Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models

arXiv:2609.08073 · quant-ph · Submitted 2026-09-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: We're looking at the paper: "Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models".

Mira: Here's the gist of the paper. Designing quantum optical experiments requires searching over discrete circuit topologies and continuous parameters, often with multiple realizations of the same target state <ref:2609.08073#pg2> The framework introduces Grinch, a bottom-up reward-driven generative framework that learns to sample optical graphs directly from fidelity-based rewards without relying on a pre-existing training dataset <ref:2609.08073#pg2> It demonstrates the framework on multipartite entangled states, graph and cluster states, and nonlocal photonic-gate targets, obtaining multiple high-fidelity optical graphs <ref:2609.08073#pg2> Pruning redundant reasoning tokens The paper introduces Grinch to address the challenge of designing quantum optical experiments by constructing circuit topologies edge by edge and learning directly from reward evaluations without needing a preexisting training dataset <ref:2609.08073#pg2> This bottom-up approach contrasts with top-down methods that optimize and prune a single dense graph, as Grinch constructs circuit topologies edge by edge <ref:2609.08073#pg2> The framework utilizes GraphFlowNets to sample circuit graphs from a reward-proportional distribution where Pθ(G) is proportional to the reward R(G) <ref:2609.08073#pg2> How the training signal is built The training objective for GFlowNets is defined by the trajectory balance loss, which balances forward and backward policies based on a learnable estimate of the partition function Zθ <ref:2609.08073#pg2> The trajectory balance loss is defined as LT B(τ) = log Zθ Qlτ t=1 PF (stst−1) R(sf) Qlτ t=1 PB(st−1st)!2 <ref:2609.08073#pg2> The reward R(sT) is calculated by optimizing the real edge weights using L-BFGS-B with bounds

−1, 1: and assigning R(sf) = F(sf, wopt) 2 <ref:2609.08073#pg2> Where the gains come from The research demonstrates that allowing Grinch to sample circuits with more edges reduces the overall number of samples required to discover high-reward solutions compared to a strict campaign <ref:2609.08073#pg2> The relaxed budget finds a unit-fidelity solution in fewer than 3,000 samples for the SRV (5, 4, 4)⟩ ⊗ 0⟩ state when using masking functions <ref:2609.08073#pg2> Furthermore, diverse sampling allows Grinch to recover multiple high-reward modes when several arrays are possible <ref:2609.08073#pg2> Comparison with other approaches Grinch's performance is compared against four reference frameworks for the seven shared target states reported in Ref.

13: and shows that Grinch can match the solutions found by such approaches while maintaining a consistent number of edges across the solutions <ref:2609.08073#pg2> The paper also notes that Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints <ref:2609.08073#pg2> Grinch is a GFlowNet-based framework that samples quantum optical circuits in proportion to a task-defined reward <ref:2609.08073#pg2> By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks <ref:2609.08073#pg2> Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions <ref:2609.08073#pg2> Modest flexibility beyond the strict terminal edge budget accelerated the discovery of high-reward circuits after which pruning removed redundant edges while maintaining fidelity <ref:2609.08073#pg2> Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints as demonstrated in the no-ancilla communication experiment reported here <ref:2609.08073#pg2> The broader applicability of this approach depends on a sufficiently efficient simulator or optimizer to evaluate terminal rewards since these evaluations are performed repeatedly during training <ref:2609.08073#pg2> Page 1 --- Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models Grinch is a GFlowNet-based framework that samples quantum optical circuits in proportion to a task-defined reward <ref:2609.08073#pg2> By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks <ref:2609.08073#pg2> Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions <ref:2609.08073#pg2> Modest flexibility beyond the strict terminal edge budget accelerated the discovery of high-reward circuits after which pruning removed redundant edges while maintaining fidelity <ref:2609.08073#pg2> Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints as demonstrated in the no-ancilla communication experiment reported here <ref:2609.08073#pg2> Page 2 --- 2 A.

Mira: Quantum Optics Experiments as Graphs Photonic quantum experiments can be encoded in weighted colored graphs G(V, E), where the graph nodes (V) correspond to spatial photon paths and the edges (E) between them encode the mode number of the photons through a specific color <ref:2609.08073#pg2> The edge weight gives the amplitude associated with the photons <ref:2609.08073#pg2> For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively <ref:2609.08073#pg2> This can be translated into the operator w ki,kj a† ka b† + h.c.

Kai: with a† ka the creation operator of a photon on path a with mode ka <ref:2609.08073#pg2> The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function Φ(w) = X m 1 m!

Paper summary: Kai: So, we're looking at this paper called "Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models". Basically, it tackles the problem of designing quantum optical experiments by searching through discrete circuit topologies and continuous parameters all at once, often needing multiple different ways to build the same target state.

Mira: And what they introduce is a framework called Grinch which is a bottom-up reward driven generative framework that learns to sample optical graphs directly from fidelity based rewards without needing any pre existing training data.

Lev: That sounds ambitious because usually these top down methods optimize one dense graph and then prune it down, but this one builds the circuit edge by edge.

Kai: Exactly. So, the main thesis is that they move away from that dense graph optimization idea and instead sample graphs from a reward proportional distribution where P theta of G is proportional to the reward R of G.

Mira: And they claim this allows them to do it without having a pre existing training dataset, which is pretty big for generative models.

Lev: I'm curious about how they actually build that training signal because sampling from a reward distribution sounds hard.

Kai: Well, the paper explains that the training objective for their GraphFlowNets is defined by something called the trajectory balance loss.

Mira: That loss balances forward and backward policies based on a learnable estimate of what they call the partition function, Z theta.

Lev: So if I'm thinking about this for real hardware, how does that Z theta estimate work in practice? Is it just some learned approximation?

Mira: It’s a learnable estimate, yes, and minimizing that trajectory balance loss essentially balances the probability of each forward construction trajectory with the reward weighted probability of its reverse trajectory.

Kai: That means different paths can contribute to the same final graph structure if they both lead to a high reward.

Lev: That makes sense for exploration because it lets them explore different topologies without getting stuck on just one path. But what about how they define that reward R of G?

Kai: The reward is calculated by optimizing the real edge weights using L-BFGS B with bounds of minus one to one, and then assigning the reward R of a final object as F of s f w opt squared.

Mira: So they are using actual fidelity values to define what's good in the circuit design.

Lev: And that sounds like it couples discrete things—the graph structure—with continuous things—those edge weights—in a very specific way. Where does that come from physically?

Kai: They model photonic quantum experiments as weighted colored graphs G of nodes V corresponding to spatial photon paths and edges E encoding the mode number through a specific color.

Mira: The edge weight itself gives the amplitude associated with those photons, like how an edge between paths a and b with modes ka and kb translates into an operator w ki,kj a dagger ka b dagger kb plus h.c.

Paper summary: Lev: So they’re connecting the graph structure directly to the underlying quantum state operators through these weights. That's where my work on error correction comes in; how does this sampling handle those continuous weights?

Kai: They use masked, edge by edge graph construction and continuous edge weight optimization, and they restrict exploration through structural masks.

Mira: The whole point of the masking is to keep the search focused while still allowing for flexibility.

Lev: If you’re restricting exploration with masks, does that limit you from finding solutions that might be really far away in the search space?

Kai: They found that a modest flexibility beyond the strict terminal edge budget actually accelerated discovery of high reward circuits before pruning removed redundant edges while maintaining fidelity.

Mira: They showed this on multipartite entangled states, graph and cluster states, and nonlocal photonic gate targets.

Lev: And they did get multiple high fidelity optical graphs from this process. That’s a good result for a bottom up approach to finding actual experimental setups.

Kai: It does, but the paper also points out that Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs, and it can derive alternative arrays for nonlocal quantum gates under hardware constraints.

Mira: That's important because sometimes the perfect graph just doesn't exist, so finding an approximation is valuable.

Lev: Finding alternatives when the exact solution is impossible might be useful when we’re constrained by real hardware limitations in building those nonlocal gates. But what are the practical implications for someone building a circuit?

Kai: The main implication is that Grinch provides a way to search for solutions where you don't have to start with a dense graph and try to prune it down.

Mira: It’s about learning directly from the target state reward, which bypasses needing a pre existing training dataset.

Lev: So, if we were trying to design an experiment for something complex, this suggests we could just let the system generate circuits based on what it wants and then refine them with rewards.

Kai: That’s the general idea behind Grinch in "Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models." It shifts the focus from optimizing a single dense graph to generating many good ones through reward sampling.

Mira: And I think for us, as theorists, it's interesting because it shows a way to generate multiple high fidelity optical graphs across different state types like multipartite entangled states.

Lev: From an error correction standpoint, the idea of identifying asymptotic solutions when exact graphs are impossible is something we need to keep in mind for our hardware constraints.

Kai: So, the authors show this framework works across several benchmarks, and it seems to handle those coupled discrete continuous design problems pretty well.

Conclusion: Kai: So we're wrapping up this look at the paper called "Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models."

Mira: It boils down to this framework, Grinch, which is a bottom-up reward driven generative model that builds quantum optical circuits edge by edge.

Kai: Right. The authors are looking at how we can design these complex experiments without needing a massive pre existing training dataset.

Mira: They show that they can sample these optical graphs directly based on fidelity rewards, meaning the system learns what a good circuit looks like through trial and error from the start.

Kai: It’s about constructing the discrete structure of the experiment, but instead of optimizing one big graph at once, you build it piece by piece.

Mira: And they use something called GraphFlowNets to do this sampling according to a reward distribution proportional to how good each potential graph is.

Kai: So what does this mean for designing real experiments? It suggests a new way to explore the huge space of possible circuit topologies.

Mira: It means they can find multiple high fidelity optical graphs for different states, like multipartite entangled states and nonlocal photonic gate targets.

Kai: But the authors also found that even with some flexibility beyond the strict edge budget, you can speed up finding those good circuits before pruning gets in the way.

Mira: They also managed to find asymptotic solutions for some states where a perfect graph just isn't possible, which is important for real hardware constraints.

Kai: It really shifts the focus from top down optimization to this generative sampling approach.

Mira: And it opens up avenues for designing more robust experimental setups across different quantum states.

Department of Chemistry, McMaster University · Département de Physique, Université de Montréal · Brockhouse Institute for Materials Research, McMaster University

quant-ph

Submitted: 2026-09-08

Updated: 2026-11-11

Comments: 21 pages, 8 figures, 8 tables

Code: https://github.com/ishume94/Grinch

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 89/100

The gist: Designing quantum optical experiments requires searching over discrete circuit topologies and continuous parameters…

Key concepts

Quantum Optical Experiments as Graphs
Photonic experiments are modeled as weighted colored graphs where nodes represent spatial photon paths and edges encode mode numbers or colors. These graph structures directly translate into the mathematical operators that describe the resulting quantum state of the optical setup.
Grinch Framework
Grinch is a generative framework that treats circuit design as sampling from a reward-proportional distribution rather than optimizing a single dense graph. It constructs circuits incrementally, allowing it to explore many possible topologies and find high-fidelity solutions without needing prior training data.
Trajectory Balance Loss
This loss function trains the generative model by balancing the forward and backward policies. It uses an estimated partition function to ensure that the probability of constructing a circuit is proportional to its calculated reward, guiding the system toward optimal designs.

Terminology

Summary

Designing quantum optical experiments requires searching over discrete circuit topologies and continuous parameters, often with multiple realizations of the same target state <ref:2609. p. 0

Where the paper says it

Huidobro-Meezs,1, ∗ Simón Paiva-Ortega,2 and Rodrigo A. Vargas-Hernández1, 3, † 1Department of Chemistry, McMaster University, Hamilton, ON, Canada 2Département de Physique, Université de Montréal, Montréal, QC, Canada 3Brockhouse Institute for Materials Research, McMaster University, Hamilton, ON, Canada (Dated: September 9, 2026) Designing quantum optical experiments requires searching over discrete circuit topologies and continuous parameters, often with multiple realizations of the same target state.

08073#pg2> The framework introduces Grinch, a bottom-up reward-driven generative framework that learns to sample optical graphs directly from fidelity-based rewards without relying on a pre-existing training dataset <ref:2609. p. 0
Where the paper says it

Graph-based methods commonly address this problem by optimizing a dense graph and pruning it toward a single circuit. We introduce Grinch, a bottom-up reward-driven generative framework that learns to sample optical graphs directly from fidelity-based rewards without relying on a pre-existing training dataset.

08073#pg2> It demonstrates the framework on multipartite entangled states, graph and cluster states, and nonlocal photonic-gate targets, obtaining multiple high-fidelity optical graphs <ref:2609. p. 0
Where the paper says it

We introduce Grinch, a bottom-up reward-driven generative framework that learns to sample optical graphs directly from fidelity-based rewards without relying on a pre-existing training dataset. We demonstrate the framework on multipartite entangled states, graph and cluster states, and nonlocal photonic-gate targets, obtaining multiple high-fidelity optical graphs.

08073#pg2>

Pruning redundant reasoning tokens

The paper introduces Grinch to address the challenge of designing quantum optical experiments by constructing circuit topologies edge by edge and learning directly from reward evaluations without needing a preexisting training dataset <ref:2609.08073#pg2> This bottom-up approach contrasts with top-down methods that optimize and prune a single dense graph, as Grinch constructs circuit topologies edge by edge <ref:2609. p. 0

Where the paper says it

Unlike existing top-down approaches that optimize and prune a single dense graph, Grinch constructs circuit topologies edge by edge and learns directly from reward evaluations without requiring a preexisting training dataset. Within the framework for AIdriven experimental design proposed in Ref.

p. 0
Where the paper says it

Unlike existing top-down approaches that optimize and prune a single dense graph, Grinch constructs circuit topologies edge by edge and learns directly from reward evaluations without requiring a preexisting training dataset. Within the framework for AIdriven experimental design proposed in Ref.

08073#pg2> The framework utilizes GraphFlowNets to sample circuit graphs from a reward-proportional distribution where Pθ(G) is proportional to the reward R(G) <ref:2609. p. 1
Where the paper says it

This motivates bottom-up construction as an alternative framework that may require fewer structural operations during the search. In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G).

p. 1
Where the paper says it

This motivates bottom-up construction as an alternative framework that may require fewer structural operations during the search. In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G).

08073#pg2>

How the training signal is built

The training objective for GFlowNets is defined by the trajectory balance loss, which balances forward and backward policies based on a learnable estimate of the partition function Zθ <ref:2609.08073#pg2> The trajectory balance loss is defined as LT B(τ) = log Zθ Qlτ t=1 PF (stst−1) R(sf) Qlτ t=1 PB(st−1st)! p. 8

Where the paper says it

For this work we consider the trajectory balance loss [25] function which gives an increased credit assignment making the training protocol more efficient. The trajectory balance loss is defined as: LT B(τ ) = log Zθ Qn t=1 Pθ(st|st−1) R(sn) Qn t=1 Pθ(st−1|st) 2 (A11) with sn ∈ Sf .

2 <ref:2609. p. 0
Where the paper says it

[4], Grinch addresses experimental-design problems involving both discrete structures and coupled discrete–continuous variables: the circuit topology is sampled discretely, while its continuous edge weights are optimized to maximize the reward. arXiv:2609.

08073#pg2> The reward R(sT) is calculated by optimizing the real edge weights using L-BFGS-B with bounds [−1, 1] and assigning R(sf) = F(sf, wopt) 2 <ref:2609. p. 2
Where the paper says it

4, using Adam [28] with a learning rate of 10−3 ; see the Appendix for further details. For each sampled terminal graph sf eligible for reward evaluation, we optimize its real edge weights using L-BFGS-B [29] with bounds [−1, 1] and assign the reward R(sf ) = F(sf , wopt) 2 , where wopt denotes the weights returned by the optimizer.

08073#pg2>

Where the gains come from

The research demonstrates that allowing Grinch to sample circuits with more edges reduces the overall number of samples required to discover high-reward solutions compared to a strict campaign <ref:2609.08073#pg2> The relaxed budget finds a unit-fidelity solution in fewer than 3,000 samples for the SRV (5, 4, 4)⟩ ⊗ 0⟩ state when using masking functions <ref:2609. p. 3

Where the paper says it

7 in the Appendix, we tested our approach on the |SRV (5, 4, 4)⟩ ⊗ |0⟩ state without the logical masking functions to isolate the effect of the trajectory length. We observe that the relaxed campaign finds a unit-fidelity solution in fewer than 3,000 samples, whereas the strict campaign requires more than 12K samples to obtain a similar solution.

08073#pg2> Furthermore, diverse sampling allows Grinch to recover multiple high-reward modes when several arrays are possible <ref:2609. p. 4
Where the paper says it

When multiple arrays are possible, each with potential hardware differences, GFlowNets can effectively explore these regions and identify multiple high-reward modes. When the solution is concentrated in a specific section of the solution space, the sampling process allows to identify that mode; however, for top-down approaches, such problems might be more difficult, as this means that if a pruning step in the topological optimization removed an edge which is necessary, the end result would not have the best possible fidelity, explaining the high variance in these approaches [13].

08073#pg2>

Comparison with other approaches

Grinch's performance is compared against four reference frameworks for the seven shared target states reported in Ref. [13] and shows that Grinch can match the solutions found by such approaches while maintaining a consistent number of edges across the solutions <ref:2609. p. 3

Where the paper says it

Table IV compares Grinch with four reference frameworks for the seven shared target states reported in Ref. We observe that Grinch can match the solutions found by such approaches while maintaining a consistent number of edges across the solutions.

08073#pg2> The paper also notes that Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints <ref:2609. p. 7
Where the paper says it

Moreover, when alternative implementations exist, the sampled solutions occupy multiple high-reward regions of graph space, demonstrating that Grinch can recover diverse circuit realizations rather than a single optimum. Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints, as demonstrated in the no-ancilla communication experiment reported here.

08073#pg2>

Summary of findings

Grinch is a GFlowNet-based framework that samples quantum optical circuits in proportion to a task-defined reward <ref:2609.08073#pg2> By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks <ref:2609. p. 7

Where the paper says it

Unlike conventional generative models that learn from an existing dataset, Grinch learns directly from reward evaluations, enabling circuit discovery when representative training data is unavailable. By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks.

p. 7
Where the paper says it

Unlike conventional generative models that learn from an existing dataset, Grinch learns directly from reward evaluations, enabling circuit discovery when representative training data is unavailable. By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks.

08073#pg2> Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions <ref:2609. p. 7
Where the paper says it

By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks. Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions.

p. 7
Where the paper says it

By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks. Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions.

08073#pg2> Modest flexibility beyond the strict terminal edge budget accelerated the discovery of high-reward circuits after which pruning removed redundant edges while maintaining fidelity <ref:2609. p. 7
Where the paper says it

Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions. Modest flexibility beyond the strict terminal edge budget accelerated the discovery of high-reward circuits, after which pruning removed redundant edges while maintaining fidelity.

p. 7
Where the paper says it

Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions. Modest flexibility beyond the strict terminal edge budget accelerated the discovery of high-reward circuits, after which pruning removed redundant edges while maintaining fidelity.

08073#pg2> Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints as demonstrated in the no-ancilla communication experiment reported here <ref:2609. p. 7
Where the paper says it

Moreover, when alternative implementations exist, the sampled solutions occupy multiple high-reward regions of graph space, demonstrating that Grinch can recover diverse circuit realizations rather than a single optimum. Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints, as demonstrated in the no-ancilla communication experiment reported here.

p. 7
Where the paper says it

Moreover, when alternative implementations exist, the sampled solutions occupy multiple high-reward regions of graph space, demonstrating that Grinch can recover diverse circuit realizations rather than a single optimum. Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints, as demonstrated in the no-ancilla communication experiment reported here.

08073#pg2> The broader applicability of this approach depends on a sufficiently efficient simulator or optimizer to evaluate terminal rewards since these evaluations are performed repeatedly during training <ref:2609. p. 7
Where the paper says it

Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints, as demonstrated in the no-ancilla communication experiment reported here. The broader applicability of this approach depends on a sufficiently efficient simulator or optimizer to evaluate terminal rewards, since these evaluations are performed repeatedly during training.

08073#pg2>

--- Page 1 ---

Bottom-Up Design of Quantum Optical Experiments Using Discrete Generative Models

Grinch is a GFlowNet-based framework that samples quantum optical circuits in proportion to a task-defined reward <ref:2609.08073#pg2> By combining masked, edge-by-edge graph construction with continuous edge-weight optimization, the framework addresses coupled discrete–continuous design problems while restricting exploration through structural masks <ref:2609.08073#pg2> Across multipartite-entangled-state, graph- and clusterstate, and nonlocal photonic-gate benchmarks, Grinch identified multiple high-fidelity optical-graph solutions <ref:2609.08073#pg2> Modest flexibility beyond the strict terminal edge budget accelerated the discovery of high-reward circuits after which pruning removed redundant edges while maintaining fidelity <ref:2609.08073#pg2> Importantly, Grinch can identify asymptotic solutions for states that cannot be exactly generated by graphs and derive alternative arrays for nonlocal quantum gates under hardware constraints as demonstrated in the no-ancilla communication experiment reported here <ref:2609.08073#pg2>

--- Page 2 ---

2

A. Quantum Optics Experiments as Graphs

Photonic quantum experiments can be encoded in weighted colored graphs G(V, E), where the graph nodes (V) correspond to spatial photon paths and the edges (E) between them encode the mode number of the photons through a specific color <ref:2609.08073#pg2> The edge weight gives the amplitude associated with the photons <ref:2609. p. 1

Where the paper says it

Quantum Optics Experiments as Graphs Photonic quantum experiments can be encoded in weighted colored graphs G(V, E), where the graph nodes (V ) correspond to spatial photon paths and the edges (E) between them encode the mode number of the photons through a specific color. The edge weight gives the amplitude associated with the photons.

p. 1
Where the paper says it

Quantum Optics Experiments as Graphs Photonic quantum experiments can be encoded in weighted colored graphs G(V, E), where the graph nodes (V ) correspond to spatial photon paths and the edges (E) between them encode the mode number of the photons through a specific color. The edge weight gives the amplitude associated with the photons.

08073#pg2> For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively <ref:2609. p. 1
Where the paper says it

The edge weight gives the amplitude associated with the photons. For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively.

p. 1
Where the paper says it

The edge weight gives the amplitude associated with the photons. For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively.

08073#pg2> This can be translated into the operator w ki,kj a† ka b† + h. p. 1
Where the paper says it

For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively. This can be translated into the operator w ki,kj a,b a † ka b † kb + h.

p. 1
Where the paper says it

For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively. This can be translated into the operator w ki,kj a,b a † ka b † kb + h.

c. with a† ka the creation operator of a photon on path a with mode ka <ref:2609. p. 1
Where the paper says it

This can be translated into the operator w ki,kj a,b a † ka b † kb + h. with a † ka the creation operator of a photon on path a with mode ka.

p. 1
Where the paper says it

This can be translated into the operator w ki,kj a,b a † ka b † kb + h. with a † ka the creation operator of a photon on path a with mode ka.

08073#pg2> The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function Φ(w) = X m 1 m! p. 1
Where the paper says it

with a † ka the creation operator of a photon on path a with mode ka. The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function [3] Φ(w) = X m 1 m!

p. 1
Where the paper says it

with a † ka the creation operator of a photon on path a with mode ka. The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function [3] Φ(w) = X m 1 m!

(X e∈E(G) w(e)x† (e)y† (e) + h.c. <ref:2609.08073#pg2>

Photonic quantum experiments can be encoded in weighted colored graphs G(V, E), where the graph nodes (V) correspond to spatial photon paths and the edges (E) between them encode the mode number of the photons through a specific color <ref:2609.08073#pg2> The edge weight gives the amplitude associated with the photons <ref:2609.08073#pg2> For example, the weight w ka,kb a,b is encoded as the edge between photonic paths a and b with modes/colors ka and kb respectively <ref:2609.08073#pg2> This can be translated into the operator w ki,kj a† ka b† + h.c. with a† ka the creation operator of a photon on path a with mode ka <ref:2609.08073#pg2> The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function Φ(w) = X m 1 m! (X e∈E(G) w(e)x† (e)y† (e) + h.c. <ref:2609.08073#pg2>

The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function Φ(w) = X m 1 m! (X e∈E(G) w(e)x† (e)y† (e) + h.c. <ref:2609.08073#pg2>

The connection between the multi-graph and the resulting quantum state from the optical setup is given by a weight function Φ(w) = X m 1 m! (X e∈E(G) w(e)x† (e)y† (e) + h.c. <ref:2609.08073#pg2>

--- Page 3 ---

3

B. Grinch

A prominent class of graph-based inverse-design methods formulates the circuit topology and its continuous edge weights as a coupled discrete–continuous optimization problem (Ge, w∗) ∈ arg min G,w [1 − F(G, w)], (2) where F(G, w) is the fidelity between the generated and target states <ref:2609.08073#pg2> Top-down methods begin with a dense graph and use optimized weights, often together with L1 regularization, to identify edges to remove <ref:2609. p. 1

Where the paper says it

Grinch A prominent class of graph-based inverse-design methods formulates the circuit topology and its continuous edge weights as a coupled discrete–continuous optimization problem [2, 3, 13], (Ge, w* ) ∈ arg min G,w [1 − F(G, w)] , (2) where F(G, w) is the fidelity between the generated and target states, while Ge and w* denote an optimal topology (discrete) and its corresponding edge weights (continuous). Top-down methods begin with a dense graph and use optimized weights, often together with L1 regularization, to identify edges to remove [2, 3].

p. 1
Where the paper says it

Grinch A prominent class of graph-based inverse-design methods formulates the circuit topology and its continuous edge weights as a coupled discrete–continuous optimization problem [2, 3, 13], (Ge, w* ) ∈ arg min G,w [1 − F(G, w)] , (2) where F(G, w) is the fidelity between the generated and target states, while Ge and w* denote an optimal topology (discrete) and its corresponding edge weights (continuous). Top-down methods begin with a dense graph and use optimized weights, often together with L1 regularization, to identify edges to remove [2, 3].

08073#pg2> In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G) <ref:2609. p. 1
Where the paper says it

This motivates bottom-up construction as an alternative framework that may require fewer structural operations during the search. In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G).

p. 1
Where the paper says it

This motivates bottom-up construction as an alternative framework that may require fewer structural operations during the search. In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G).

p. 1
Where the paper says it

This motivates bottom-up construction as an alternative framework that may require fewer structural operations during the search. In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G).

08073#pg2>

A prominent class of graph-based inverse-design methods formulates the circuit topology and its continuous edge weights as a coupled discrete–continuous optimization problem (Ge, w∗) ∈ arg min G,w [1 − F(G, w)], (2) where F(G, w) is the fidelity between the generated and target states <ref:2609.08073#pg2> In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G) <ref:2609.08073#pg2>

GFlowNets are designed to construct compositional objects through sequences of stochastic actions (policy) <ref:2609.08073#pg2> The forward and backward policies, PF and PB, are trained so that terminal graphs are sampled in proportion to their rewards, Eq. p. 2

Where the paper says it

For a target expansion with K nonzero computational-basis terms on N nodes, including ancillas, the quantity NK/2 serves as a heuristic reference for selecting the edge budget. The forward and backward policies, PF and PB, are trained so that terminal graphs are sampled in proportion to their rewards, Eq.

3 <ref:2609. p. 0
Where the paper says it

[4], Grinch addresses experimental-design problems involving both discrete structures and coupled discrete–continuous variables: the circuit topology is sampled discretely, while its continuous edge weights are optimized to maximize the reward. arXiv:2609.

p. 0
Where the paper says it

[4], Grinch addresses experimental-design problems involving both discrete structures and coupled discrete–continuous variables: the circuit topology is sampled discretely, while its continuous edge weights are optimized to maximize the reward. arXiv:2609.

p. 0
Where the paper says it

[4], Grinch addresses experimental-design problems involving both discrete structures and coupled discrete–continuous variables: the circuit topology is sampled discretely, while its continuous edge weights are optimized to maximize the reward. arXiv:2609.

p. 0
Where the paper says it

[4], Grinch addresses experimental-design problems involving both discrete structures and coupled discrete–continuous variables: the circuit topology is sampled discretely, while its continuous edge weights are optimized to maximize the reward. arXiv:2609.

08073#pg2> Minimizing Eq. p. 2
Where the paper says it

(4) where Zθ is a learnable estimate of the partition function and R(x) is the reward of the terminal object. Minimizing Eq.

4 balances the probability of each forward construction trajectory with the reward-weighted probability of its reverse trajectory <ref:2609. p. 2
Where the paper says it

Minimizing Eq. 4 balances the probability of each forward construction trajectory with the reward-weighted probability of its reverse trajectory, allowing different trajectories to contribute to the same terminal graph.

08073#pg2>

The forward and backward policies, PF and PB, are trained so that terminal graphs are sampled in proportion to their rewards, Eq. 3 <ref:2609.08073#pg2> This sampling formulation is motivated by the existence of several inequivalent topologies with comparable fidelity <ref:2609. p. 1

Where the paper says it

For quantum circuit design, the reward can be the fidelity or a combination of experimental metrics, while the continuous edge weights can still be optimized using standard methods. This sampling formulation is motivated by the existence of several inequivalent topologies with comparable fidelity, whereas Eq.

08073#pg2> When multiple arrays are possible, each with potential hardware differences, GFlowNets can effectively explore these regions and identify multiple high-reward modes <ref:2609. p. 4
Where the paper says it

These examples illustrate how diverse sampling can be beneficial. When multiple arrays are possible, each with potential hardware differences, GFlowNets can effectively explore these regions and identify multiple high-reward modes.

08073#pg2>

The forward and backward policies, PF and PB, are trained so that terminal graphs are sampled in proportion to their rewards, Eq. 3 <ref:2609.08073#pg2> Grinch provides a complementary approach to these limitations by learning directly from a state-dependent reward rather than reconstructing a distribution over archived experiments <ref:2609. p. 6

Where the paper says it

The QOVAE study also notes that incorporating continuously parametrized devices requires discretization and that generated experiments may remain challenging to implement in the laboratory. Grinch provides a complementary approach to these limitations by learning directly from a state-dependent reward rather than reconstructing a distribution over archived experiments.

p. 6
Where the paper says it

The QOVAE study also notes that incorporating continuously parametrized devices requires discretization and that generated experiments may remain challenging to implement in the laboratory. Grinch provides a complementary approach to these limitations by learning directly from a state-dependent reward rather than reconstructing a distribution over archived experiments.

08073#pg2>

The framework introduces Grinch to address the challenge of designing quantum optical experiments by constructing circuit topologies edge by edge and learning directly from reward evaluations without needing a preexisting training dataset <ref:2609.08073#pg2> This bottom-up approach contrasts with top-down methods that optimize and prune a single dense graph, as Grinch constructs circuit topologies edge by edge <ref:2609.08073#pg2> The framework utilizes GraphFlowNets to sample circuit graphs from a reward-proportional distribution where Pθ(G) is proportional to the reward R(G) <ref:2609.08073#pg2>

--- Page 4 ---

4

B. Grinch

A prominent class of graph-based inverse-design methods formulates the circuit topology and its continuous edge weights as a coupled discrete–continuous optimization problem (Ge, w∗) ∈ arg min G,w [1 − F(G, w)], (2) where F(G, w) is the fidelity between the generated and target states <ref:2609.08073#pg2> Top-down methods begin with a dense graph and use optimized weights, often together with L1 regularization, to identify edges to remove <ref:2609.08073#pg2> In Grinch, we recast the discrete search as sampling circuit graphs from a reward-proportional distribution, G ∼ Pθ(G) where Pθ(G) ∝ R(G) <ref:2609.08073#pg2>

The forward and backward policies, PF and PB are trained so that terminal graphs are sampled in proportion to their rewards, Eq. 3 <ref:2609.08073#pg2> Grinch provides a complementary approach to these limitations by learning directly from a state-dependent reward rather than reconstructing a distribution over archived experiments <ref:2609.08073#pg2>

--- Page 5 ---

5

C. Nonlocal photonic gates

Entangling photonic gates are building blocks for optical quantum computation <ref:2609.08073#pg2> Extending such operations to spatially separated systems is relevant to distributed quantum processing, in which nonlocal gates connect distinct nodes in quantum networks <ref:2609. p. 5

Where the paper says it

Linear-optical schemes can realize effective interactions through interference, ancillary photons, and conditional measurements, avoiding the need for strong direct photon–photon coupling [34]. Extending such operations to spatially separated systems is relevant to distributed quantum processing, in which nonlocal gates connect distinct nodes in quantum networks [35].

08073#pg2> The synthesis of CNOT, CZ, and multiply controlled transformations therefore extends the design task from preparing entangled resources to reproducing coherent logical operations <ref:2609. p. 5
Where the paper says it

Extending such operations to spatially separated systems is relevant to distributed quantum processing, in which nonlocal gates connect distinct nodes in quantum networks [35]. The synthesis of CNOT, CZ, and multiply controlled transformations therefore extends the design task from preparing entangled resources to reproducing coherent logical operations.

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. Reward-driven generative framework for quantum optical experiments

The Grinch framework learns to sample optical graphs directly from fidelity-based rewards without relying on a pre-existing training dataset, allowing it to construct circuit topologies edge by edge and learn directly from reward evaluations.

  1. Handling coupled discrete and continuous design problems

The system addresses experimental-design problems involving both discrete structures and coupled discrete–continuous variables: the circuit topology is sampled discretely, while its continuous edge weights are optimized to maximize the reward.

  1. Exploring alternative circuit topologies via diversity sampling

Grinch can recover diverse circuit realizations rather than a single optimum, as shown by how different trajectories with distinct final states can lead to the same pruned state as shown in the figure, enabling exploration of regions where several distinct circuits may generate the same target state.

  1. Incorporating hardware constraints directly into the search

The framework allows for incorporating connectivity constraints into the search process by augmenting actions and applying structural masks, which are defined as a mechanism to restrict the search space, increasing the efficiency of the algorithm by preferring valid and meaningful states.

  1. Identifying asymptotic solutions for complex states

Grinch can identify asymptotic solutions for states that cannot be generated exactly by graphs, which is crucial for preparing targets like SRV (5, 4, 4)∗⟩ where high-fidelity solutions approach unity within a certain tolerance.

Abstract

Designing quantum optical experiments requires searching over discrete circuit topologies and continuous parameters, often leading to multiple possible realizations of the same target state. Graph-based methods commonly address this problem by optimizing a dense graph and pruning it toward a single circuit. We introduce Grinch, a bottom-up reward-driven generative framework that learns to sample optical graphs directly from fidelity-based rewards without relying on a pre-existing training dataset. We demonstrate the framework on multipartite entangled states, graph and cluster states, and nonlocal photonic-gate targets, obtaining multiple high-fidelity optical graphs. For nonlocal photonic gates, we obtain Toffoli constructions requiring two ancillas and alternative CNOT and Toffoli designs when connections between ancilla nodes are restricted. For state preparation, we identify asymptotic solutions for targets that cannot be generated exactly by graphs. Using the discovered topologies, we derive analytic families of optical experiments with explicit interference mechanisms and fidelity--rate relations, showing how generative design can support interpretable scientific discovery. These findings illustrate a reward-driven approach to quantum optical inverse design, exploring alternative circuit topologies while directly incorporating connectivity constraints into the search process.

Sources

Related papers