LegoNE: Modular Tools for Equilibria, MARL & Education

• LegoNE is a suite of modular frameworks that employs LEGO-style building blocks for algorithm synthesis, multi-agent policy learning, and nuclear physics education.
• The algorithmic variant automates Nash equilibrium discovery via constrained optimization while the MARL variant uses equivariant GNNs for robust, zero-shot generalizable control.
• The educational toolkit constructs a 3D nuclide chart that visually represents nuclear stability and nucleosynthesis, enhancing hands-on learning through color and height coding.

LegoNE refers to distinct frameworks and educational tools, each leveraging the modular and compositional philosophy of "LEGO"-style building blocks. This term is notably used in three research domains: (1) a provable algorithm synthesis framework for Nash equilibria, (2) a graph neural network (GNN) architecture for multi-agent reinforcement learning (MARL), and (3) an educational toolkit for nuclear physics (“LEGO-based Nuclear Education”). Each variant embodies programmatic modularity, compositional reasoning, and domain-specific encoding, but serves fundamentally different scientific and educational aims.

1. Algorithmic LegoNE: Automated Nash Equilibrium Discovery

LegoNE, as introduced in "Discovering Expert-Level Nash Equilibrium Algorithms with LLMs" (Li et al., 16 Aug 2025), is a framework for synthesizing and certifying approximate Nash equilibrium algorithms in finite normal-form games. Given the intractability (PPAD-completeness) of computing exact Nash equilibria beyond two-player games, the focus is on $\epsilon$-approximate Nash equilibria (ANE), where each player can gain at most $\epsilon$ by unilaterally deviating.

Architectural Components

• Specification Language: Algorithms are described using a Python-like specification that restricts code to high-level "building blocks" such as BestResponse$_i$, UniformMixing, StationaryPoint, and OptimalMixing. Each block exposes axiomatic, Hoare-style contracts—for example, BestResponse$_2(x^1)=x^2$ guarantees $u_2(x^1, y^2)\le u_2(x^1,x^2)$ for any $y^2$—encoding mathematical properties of equilibrium-relevant operations.
• Automated Compilation: The LegoNE analyzer compiles code into a conjunction of logical formulas over payoff and regret variables, systematically instantiating quantifiers only on those strategies and payoffs constructed in the algorithm. Infinite-dimensional functions are "forgotten" via the Forgetting Principle—replacing function terms with real variables—yielding a finite algebraic system.
• Proof-Driven Optimization: The performance guarantee (e.g., achieving an $\epsilon$-ANE) is formalized as the universal validity of such formulas. LegoNE reduces correctness to solving a constrained optimization (typically QP/LP/SMT), extracting the tightest possible $\epsilon$ for the current algorithm.
• LLM-Driven Search: An LLM acts as an "explorer," proposing new block sequences. Candidate algorithms are certified if the formal solver can validate the regret bound. Human experts intervene by steering block selection, prompting, or requiring constraints (such as inclusion of StationaryPoint blocks), but the search and proof process are automated.

Empirical and Theoretical Results

• The framework rediscovered the post-2022 state-of-the-art $(1/3+\delta)$ approximation bound for two-player games within two search rounds, simulating in hours what took the theory community over a decade.
• For three-player games, LegoNE discovered a native algorithm (using StationaryPoint and asymmetric OptimalMixing) certified at $0.5+\delta$, improving over the best "extension" methods which previously achieved only $\epsilon$0. Of 12 distinct algorithms discovered, half achieved this $\epsilon$1 bound, several employing structurally novel architectures. Human proofs spanned multi-page derivations; LegoNE’s code spanned 30–60 lines, proved in hours.
• The Instantiation Lemma guarantees that universal properties over all strategies are faithfully captured via finitely many inequalities; the Correctness Theorem asserts that the output bound is certified for all games.

A key implication is the emergence of a human–machine symbiosis in theoretical science: the "language of thought" (block design) is provided by humans, while AI automates exploration and proof. Because the foundational principles (instantiation + forgetting) are domain-agnostic, extensions to polymatrix games, cryptographic protocols, and combinatorial optimization are plausible directions.

2. LEGO Framework in MARL: Equivariant GNNs for Swarm Control

The LEGO architecture (Local-Canonicalization Equivariant Graph Neural Networks), as developed for MARL in (Wang et al., 17 Sep 2025), targets sample-efficient, generalizable policy learning for multi-agent swarms, particularly in partially observed, competitive, and role-heterogeneous settings.

Formal Model

• Setting: $\epsilon$2-agent Markov game, where each agent $\epsilon$3 observes $\epsilon$4 (position, velocity) of all (or visible) others. Action space $\epsilon$5 (planar force). Agents may have preassigned roles (pursuer, evader, obstacle).
• Policy Network Pipeline: "Canonicalize → Encode → Decanonicalize."
  • Canonicalization: For each agent, constructs a local Euclidean frame via (i) normalizing velocity as local $\epsilon$6-axis, (ii) direction to center-of-mass as $\epsilon$7-axis, and (iii) applying the inverse pose to positions and velocities, providing E(2)-invariant node features ($\epsilon$8, $\epsilon$9).
  • Graph Representation: Heterogeneous graphs partitioned by roles; dense intra-role connections; no explicit edge features, with relative geometry encoded in node features.
  • GNN Encoding: Role-specific subgraphs are processed independently with L layers of a Graphormer-style encoder. Plural pooling yields permutation-equivariant features at the role level, concatenated for final node representation.
  • Decanonicalization: Global action is obtained by rotating the local action output back into world coordinates.
• Equivariance Guarantees: For any transformation $_i$0 (rotation, translation), the policy satisfies $_i$1.

Learning and Evaluation

• LEGO-MAPPO: Integration with MAPPO under centralized training and decentralized execution (CTDE). The full algorithm is loss-augmented PPO with value, surrogate, and entropy losses; no explicit equivariance regularization is needed due to architectural invariance/enforcement by construction.
• Empirical Results:
  • LEGO-MAPPO consistently outperforms baselines that employ only geometric equivariance or only permutation equivariance. On the MPE Spread benchmark, convergence to near-optimal solutions is achieved in $_i$2 steps (vs $_i$3 for MAPPO).
  • Zero-shot generalization holds across variable team sizes and out-of-distribution initializations, with performance robustly exceeding MAPPO-GNN and non-equivariant variants.
  • In real-world drone experiments, LEGO enables robust adaptation to agent failures (e.g., the remaining pursuer adapts when one fails), with mission objectives maintained across all trial runs.

The fusion of local canonicalization, permutation-equivariant GNN encoding, and heterogeneous role partitioning leads to policies that are both mathematically guaranteed to respect group symmetries and empirically robust to variations in agent count, team failure, and deployment domain.

3. LEGO-Based Nuclear Education (LegoNE) as a Pedagogical Toolkit

The "Binding Blocks" initiative, also referred to as LEGO-based Nuclear Education (LegoNE), represents a physically embodied approach to teaching nuclear physics, astrophysics, and chemistry via the construction of a large three-dimensional nuclide chart with LEGO® bricks (Diget et al., 2016).

Physical and Pedagogical Design

• Construction: The core activity is assembling a $_i$4 chart of observed isotopes (A,Z plane), using over 26,000 2×4 LEGO® bricks. Participants build "towers" at each $_i$5 coordinate.
  • Color-Coding: Dominant nuclear decay mode identifies each isotope: e.g., black (stable), yellow (α-emitter), light blue (β$_i$6), red (β$_i$7 or EC), dark blue (neutron-unbound), orange (proton-unbound), green (spontaneous fission).
  • Height Encoding: Tower height is proportional to mass-excess per nucleon, with a shift so that $_i$8 sits at unity. Each layer: 25,000 GJ/kg, facilitating energy-scale comparisons.
• Process Visualization: The chart makes visible, at a glance:
  • The valley of stability
  • Pathways for s-process/r-process nucleosynthesis and stellar fusion
  • Roles of isotopes in medical physics (e.g., $_i$9, $_2(x^1)=x^2$0, $_2(x^1)=x^2$1).
• Curricular Integration: Mapped to GCSE/A-level (UK) curricula:
  • Nuclear decay identification and half-life calculation by chart color
  • Binding energy computations from tower heights
  • Diagrammatic tracing of nucleosynthesis pathways and decay chains

Implementation and Outcomes

• The full chart requires 38 plates, with towers rising up to 36 layers.
• Modular subsets allow tailoring to class size: e.g., the Ne region (A≤20, ~2,000 bricks) is portable for single sessions; Fe region (A≤56, ~7,000 bricks) supports group learning.
• Events such as the STFC Daresbury Open Day with over 1,000 participants report increased engagement, better conceptual understanding, and higher-quality classroom discussions compared to traditional 2D representations.
• Anecdotal and instructor feedback point to inquiry-driven learning, improved grasp of nuclear stability, and effective illustration of decay energetics.

4. Comparative Table of Major LegoNE Variants

Domain	Core Principle	Key Output/Impact
Nash Equilibrium Algorithms (Li et al., 16 Aug 2025)	Block-based symbolic algorithm search	SOTA-certified $_2(x^1)=x^2$2-ANE; machine-certified proofs; new 3-player ANE algorithms
MARL GNNs for Swarm Control (Wang et al., 17 Sep 2025)	GNN with local canonicalization	E(2)- and permutation-equivariant, robust, zero-shot generalization
Nuclear Education Kit (Diget et al., 2016)	Tactile, coded model of nuclides	3D chart elucidates nuclear stability, nucleosynthesis, decay Hunter

This table is limited to information present in the cited research.

5. Significance and Broader Implications

The recurring "LEGO" motif in these frameworks indicates a shift towards compositional, modular, and physically or mathematically grounded educational and algorithmic tools. In algorithmic discovery, LegoNE demonstrates that human–machine co-design—where people specify high-level symbolic languages and machines exhaustively search and rigorously prove—can compress theoretical advances that previously required years. In MARL, the compositional GNN approach achieves both group-theoretic optimality (equivariance) and robust performance in nonstationary, partially observed, and role-diverse domains. The educational variant shows physical modularity’s efficacy in bridging abstract concepts and concrete learning objectives.

A plausible implication is that the modular, compositional architecture underpinning these LegoNE instances will influence future frameworks in other domains, including cryptography, optimization, systems biology, and STEM education, wherever formal guarantees, interpretability, or engagement with high-dimensional structure are required.