EGG-SR: Embedding Symbolic Equivalence into Symbolic Regression via Equality Graph¶
Conference: ICLR 2026 arXiv: 2511.05849 Code: Project Page Area: Reinforcement Learning Keywords: Symbolic Regression, Symbolic Equivalence, Equality Graph, Monte Carlo Tree Search, Deep Reinforcement Learning
TL;DR¶
This paper proposes Egg-SR, a unified framework that embeds symbolic equivalence via equality graphs (e-graphs) into three categories of symbolic regression methods—MCTS, DRL, and LLM—achieving subtree pruning, policy gradient variance reduction, and feedback prompt enrichment, respectively. Theoretical results prove that Egg-MCTS tightens the regret bound and Egg-DRL reduces gradient estimation variance, while experiments consistently validate improved expression discovery accuracy.
Background & Motivation¶
Symbolic Regression (SR) aims to discover closed-form physical laws from experimental data and is a key task in AI-driven scientific discovery. The search space grows exponentially with data dimensionality, posing significant computational challenges.
A promising but underexplored direction is symbolic equivalence: many syntactically distinct expressions define the same function. For example: - \(\log(x_1^2 x_2^3)\) - \(\log(x_1^2) + \log(x_2^3)\) - \(2\log(x_1) + 3\log(x_2)\)
These three expressions are mathematically identical, yet existing algorithms treat them as distinct outputs, leading to:
Redundant exploration: MCTS independently searches equivalent subtrees, wasting computation.
Slow training: DRL cannot share reward signals across equivalent sequences, resulting in high gradient variance.
Insufficient feedback: LLMs observe only a single expression form and miss information from equivalent variants.
The core challenge is: how to represent symbolically equivalent expressions in a unified and scalable manner and embed them into modern learning frameworks? Explicitly enumerating equivalent variants is infeasible, as their count grows exponentially with expression length.
Method¶
Overall Architecture¶
The Egg-SR framework consists of a core module Egg (Equality graph for Grammar-based expressions) and three integration interfaces:
- Egg-MCTS: Shares search statistics across equivalent paths and prunes redundant subtrees.
- Egg-DRL: Aggregates probabilities over equivalent sequences to reduce policy gradient variance.
- Egg-LLM: Enriches feedback prompts with equivalent expressions to guide subsequent predictions.
Key Design 1: The Egg Module — Grammar-Based Equality Graph¶
Expression Representation: Symbolic expressions are represented using a context-free grammar \(\langle V, \Sigma, R, S \rangle\), where \(V=\{A\}\) denotes non-terminals, \(\Sigma\) denotes terminals (variables and constants), and \(R\) denotes production rules. Expressions are constructed by sequentially applying rules to the leftmost non-terminal.
E-graph Data Structure: - Composed of equivalence classes (e-classes), each containing a set of equivalent e-nodes. - Each e-node encodes a mathematical operation and references child e-classes. - Common subexpressions are shared, enabling compact representation.
Construction Process (Equality Saturation): 1. Initialize the e-graph with the production sequence of the input expression. 2. For each rewrite rule (e.g., \(\log(ab) \leadsto \log(a) + \log(b)\)), match the LHS pattern. 3. Create new e-classes and e-nodes corresponding to the RHS at matched locations. 4. Merge equivalent e-classes. 5. Iterate until saturation or a maximum number of iterations is reached.
Extraction Strategies: (1) Cost-based extraction — selects the most simplified expression; (2) Random walk sampling — generates a batch of equivalent expressions.
Key Design 2: Egg-MCTS — Equivalence-Aware Backpropagation¶
Standard MCTS consists of four steps: selection (UCT) → expansion → simulation → backpropagation. Egg-MCTS modifies the backpropagation step:
- Converts the current path (partially completed expression) into an e-graph and saturates it.
- Samples multiple equivalent sequences from the saturated e-graph.
- Checks whether corresponding paths exist in the search tree.
- Simultaneously updates the reward and visit counts of all equivalent paths.
This changes the semantics of the UCT formula: \(\mathtt{visits}(s)\) no longer counts visits to a specific node, but rather visits to any representative of the equivalence class, analogous to a transposition table.
Example: Path 1 corresponds to \(\log(x_1 \times A)\) and Path 2 to \(\log(x_1) + \log(A)\); the two are equivalent under \(\log(ab) \leadsto \log(a)+\log(b)\). Standard MCTS explores both subtrees independently, whereas Egg-MCTS shares statistics to avoid redundancy.
Key Design 3: Egg-DRL — Equivalence-Aware Policy Gradient¶
Standard DRL uses a sequence decoder to sample production sequences \(\tau\), with the policy gradient estimator:
Egg-DRL improvement: for each sampled sequence \(\tau_i\), an e-graph is constructed and \(K-1\) equivalent sequences are sampled, yielding the equivalence-aware estimator:
The key modification replaces the single-sequence probability with the sum of probabilities over equivalent sequences \(\sum_k p_\theta(\tau_i^{(k)})\). Since equivalent sequences share the same reward, averaging over sequences with equal rewards reduces within-group variability.
Key Design 4: Egg-LLM — Equivalence-Augmented Feedback Prompts¶
The iterative LLM-based SR pipeline proceeds as: hypothesis generation → data evaluation → experience management. Egg-LLM intervenes at the feedback stage by parsing generated Python expressions into symbolic form, converting them into e-graphs to extract equivalent variants, and incorporating these variants into the feedback prompt, allowing the LLM to observe richer functional equivalences.
Theoretical Guarantees¶
Theorem 3.1 (Egg-MCTS Regret Bound): Let \(\kappa\) denote the near-optimal branching factor of standard MCTS and \(\kappa_\infty\) that of Egg-MCTS; then \(\kappa_\infty \leq \kappa\), and the regret bound tightens from \(\tilde{\mathcal{O}}(n^{-\frac{\log(1/\gamma)}{\log\kappa}})\) to \(\tilde{\mathcal{O}}(n^{-\frac{\log(1/\gamma)}{\log\kappa_\infty}})\).
Theorem 3.2 (Egg-DRL Variance Reduction): The Egg-DRL gradient estimator is unbiased and its variance is strictly no greater than that of standard DRL: \(\mathbb{V}\mathrm{ar}[g_\mathtt{egg}(\theta)] \leq \mathbb{V}\mathrm{ar}[g(\theta)]\).
Key Experimental Results¶
Main Results: Trigonometric Dataset NMSE¶
| Method | (2,1,1) Noiseless | (3,2,2) Noiseless | (4,4,6) Noiseless | (2,1,1) Noisy | (3,2,2) Noisy |
|---|---|---|---|---|---|
| MCTS | 0.006 | 0.033 | 0.144 | 0.015 | 0.007 |
| Egg-MCTS | <1E-6 | <1E-6 | 0.006 | 0.005 | 0.012 |
| DRL | 0.030 | 0.277 | 2.990 | 0.09 | 0.44 |
| Egg-DRL | 0.020 | 0.161 | 2.381 | 0.07 | 0.35 |
Egg-MCTS reduces NMSE from 0.006/0.033 to <1E-6 on simple noiseless cases, an improvement of several orders of magnitude.
Main Results: Scientific Benchmark LLM Comparison¶
| Method | Oscillation I IID | Oscillation II OOD | Bacterial OOD | Stress-Strain OOD |
|---|---|---|---|---|
| LLM-SR (GPT3.5) | <1E-6 | 3.81E-5 | 0.0264 | 0.0516 |
| Egg-LLM (GPT3.5) | <1E-6 | <1E-6 | 0.0198 | 0.0419 |
| LLM-SR (Mistral) | <1E-6 | 0.0291 | 0.0037 | 0.0946 |
| Egg-LLM (Mistral) | <1E-6 | 0.0114 | 0.0107 | 0.0754 |
Efficiency Analysis¶
Space Efficiency: For \(\log(x_1 \times \cdots \times x_n)\), there are \(2^{n-1}\) equivalent variants; array-based storage grows exponentially, whereas e-graphs achieve significant compression through subexpression sharing.
Time Efficiency: The overhead introduced by the Egg module is negligible; compared to the cost of coefficient fitting and parameter updates in DRL, e-graph construction and equivalent sequence sampling account for a minimal fraction of total runtime.
Key Findings¶
- Egg consistently improves performance across all three algorithm classes (MCTS/DRL/LLM), validating the framework's universality.
- Egg-MCTS maintains broader and deeper search trees, exploring a larger and more diverse search space.
- The gradient estimation variance of Egg-DRL decreases substantially during training (Figure 3, right).
- Trigonometric expressions benefit most due to the abundance of equivalent variants (trigonometric identities).
Highlights & Insights¶
- Unified Framework: A single Egg module serves all three method classes—MCTS, DRL, and LLM—through a consistent interface.
- Dual Validation via Theory and Experiment: Theorems 3.1 and 3.2 formally prove MCTS regret tightening and DRL variance reduction, respectively, and are fully corroborated by experiments.
- The core advantage of e-graphs lies in shared subexpressions enabling exponential compression; at \(n=8\) variables, the memory advantage exceeds 100×.
- Equivalences not captured by e-graph rewriting (e.g., coefficient equivalence in SymNet) point to interesting open problems.
Limitations & Future Work¶
- Rewrite rule sets require manual definition; current coverage includes logarithmic and trigonometric identities, and extension to more complex mathematical identities remains future work.
- For expressions lacking rich equivalent variants (e.g., pure polynomials), the benefit of Egg is limited.
- Only grammar-based expression representations are currently supported; layered representations based on SymNet require further extension.
- How to optimally leverage Egg at inference time (e.g., in LLMs) remains an open question.
- The termination condition (maximum iteration count) for e-graph saturation requires careful tuning.
Related Work & Insights¶
- E-graphs originate from program synthesis and compiler optimization; this work is the first to systematically embed them into the learning algorithms of symbolic regression.
- Unlike the use of e-graphs in GP for deduplication and simplification, Egg-SR exploits equivalent variants for equivalence-aware learning.
- Sharing statistics across equivalent paths in MCTS is analogous to transposition tables in game playing, but requires pattern matching for symbolic equivalence.
- The idea of probability aggregation in DRL generalizes to any sequence generation task with output equivalence.
- Knowledge-guided scientific discovery is an important direction in SR; Egg-SR is orthogonal to symmetry constraints and unit constraints as in AI-Feynman.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ (Unified framework + theoretical guarantees; first systematic embedding of symbolic equivalence)
- Experimental Thoroughness: ⭐⭐⭐⭐ (Covers MCTS/DRL/LLM with efficiency analysis)
- Writing Quality: ⭐⭐⭐⭐⭐ (Well-crafted figures; tight theory–experiment integration)
- Value: ⭐⭐⭐⭐ (Substantial advancement for SR; generalizable to other equivalence settings)