acceptodds
Under review as a conference paper at ICLR 2027

Symmetry- and Semantic-Aware Symbolic Discovery of Ordinary Differential Equations

Abstract

Recovering symbolic ordinary differential equations (ODEs) from trajectories requires solutions that are both expressive and consistent with known physics. Library-based sparse regression methods provide structural guarantees only within predefined invariant/equivariant bases, while free-form symbolic regression explores many symmetry-violating expressions and obtains behavioral feedback mainly from completed candidates. From this perspective, we develop Symmetry- and Semantic-Aware Graph Symbolic Regression (SGSR) by searching over canonical, symmetry-typed expression graphs equipped with sub-expression-level semantics. The local typing rules of expression graphs guarantee the prescribed symmetry of the discovered ODEs by construction, and path-conditioned semantic features compare closed sub-expressions with position-specific behavioral targets. This framework makes invariant-basis methods a special case of SGSR while supporting *free-form, non-polynomial* dynamics, and further integrates path-conditioned semantic guidance into a neural-guided Monte Carlo Tree Search pipeline. Experiments on polynomial and non-polynomial symmetry-constrained systems and on ODEBench demonstrate improved symbolic recovery and trajectory reconstruction of SGSR while maintaining physical consistency.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
symbolic regression, ordinary differential equations, symmetry typing, expression graph, semantic feature

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_n9HHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed a neural-guided symbolic search method. The method uses symmetry-typed expression graphs to enforce prescribed ODE symmetries and sub-expression semantics to guide search. The author claims that their design supports free-form, non-polynomial dynamics, and can be further integrates into a neural-guided Monte Carlo Tree Search.

Critical strengths
  1. The problem is clear and well motivated, the gap between fixed invariant libraries and unconstrained symbolic search is interesting.
  2. The design is direct: aonstraints are built into representation.
  3. The reported results are comprehensive.
Critical weaknesses
  1. In my opinion, the method is a combination of symmetry-typed expression-graph construction with path-conditioned functional targets for sub-expressions, integrating both into neural-guided MCTS for ODE discovery. The individual ingredients have clear precedents.
  2. Equation (8) defines a child target by applying the inverse of its parent operator. But the mask in Equation (8) is set using only whether the parent target and sibling semantics are available. The target-availability mask seems not check whether an operator can actually be inverted.
  3. The ODEBench comparison does not isolate the semantic components: I suppose that section 4.4 attributes S2GSR’s gains over GSR to sub-expression semantics and path-conditioned guidance. But that is a comparison of the full methods.
  4. The reported recovery rates lack uncertainty and an operational definition.
  5. Given the Figure 2, “significantly outperforms library-based SINDy methods” is too strong and not sufficiently verified.
Questions for authors
  1. What are the core innovations of S2GSR? In what specific ways does the combination of "symmetry-aware graphs," "sub-expression guidance," and "Monte Carlo Tree Search" (MCTS) surpass existing methods?
  2. In Equation (8), how is it ensured that the parent operator is invertible, thereby allowing for the derivation of a child target?
  3. Could ablation studies be added for the ODEBench experiments to clarify which specific components of S2GSR contribute to these performance improvements?
  4. How is it determined whether a method has successfully recovered the complete ground-truth equation? Are algebraically equivalent expressions considered correct?
  5. Figure 2 indicates that S2GSR performs "significantly better" than library-based methods like SINDy. What statistical evidence supports this conclusion? Does "significantly" here imply statistical significance, or does it simply mean that S2GSR achieved higher scores in the experiments?
5 comments
AuthorsOpenReview ↗

Responses to Reviewer n9HH [1/2]

We thank the reviewer for the careful review and for recognizing the motivation of bridging fixed invariant libraries and unconstrained symbolic search, the direct design that builds constraints into the representation, and the comprehensiveness of our results. We also value the Weaknesses (W) and Questions (Q) raised in the review. Below, we respond to each concern in order; all revisions are marked in blue in the updated PDF.


R1. Core innovations of S$^2$GSR

[W1, Q1] “The individual ingredients have clear precedents.” and “What are the core innovations of S2GSR? In what specific ways does the combination [...] surpass existing methods?”

We thank the reviewer for this question. S$^2$GSR differs from each of these existing methods in important ways and unifies them to address two challenges that matter in ODE discovery: enforcing physical symmetries exactly in a free-form search space, and guiding the search toward hard, non-polynomial dynamics. Specifically:

  • Symmetry-typed expression graphs vs. expression graphs in GSR. GSR (Xiang et al., 2025) represents expressions as graphs that encode algebraic equivalences, but imposes no symmetry constraints for dynamical systems. S$^2$GSR attaches a signature (group representation, scaling degree, time-reversal parity) to each node and admits an operator only if it respects the group action (Eq. (5)); Theorem 3.2 guarantees that the discovered vector field satisfies the prescribed symmetry.
  • Sub-expression targets as evidence vs. semantic backpropagation. Semantic backpropagation uses the inverted target as a specification for replacing sub-expressions. In ODE discovery, the target is a noisy derivative estimate, so S$^2$GSR uses it as evidence instead: its alignment with each closed sub-expression is an input to the learned policy and value networks (see General Response G1).
  • Joint guidance vs. hnMCTS in GSR. S$^2$GSR keeps the hnMCTS backbone of GSR, but adds symmetry-aware action masking and sub-expression semantics to its structural embeddings, so the policy and value networks reason over all three jointly.

Together, these improve recovery beyond both library-based and free-form methods: on the non-polynomial systems, S$^2$GSR reaches RR 0.36 and 0.58, compared with at most 0.12 for all the tested baselines, including the symmetry-aware EquivSINDy-c (0.00) and GSR (0.05 and 0.12) (Table 3). Table 4 shows that each component contributes: on NP-SO(2), symmetry typing raises RR from 0.06 to 0.22, and semantic guidance further raises it to 0.36.


R2. Invertibility in Eq. (8)

[W2, Q2] “The target-availability mask seems not check whether an operator can actually be inverted.” and “how is it ensured that the parent operator is invertible [...]?”

We thank the reviewer for the interest in this point. Invertibility is checked in our implementation, which has been discussed in Appendix C.2, where the manuscript line 257 (line 289 in the revised PDF) referred to it. There, a two-tier dispatch computes the target in closed form only for operators that are invertible at the child node's slot (additive slots, multiplication and division away from zero, and strictly monotone unary operators such as exp⁡\exp, log⁡\log, and odd powers); for all others (e.g., even powers, ∥⋅∥2\|\cdot\|^2, sin⁡\sin, and cos⁡\cos), the target is masked for the whole subtree below the operator, which then uses the signature-indexed embedding.

AuthorsOpenReview ↗

Responses to Reviewer n9HH [2/2]

R3. ODEBench ablation

[W3, Q3] “The ODEBench comparison does not isolate the semantic components” and “Could ablation studies be added for the ODEBench experiments [...]?”

We agree that the comparison with GSR is a comparison of full semantic features. We added an ODEBench ablation with an identical backbone, which isolates the semantic components. Under trivial symmetry, we compare structural guidance only (None), sub-expression semantics (Sem-only), and the full model (Sem+Align), keeping the graph transformer, canonicalization, constant optimization, and search budget fixed, on all 63 systems with noise σ=0.05\sigma = 0.05 (percentage of systems with R2>0.9R^2 > 0.9):

Variantρ=0\rho = 0ρ=0.5\rho = 0.5
None[50.79%][55.56%]
Sem-only[55.56%][63.49%]
Sem+Align (full)[58.73%][65.08%]

Each semantic component raises the success rate in both settings (None → Sem-only → Sem+Align: 50.79% → 55.56% → 58.73% at ρ=0\rho = 0, and 55.56% → 63.49% → 65.08% at ρ=0.5\rho = 0.5), while the structure-only variant is on par with GSR (52.38% and 55.56%), so the gains over GSR come from the semantic components rather than the backbone.

Revision: we added this ablation table to Appendix E.5 (Table 10), and Sec. 4.4 now attributes the gains over GSR only as supported by it.


R4. Definition and uncertainty of the recovery rate

[W4, Q4] “The reported recovery rates lack uncertainty and an operational definition.” and “Are algebraically equivalent expressions considered correct?”

We thank the reviewer for this comment. Both were given in Appendix E.1 and Table 9. Following the EquivSINDy evaluation (Yang et al., 2024), a seed counts as recovered only if, after algebraic simplification, the inferred and ground-truth fields contain exactly the same symbolic terms in every component, regardless of coefficient values; algebraically equivalent expressions are therefore counted as correct. Table 9 reports Wilson 95% intervals over the 100 seeds for the recovery rate (RR).


R5. Use of “significantly” in Sec. 4.4

[W5, Q5] “‘significantly outperforms library-based SINDy methods’ is too strong” and “Does ‘significantly’ here imply statistical significance [...]?”

We thank the reviewer for this comment. We meant a large margin (18-62 percentage points higher) rather than statistical significance, and we have removed the word.

We thank the reviewer again for the time and constructive comments, which helped us improve the precision of our definitions and claims. We hope our responses address your concerns, and we look forward to your feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up on Responses to Reviewer n9HH

Dear Reviewer n9HH,

Thank you for your careful review. As the discussion period closes today, we wanted to follow up on our response. In brief, we:

  1. clarified the core innovations of S$^2$GSR relative to GSR, semantic backpropagation, and hnMCTS;
  2. pointed to the invertibility check of Eq. (8), described in Appendix C.2;
  3. added an ODEBench ablation with an identical backbone, which isolates the semantic components (Appendix E.5, Table 10);
  4. pointed to the operational definition of the recovery rate and its Wilson intervals (Appendix E.1, Table 9);
  5. removed "significantly" from Sec. 4.4.

We would be glad to answer any remaining questions. If our response has addressed your concerns, we would appreciate it if you could consider updating your assessment. We would like to express our sincere appreciation again for your time and effort in reviewing our work.

Best regards,

The Authors

Reviewer_Z6EwOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes S2GSRS^2GSR, a framework for discovering symbolic ODEs from observed trajectories under prescribed physical symmetries. It builds on the graph-based neural-guided MCTS of GSR and adds two main components: 1) Symmetry-Typed Expression Graphs. Each node carries a signature made of a group representation, a scaling degree, and a time-reversal parity, and local typing rules mask out symmetry-incompatible expansions during search. Theorem 3.2 shows that, for linear group actions, local type-soundness implies that the final vector field satisfies the prescribed symmetry. 2) Path-conditioned semantic guidance. For closed sub-expressions, the authors invert the parent operators along the path from the root, given the sibling values, to obtain a slot-specific derivative target. A cosine alignment between the sub-expression’s outputs on a probe set and this target is then fed, together with structural and semantic features, into a graph transformer that provides policy and value estimates for MCTS. The authors evaluate the method on four symmetry-constrained systems (polynomial and non-polynomial, under rotation, scaling, and time-reversal symmetries) over 100 seeds. Ablations separate symmetry typing from semantic guidance, an experiment selects among candidate symmetries using validation reconstruction, and the ODEBench experiments use trivial symmetry. On the non-polynomial systems, S²GSR reaches recovery rates of 0.36 and 0.58, while GSR reaches 0.05 and 0.12 and the equivariant SINDy baseline reach zero.

Critical strengths

1. Symmetry enforced by construction in a free-form search space

I appreciate the way the authors turn physical symmetry into a hard constraint on the search space instead of a soft penalty or a post hoc filter. The typing over group representation, scaling degree, and time-reversal parity is clean, and the local-to-global argument in Theorem 3.2 is easy to follow. The guarantee covers linear group actions (Appendix B.5), and the authors are upfront about this scope. What I find most valuable is that this keeps free-form expressivity: the method can recover non-polynomial factors like sin⁡(r)e−r\sin(r)e^{-r} that fixed invariant-basis libraries cannot represent, and the non-polynomial results in Table 3 show this clearly. I think this is a meaningful contribution to physics-constrained symbolic discovery.

2. Clean ablation design that separates the two mechanisms

I also appreciate that the ablation in Table 4 is set up so the contribution of each component can actually be read off. Symmetry typing is toggled by switching to the trivial group, while canonicalization and the search procedure stay unchanged. Semantic information is then added in two steps, from sub-expression semantics to path-conditioned alignment. This makes it clear that typing is responsible for most of the gain on the polynomial system, while semantic guidance contributes comparably on the harder non-polynomial one. The ODEBench results with trivial symmetry further suggest that semantic guidance helps even without symmetry pruning, since S2GSRS^2GSR improves over GSR in all six conditions.

3. Minor: Good practice for diagram display

Figures are well-made, visually clear and considerate to help readers easily understand the framework. Good visualization is often taken for granted in scientific publications, but I believe thoughtful figure design deserves recognition, especially when illustrating technical methods.

Critical weaknesses

1. Unclear novelty of path-conditioned guidance relative to semantic backpropagation

My main concern is the novelty of the path-conditioned semantic guidance, which the paper presents as one of its central contributions. The recursion in Eq. (8) propagates a desired target from the root down to an operand slot by locally inverting parent operators given the sibling values. As far as I can tell, this is closely related to semantic backpropagation in the genetic programming literature, where desired outputs are propagated through a program via operator inversion to guide the construction of subtrees. The related-work section describes semantic GP methods as comparing only complete expressions, which I think may understate this connection.

I acknowledge that using the propagated target as a cosine-alignment feature inside a learned policy/value network for MCTS, combined with symmetry-typed graphs, could still be new. However, the paper does not currently separate what is inherited from what is novel, and this makes it difficult for me to assess the methodological contribution of this component. I consider this the most important issue, because it directly affects how much of the paper’s contribution is new rather than a combination of existing ideas. The authors should justify and clarify their novelty.

2. Limited understanding of when path-conditioned alignment is actually available

The alignment signal requires a valid local inverse at every step along the path from the root. As described in Appendix C.2, operators such as sin⁡\sin, cos⁡\cos, even powers, and absolute value, as well as multiplication or division near zero, set the target mask to zero. Every descendant of such an operator then inherits the mask and falls back to a signature-indexed embedding. One case I can think of is the NP-SO(2) ground truth itself: everything under the sin in sin⁡(r)e−r\sin(r)e^{-r} receives no alignment target. As the authors acknowledge in Appendix E.2, the Sem-only ablation removes the alignment stream entirely, but it does not measure how often alignment is available in the full model.

I understand that Table 4 shows alignment improves recovery. Nevertheless, without statistics on how often this alignment is available, I find it hard to judge how broadly this mechanism applies beyond the tested systems, especially for deeper expressions with nested uninvertible or numerically-unstable operators. I consider this a limitation in understanding the practical scope of the central component.

3. Minor: presentation issue In Figure 1(b), the illustrated states jump from s1s_1 to s4s_4, and both the pre- and post-canonicalization graphs are labeled s_4. It would be helpful to clarify whether intermediate expansion steps are omitted and whether canonicalization is considered part of the same search state.

Questions for authors

I don't have other questions than the concerns I raised above. I'd be willing to raise my score if the authors can adequately address the novelty concern in particular.

5 comments
AuthorsOpenReview ↗

Responses to Reviewer Z6Ew [1/2]

We thank the reviewer for the detailed and positive review, and for recognizing that S$^2$GSR enforces symmetry by We thank the reviewer for the detailed and positive review, and for recognizing that S$^2$GSR enforces symmetry by construction in a free-form search space with a local-to-global guarantee, as well as the clean ablation design and the clarity of our figures. We also value the Weaknesses (W) raised in the review. Below, we respond to each concern in order; all revisions are marked in blue in the updated PDF.


R1. Novelty relative to semantic backpropagation

[W1] “closely related to semantic backpropagation [...] does not currently separate what is inherited from what is novel”

We thank the reviewer for this important point and agree that our related work understated this connection. In short, Eq. (8) inherits the inversion principle of semantic backpropagation, and our novelty lies in how the target is used: S$^2$GSR treats it as evidence rather than a specification. As the reviewer anticipated, the target's alignment with each closed sub-expression is an input to the learned policy and value networks, which learn how much to trust it during MCTS, instead of a specification for replacing the sub-expression (Table 4: RR 0.28 → 0.36 on NP-SO(2)). Please see General Response G1 for the full comparison and the corresponding revisions.


R2. Availability of path-conditioned alignment

[W2] “everything under the sin [...] receives no alignment target” and “without statistics on how often this alignment is available [...], especially for deeper expressions with nested uninvertible [...] operators”

We thank the reviewer for this question. Masking affects only the interior of a non-invertible operator, not the operator itself, and we now report how often targets are available. In the NP-SO(2) field sin⁡(r)e−r Jx\sin(r)e^{-r}\,Jx with r=∥x∥2r = \|x\|^2, the coefficient sin⁡(r)e−r\sin(r)e^{-r} receives a target through the scalar–vector product, and sin⁡(r)\sin(r) as a whole receives one through its parent product with e−re^{-r}. Only the edges inside sin⁡\sin are masked. Even then, candidates for the argument of sin⁡\sin are still evaluated: once the argument is closed, sin⁡(⋅)\sin(\cdot) is aligned against the target on its own edge, which the value network uses to score the resulting state, so MCTS favors arguments that make sin⁡(⋅)\sin(\cdot) fit its slot. Constants inside sin⁡\sin are likewise fitted by maximizing this alignment (Appendix D.4). Since only terminals are shared, the rr inside e−re^{-r} is a separate sub-expression and receives a target through exp⁡−1\exp^{-1} where it is defined, and the shared terminal xx receives a target on its edge from JJ. The sub-expression semantics, which do not require inversion, remain available for every closed node.

Appendix C.3 (Table 7) now reports how often a valid target is available on the ground-truth expressions (share of edges, excluding constants) and during search (share of sampled nodes):

Alignment Availability RatioP-SO(2)P-SENP-SO(2)NP-TR
Ground-truth expressions100% (5/5)100% (6/6)67% (6/9)88% (7/8)
During search (when sampling a node)72%78%36%45%

All edges have a target in the polynomial systems; in the non-polynomial systems, the only edges without a target are those inside sin⁡\sin and ∥⋅∥2\|\cdot\|^2. During search, the target availability is lower, since sampled nodes often still have open siblings, and lowest on the non-polynomial systems. Even so, alignment improves recovery where it is least available: on NP-SO(2) (36%), RR rises from 0.28 to 0.36 (Table 4). Coverage decreases as expressions contain more non-invertible operators; the affected edges then use the signature-indexed embedding, while the structural and semantic streams remain. Extending inversion to multi-valued operators (e.g., even powers and periodic functions) is future work.

Revision: Appendix C.3 reports these statistics, and Sec. 3.3 and the limitations now state the scope of the alignment signal.

AuthorsOpenReview ↗

Responses to Reviewer Z6Ew [2/2]

R3. States in Figure 1(b)

[W3] “the illustrated states jump [...] and both the pre- and post-canonicalization graphs are labeled s4s_4”

We thank the reviewer for catching this. Intermediate states were omitted, and canonicalization is part of the same transition. From s1s_1, the actions φ1,…,φ4\varphi_1, \dots, \varphi_4 pass through s2s_2–s4s_4, which the figure omits. The middle panel is the graph after φ4\varphi_4 before canonicalization, T(s4,φ4)T(s_4, \varphi_4), and the right panel is the next search state s5=Canon(T(s4,φ4))s_5 = \mathrm{Canon}(T(s_4, \varphi_4)). Following Eq. (23), canonicalization is part of the transition, so only canonical graphs are MCTS states.

Revision: we relabeled the panels accordingly and added this explanation to the caption of Figure 1.

We thank the reviewer again for the time and constructive comments, which helped us clarify the novelty and the scope of the path-conditioned guidance. We hope our responses address your concerns, especially regarding novelty, and we look forward to your feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up on Responses to Reviewer Z6Ew

Dear Reviewer Z6Ew,

Thank you for your detailed and positive review. As the discussion period closes today, we wanted to follow up on our response. In brief, we:

  1. separated what S$^2$GSR inherits from semantic backpropagation (the inversion principle) from what is new (using the target as evidence for the learned policy and value networks rather than as a specification), with details in General Response G1;
  2. reported how often path-conditioned targets are available, both on the ground-truth expressions and during search (Appendix C.3, Table 7);
  3. relabeled the states in Figure 1(b) and explained in the caption that canonicalization is part of the transition.

We would be glad to answer any remaining questions, especially on novelty. If our response has addressed your concerns, we would appreciate it if you could consider updating your assessment. We would like to express our sincere appreciation again for your time and effort in reviewing our work.

Best regards,

The Authors

Reviewer_yfeiOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces S²GSR for symbolic ODE discovery. It combines symmetry-based typing with sub-expression semantics and position-specific targets in graph-based neural MCTS. Experiments on four symmetry-constrained systems and ODEBench show gains in structural recovery and trajectory reconstruction.

Critical strengths
  1. The method separates hard symmetry constraints from learned search guidance. The ablations show that both contribute, with additional semantic gains on the non-polynomial rotation task.
  2. The symmetry experiments use 100 seeds and report uncertainty. Improvements over GSR on ODEBench with trivial symmetry also support the value of semantic guidance beyond known physical constraints.
Critical weaknesses
  1. The typing rules appear inconsistent with NP-SO(2): Table 1 gives r=∥x∥2r=\|x\|^2 scaling degree two, but Table 6 requires degree-zero inputs for sine and exponential.
  2. The paper does not explain how scaling constraints are disabled.
  3. Novelty and comparisons need closer treatment of semantic backpropagation and symmetry-invariant symbolic search. Polynomial-library baselines cannot establish an advantage over symmetry-aware methods that also allow non-polynomial expressions.
Questions for authors

How is scaling disabled for NP-SO(2), and how are path-specific targets assigned to shared DAG nodes?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer yfei [1/2]

We thank the reviewer for the careful review and for recognizing that S$^2$GSR separates hard symmetry constraints from learned search guidance, as well as for the positive comments on our symmetry experiments over 100 seeds with reported uncertainty, the ablations, and the ODEBench results. We also appreciate the Weaknesses (W) and Questions (Q) raised in the review. Below, we respond to each concern in order; all revisions are marked in blue in the updated PDF.


R1. Scaling constraints for NP-SO(2)

[W1, W2, Q] “The typing rules appear inconsistent with NP-SO(2) [...]” and “How is scaling disabled for NP-SO(2)?”

We thank the reviewer for pointing this out. Scaling is not part of the specification for NP-SO(2), so its rules are not applied. Each signature component is active only if its symmetry is in Σ\Sigma; otherwise, its conditions are skipped, which leaves Theorem 3.2 unchanged. Thus, sin⁡(∥x∥2)\sin(\|x\|^2) is admissible under ΣSO(2)\Sigma_{\mathrm{SO}(2)}, and the degree-zero rule applies only when scaling is imposed (P-SE).

Revision: we revised Sec. 3.2 and Figure 1(b) to state this rule explicitly, and added the active components of each experiment (table below) to Appendix E.1 (Table 8).

Experimentρ\rhoκ\kappaη\eta
P-SO(2), NP-SO(2)vector representation∗*∗*
P-SEtrivialweights (2,1)(2, 1)∗*
NP-TRtrivial∗*parities (+1,−1)(+1, -1)
Relaxed typing, ODEBenchtrivial∗*∗*

R2. Targets at shared nodes

[Q] “How are path-specific targets assigned to shared DAG nodes?”

We thank the reviewer for this question. Each target is attached to an edge, so a shared node receives one target per incoming path. Open placeholders are never shared, and after canonicalization only terminals (state variables and constants) are shared. Each root-to-node path therefore ends in a unique incoming edge, identified by its parent and operand slot. We compute the target by Eq. (8) along that path and store the resulting alignment feature and mask on that edge, where the graph transformer uses it together with the existing edge features. For nodes that are not shared, this is identical to the node-level description in the paper.

Revision: we now define the path-conditioned target, its mask, and the alignment feature on edges, as in our implementation (all results are unchanged), and revised Figure 1(a), Sec. 3.3, Sec. 3.4, and Appendix C.2 accordingly.


R3. Relation to semantic backpropagation

[W3] “Novelty and comparisons need closer treatment of semantic backpropagation [...]”

We appreciate the opportunity to clarify how S$^2$GSR differs from semantic backpropagation. Eq. (8) follows the inversion principle of semantic backpropagation; our novelty lies in how the target is used. We detail this difference in General Response G1.

AuthorsOpenReview ↗

Responses to Reviewer yfei [2/2]

R4. Library-based baselines and symmetry-invariant symbolic search

[W3] “Polynomial-library baselines cannot establish an advantage over symmetry-aware methods that also allow non-polynomial expressions.”

We thank the reviewer for this comment and apologize for the imprecise description in Sec. 4.1 and Appendix E.1, which called these libraries polynomial. In Table 3, the SINDy-style baselines on the NP systems, including EquivSINDy-c, already use non-polynomial libraries: following EquivSINDy, their terms are built from the symmetry invariants with the same elementary operators as S$^2$GSR, together with JJ. The limitation of library-based methods is therefore structural: the library terms need to be enumerated before fitting, and the discovered expression is restricted to a coefficient-fitted linear combination of these terms. A term such as sin⁡(r)e−r\sin(r)e^{-r} or −sin⁡(x1)/(1+x12)-\sin(x_1)/(1+x_1^2) is thus recovered only if this exact term is in the library, whereas free-form search constructs it during search. The libraries in Table 3 contain non-polynomial terms and building blocks such as sin⁡(∥x∥2)\sin(\|x\|^2), but not the exact nested ground-truth terms, which cannot be enumerated without knowing the answer in advance. This is why EquivSINDy-c reaches reasonable reconstruction (R2R^2 of 0.71 and 0.85) but zero recovery on the NP systems. Table 3 therefore already compares S$^2$GSR with a symmetry-aware method that allows non-polynomial expressions.

Revision: we corrected the library description in Sec. 4.1 and Appendix E.1, and replaced the “polynomial vs. non-polynomial” contrast with the distinction between fixed libraries and free-form search in the Abstract and Sec. 1, 2, and 4.

References:

  • Brunton, S. L., Proctor, J. L., & Kutz, J. N. (2016). Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proceedings of the National Academy of Sciences, 113(15), 3932–3937.
  • Yang, J., Rao, W., Dehmamy, N., Walters, R., & Yu, R. (2024). Symmetry-informed governing equation discovery. In Advances in Neural Information Processing Systems (NeurIPS).

[W3] “Novelty and comparisons need closer treatment of [...] symmetry-invariant symbolic search.”

Yang et al. (2026) enforce symmetry by using symmetry invariants as the inputs of free-form symbolic regression, which also allows such compositions. S$^2$GSR differs in three ways: (i) it does not require a precomputed set of invariants or equivariant basis fields; instead, invariant and equivariant sub-expressions, and the way they are combined, are found in a single search from typed operators such as ∥⋅∥2\|\cdot\|^2, the dot product, and JJ; (ii) the typed search space is combined with sub-expression semantic guidance, which adds to the gain from typing (Table 4); and (iii) invariant-basis constructions are contained in the SEG hypothesis space (Prop. B.1).

Revision: Sec. 1 and Sec. 2 now describe Yang et al. (2026) accurately, including its support for non-polynomial expressions, and state these differences.

References:

  • Yang, J., Bhat, M., Hu, B., Cao, Y., Dehmamy, N., Walters, R., & Yu, R. (2026). Discovering symbolic differential equations with symmetry invariants. Transactions on Machine Learning Research.

We thank the reviewer again for the time and constructive comments, which helped us improve the clarity and positioning of the paper. We hope our responses address your concerns, and we look forward to your feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Follow-up on Responses to Reviewer yfei

Dear Reviewer yfei,

Thank you for your careful review. As the discussion period closes today, we wanted to follow up on our response. In brief, we:

  1. clarified that scaling constraints are inactive under the NP-SO(2) specification, and listed the active signature components of each experiment (Appendix E.1, Table 8);
  2. explained that path-conditioned targets are attached to edges, so a shared terminal receives one target per incoming path;
  3. clarified the relation to semantic backpropagation (General Response G1);
  4. corrected the description of the library baselines, which already use non-polynomial invariant libraries, and stated how S$^2$GSR differs from Yang et al. (2026).

We would be glad to answer any remaining questions. If our response has addressed your concerns, we would appreciate it if you could consider updating your assessment. We would like to express our sincere appreciation again for your time and effort in reviewing our work.

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their careful reviews and constructive comments. In this general response, we address a concern shared by several reviewers; the remaining points are answered under each review. All revisions are marked in blue in the updated PDF.

G1. Relation to semantic backpropagation (Reviewers yfei, Z6Ew, and n9HH)

We agree that our original related work understated this connection. Eq. (8) inherits the inversion principle of semantic backpropagation; our novelty lies in how the target is used. Semantic backpropagation (Krawiec & Pawlak, 2013; Pawlak et al., 2015) treats the inverted target as a specification: the subtree is replaced by the library expression that best matches it. In ODE discovery, however, the root target is a derivative estimated from noisy data, and inversion can amplify this noise. S$^2$GSR therefore treats the target as evidence rather than a specification: its alignment with each closed sub-expression is a masked input to the policy and value networks, which learn how much to trust it during MCTS, and no expression is substituted. Because inversion follows the SEG typing rules, each target also keeps the type of its slot (e.g., an invariant scalar target for the coefficient of JxJx), so semantic guidance acts within the symmetry-constrained space. Table 4 isolates this signal: adding path-conditioned alignment to sub-expression semantics raises RR on NP-SO(2) from 0.28 to 0.36.

Revision: Sec. 2 now discusses semantic backpropagation and its follow-up work and separates the inherited principle from our contribution, and Sec. 1 now states the contribution more precisely.

References:

  • Krawiec, K., & Pawlak, T. (2013). Approximating geometric crossover by semantic backpropagation. In Proceedings of the 15th Annual Conference on Genetic and Evolutionary Computation (GECCO '13), 941–948.
  • Pawlak, T. P., Wieloch, B., & Krawiec, K. (2015). Semantic backpropagation for designing search operators in genetic programming. IEEE Transactions on Evolutionary Computation, 19(3), 326–340.
  • Pawlak, T. P., & Krawiec, K. (2018). Competent geometric semantic genetic programming for symbolic regression and Boolean function synthesis. Evolutionary Computation, 26(2), 177–212.
  • Virgolin, M., Alderliesten, T., & Bosman, P. A. N. (2019). Linear scaling with and within semantic backpropagation-based genetic programming for symbolic regression. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO), 1084–1092.

We thank the reviewers again for their time and look forward to further discussion of any remaining concerns.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.