. Coding agents generate new programs based on guidance from previous successful structures and diagnostic summaries.

Terminology

Summary

Agentic Symbolic Search (ASYS) proposes a prior-guided framework that translates PDE theory and constraints into testable differentiable symbolic programs to characterize PDE solutions beyond traditional methods. p. 0

Core Methodology

The method operates through two nested loops with strictly decoupled optimization objectives, where the outer loop is scientifically a hypothesis generator over an ansatz space and the inner loop fits continuous parameters by quasi-Newton optimization. Each representation is a differentiable program composed of coordinate transformations, analytic building blocks, piecewise branches, and trainable parameters theta written so that automatic differentiation can compute ∂u theta /∂theta for gradient-based fitting. p. 2 The search space is discrete at the level of program structure and continuous at the level of parameter values. p. 2

Scoring and Evaluation

Evaluation in ASYS is driven entirely by the governing equation and its publicly specified constraints, strictly excluding any observed solution data. The evaluator constructs a four-dimensional score vector s(u theta) = (sphys, sic, sbc, scomp). p. 3 This vector includes the physics residual (sphys), initial value match (sic), boundary consistency (sbc), and a crucial compatibility condition score (scomp) which evaluates kinematic consistency at the initial time. p. 4

Iterative Refinement

ASYS runs inside Evolutionary Ensemble of Agents (EvE) as an outer-loop search mechanism, maintaining a population of scored candidates and a strategy population that guides the search. New programs are generated by coding agents based on problem guidance, previous high-scoring representations, and diagnostic summaries. p. 3 Selection is gradual through rank-exponential-sum weights across the four dimensions to bias the search toward representations that perform well across multiple criteria. p. 3

Structural Characterization of Solutions

ASYS demonstrates the possibility of a new paradigm for characterizing PDE solutions by producing interpretable representations, including a geometric interface formula for Allen–Cahn 2D dynamics and a nine-parameter contraction law for Keller–Segel chemotactic blow-up. For instance, in the Keller–Segel case, ASYS constructs an explicit contraction law blending an early Taylor branch with a late blow-up branch composed of a critical core and two Gaussian halos using only nine interpretable trainable scalars. p. 9

Limitations and Future Directions

Three limitations shape the current results: first, ASYS is a structural characterizer for difficult individual evolutions, not an amortized solver for families of initial conditions. Second, the quality of the search depends on the coding agent’s ability to propose useful ansatz families within a limited number of iterations. p. 1 Third, the inner-loop optimizer (L-BFGS) operates on a non-convex landscape whose conditioning deteriorates near singularities, suggesting that more flexible optimization strategies might be needed. p. 2 Extensions include incorporating observed data as a third scoring signal to extend the method to inverse problems and connecting agent-generated representations to formal verification. p. 5

Results Across Problem Types

The experiments show ASYS recovers known analytical forms when they exist, such as the exact breather core for the NLS equation. For Allen–Cahn 2D, it builds a geometric signed-distance formula representing peanut-to-oval dynamics with 23 trainable scalars. p. 7 In Keller–Segel blow-up, ASYS recovers a nine-parameter candidate with a relative L2 error of 0. p. 0188, whereas the specialized SS-PINN baseline achieves an error of 0. p. 5258. For Graveleau PME focusing, the search instantiates a second-kind self-similar free-boundary framework with a learned exponent βmodel = 0. p. 0928, matching the reference fit βref = 0. p. 9877. The gCLM stress test showed that ASYS failed to construct the global self-similar reduction within ten iterations, highlighting the search's structural reach.

Validation Metrics

Final solution accuracy is evaluated separately after the run against independent numerical references, and this offline validation error determines entries in Table 2. The four score dimensions are defined as s(u theta) = (sphys, sic, sbc, scomp). p. 3 The physics residual for singular cases uses a scale-free geometric-mean relative residual to measure equation satisfaction. p. 4

Computational Cost

All experiments were run on a single Apple M4 laptop CPU with no GPU acceleration. The dominant cost is the outer-loop structural search, not large-scale numerical training. p. 5 Each representation is fitted by L-BFGS with a wall-clock budget of 30 seconds for bounded cases and 300 seconds for blow-up and free-boundary cases. p. 5 The agent sessions were executed through a subscription-covered interactive coding-agent interface. p. 5

Candidate Formulas

The results section presents the best candidate in compact form, decomposing each expression into its functional roles like initial anchor, structural branch, transition gate, boundary factor, and residual correction. For the Keller–Segel case, the nine-parameter candidate is u(r, t) = (1 − alpha(t))uT(r, t) + alpha(t)uB(r, t), where uB incorporates a learned power law for the contraction scale. p. 9 The Allen–Cahn candidate uses a signed-distance formula involving terms like Sunion and Sneck to capture geometric dynamics. p. 7 The Graveleau candidate combines an anchor, a self-similar branch, and an MLP correction to fit the focusing free-boundary shape.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Agentic Symbolic Search (ASYS) can recover known analytical forms for PDEs naturally when they exist, such as in NLS problems, by proposing mathematical structures that are directly read from literature rather than relying on blind search.

  2. ASYS constructs new analytical approximations for problems without closed-form solutions, guiding mathematicians toward further analysis by producing interpretable representations like a geometric interface formula for Allen–Cahn 2D dynamics.

  3. The framework enables the automation of inductive bias injection by allowing a coding agent to translate prior knowledge from PDE theory into testable differentiable symbolic programs, moving beyond blind symbolic regression.

  4. ASYS can produce explicit mathematical structures, which is described as the only paradigm that produces it through an automated search, offering interpretable mechanisms like scaling laws or contraction laws where no closed-form description was previously available.

  5. The system uses a four-dimensional score vector (physics residual, initial value, boundary consistency, and compatibility condition) to rigorously assess candidates based strictly on the governing equation and public constraints, excluding observed solution data from the search process.

  6. The framework allows for decoupling hypothesis generation (outer loop over ansatz space) from parameter fitting (inner loop over continuous parameters), ensuring that the outer loop controls mathematical structure while the inner loop optimizes continuous values.

  7. ASYS can generate high-dimensional symbolic programs, such as a nine-parameter contraction law for Keller–Segel blow-up, which compresses complex dynamics into interpretable trainable scalars.

  8. The system incorporates an Evolutionary Ensemble of Agents (EvE) to maintain a population of scored candidates and search strategies, where selection is biased toward representations that perform well across the four score dimensions without committing to a single best program or fixed scalar objective.

  9. ASYS can identify the limits of current search capabilities, such as in the gCLM stress test, by showing when it fails to construct a global self-similar reduction within a limited number of iterations, directing future research toward hybrid representations.

  10. The framework supports extensions for inverse problems by incorporating observed data as a third scoring signal without redesigning the core evaluator structure, allowing recovery of unknown terms or parameters from measurements.

Sources

Related papers