Diagonal Flow Matching for Inverse Design

• The paper introduces Diag-CFM, a zero-anchored flow matching method that pairs design coordinates with noise and labels with zero to eliminate sensitivity to coordinate ordering and scaling.
• It leverages permutation-equivariant training to ensure robust, bidirectional synthesis and analysis in inverse design tasks.
• Empirical results show Diag-CFM achieves over 10× improvement in round-trip error compared to standard methods across scalable benchmarks.

to=functions.shell wuregistry 天天中彩票被{"command":"bash -lc 'python - <<\"PY\"\nimport urllib.parse, urllib.request\nqueries=[\"Diagonal Flow Matching\", \"Flow Matching for Generative Modeling\", \"Generator Matching flow matching diffusion\"]\nfor q in queries:\n url='http://export.arxiv.org/api/query?search_query=all:'+urllib.parse.quote(q)+'&start=0&max_results=3'\n print(\"QUERY\", q)\n data=urllib.request.urlopen(url, timeout=20).read().decode('utf-8','ignore')\n print(data[:1200].replace('\n',' ')[:1200])\n print('\n---\n')\nPY'"}＿奇米影视 to=functions.shell 开号地址{"output":"QUERY Diagonal Flow Matching\n<?xml version=\"1.0\" encoding=\"UTF-8\"?> <feed xmlns=\"http://www.w3.org/2005/Atom\"> <link href=\"http://arxiv.org/api/query?search_query%3Dall%3ADiagonal%2520Flow%2520Matching%26id_list%3D%26start%3D0%26max_results%3D3\" rel=\"self\" type=\"application/atom+xml\"/> <title type=\"html\">ArXiv Query: search_query=all:Diagonal%20Flow%20Matching&id_list=&start=0&max_results=3</title> <id>http://arxiv.org/api/cHxN44Xca9Ty9Eb4nuo7vhslq8A</id> <updated\>2026-07-05T00:00:00-04:00</updated> <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>1/opensearch:totalResults <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>3/opensearch:itemsPerPage <entry> <id>http://arxiv.org/abs/([2603.15925](/papers/2603.15925))v1</id> <updated\>2026-03-20T21:38:32Z</updated> <published\>2026-03-20T21:38:32Z</published> <title>Generative Inverse Design with Abstention via Diagonal Flow Matching</title>\n---\n\nQUERY Flow Matching for Generative Modeling\n<?xml version=\"1.0\" encoding=\"UTF-8\"?> <feed xmlns=\"http://www.w3.org/2005/Atom\"> <link href=\"http://arxiv.org/api/query?search_query=all:Flow%20Matching%20for%20Generative%20Modeling&id_list=&start=0&max_results=3\" rel=\"self\" type=\"application/atom+xml\"/> <title type=\"html\">ArXiv Query: search_query=all:Flow Matching for Generative Modeling&id_list=&start=0&max_results=3</title> <id>http://arxiv.org/api/DgAtv2mhhzBikjvqSWAF+E8R3XA</id> <updated\>2026-07-05T00:00:00-04:00</updated> <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>157040/opensearch:totalResults <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>3/opensearch:itemsPerPage <entry> <id>http://arxiv.org/abs/([2210.02747](/papers/2210.02747))v1</id> <updated\>2022-10-06T17:59:16Z</updated> <published\>2022-10-06T17:59:16Z</published> <title>Flow Matching for Generative Modeling</title> <summary> Flow Matching (FM) is a family of simulation-f\n---\n\nQUERY Generator Matching flow matching diffusion\n<?xml version=\"1.0\" encoding=\"UTF-8\"?> <feed xmlns=\"http://www.w3.org/2005/Atom\"> <link href=\"http://arxiv.org/api/query?search_query=all:Generator%20Matching%20flow%20matching%20diffusion&id_list=&start=0&max_results=3\" rel=\"self\" type=\"application/atom+xml\"/> <title type=\"html\">ArXiv Query: search_query=all:Generator Matching flow matching diffusion&id_list=&start=&max_results=3</title> <id>http://arxiv.org/api/m4iQkp4B+xp0AsG21go41reJI44</id> <updated\>2026-07-05T00:00:00-04:00</updated> <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>321989/opensearch:totalResults <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>3/opensearch:itemsPerPage <entry> <id>http://arxiv.org/abs/([1504.00088](/papers/1504.00088))v1</id> <updated\>2015-03-31T18:00:00Z</updated> <published\>2015-03-31T18:00:00Z</published> <title>Matching while Learning</title> <summary> \n---\n","metadata":{"exit_code":0,"duration_seconds":1.98}} RTLU to=functions.shell 天天彩票app code='{"command":"bash -lc "PY\"\nimport urllib.parse, urllib.request, re\nq=\"(Lipman et al., 2022)\"\nurl=\"http://export.arxiv.org/api/query?id_list=\"+q\nprint(urllib.request.urlopen(url).read().decode()[:800])\nPY"}' ／久久 to=functions.shell атәыла{"output":"<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n<feed xmlns=\"http://www.w3.org/2005/Atom\">\n <link href=\"http://arxiv.org/api/query?search_query=&id_list=([2210.02747](/papers/2210.02747))&start=0&max_results=10\" rel=\"self\" type=\"application/atom+xml\"/>\n <title type=\"html\">ArXiv Query: search_query=&id_list=(Lipman et al., 2022)&start=0&max_results=10</title>\n <id>http://arxiv.org/api/q0wyKR29jlG9gH8ob2+vpSgQe1s</id>\n <updated\>2026-07-05T00:00:00-04:00</updated>\n <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>1/opensearch:totalResults\n <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex\n <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>10/opensearch:itemsPerPage\n","metadata":{"exit_code":0,"duration_seconds":0.47}} Diagonal Flow Matching (Diag-CFM) is a variant of conditional flow matching for generative inverse design that defines a continuous-time, invertible flow in an augmented space $[z; y] \leftrightarrow [x; 0]$ and trains a neural velocity field against the diagonal target $[x-z; -y]$. In the formulation introduced for inverse design, the method uses a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, with the stated goals of removing accidental dependence on coordinate ordering and scaling while preserving bidirectional generation and prediction (Campos et al., 16 Mar 2026). Within the broader Generator Matching view, diagonal structure refers more generally to imposing coordinate-wise restrictions on drift $u_t$ or diagonal diffusion $a_t$, which provides the main conceptual bridge between Diag-CFM and the wider diffusion/flow matching landscape (Patel et al., 2024).

1. Inverse-design setting and the limitation of standard conditional flow matching

Diag-CFM is formulated for parametric engineering design problems with design parameters $x \in \mathbb{R}^P$ and performance labels $y \in \mathbb{R}^L$, typically with $L < P$. The forward relation is written as

$y = f(x) + \varepsilon,$

where $f : \mathbb{R}^P \to \mathbb{R}^L$ is a simulator or surrogate and $\varepsilon$ is noise. The feasible label set is

$[x-z; -y]$0

and, for a target $[x-z; -y]$1 and tolerance $[x-z; -y]$2, the success set is

$[x-z; -y]$3

The objective is to sample diverse designs $[x-z; -y]$4 that satisfy $[x-z; -y]$5, ideally covering the possibly multi-modal success set rather than returning only a single optimizer output (Campos et al., 16 Mar 2026).

Standard conditional flow matching for this setting works in an augmented $[x-z; -y]$6-dimensional space. It uses a latent $[x-z; -y]$7, forms

$[x-z; -y]$8

and linearly interpolates

$[x-z; -y]$9

The corresponding target velocity is

$u_t$0

The paper identifies the central weakness of this construction: different coordinates are matched against different types of quantities, namely “design vs noise” in some coordinates and “design vs label” in others. Because the loss is component-wise, the approximation difficulty depends on the arbitrary decomposition into the first $u_t$1 and last $u_t$2 coordinates. This introduces sensitivity to coordinate ordering and scaling, and the paper reports that standard CFM can be unstable and significantly less accurate than Diag-CFM across all three benchmarks considered (Campos et al., 16 Mar 2026).

2. Zero anchoring and the diagonal construction

Diag-CFM changes the source–target pairing rather than the basic flow-matching recipe. It operates in an augmented space of dimension $u_t$3, samples noise $u_t$4, and defines

$u_t$5

The linear path is

$u_t$6

with target velocity

$u_t$7

This is the sense in which the method is “diagonal”: each design coordinate is paired only with a noise coordinate, and each label coordinate is paired only with zero (Campos et al., 16 Mar 2026).

Construction	Standard CFM	Diag-CFM
Source state	$u_t$8	$u_t$9
Target state	$a_t$0	$a_t$1
Target velocity	$a_t$2	$a_t$3

The training objective is standard flow-matching regression applied to this new state design: $a_t$4 The velocity field is a neural network

$a_t$5

typically implemented in the paper as an MLP. At convergence, integrating

$a_t$6

from $a_t$7 to $a_t$8 approximately transports the distribution of $a_t$9 to the empirical distribution of $x \in \mathbb{R}^P$0 (Campos et al., 16 Mar 2026).

The practical consequence of zero anchoring is that labels are no longer compared directly against a subset of design coordinates. The paper states that this resolves the strong sensitivity of inverse-design CFM to arbitrary ordering and scaling, while still supporting bidirectional use through the same ODE (Campos et al., 16 Mar 2026).

3. Permutation equivariance, robustness, and what “diagonal” means

The main theoretical claim for Diag-CFM is permutation robustness at the level of the target-velocity distribution. Let

$x \in \mathbb{R}^P$1

where $x \in \mathbb{R}^P$2 and $x \in \mathbb{R}^P$3. For permutation matrices $x \in \mathbb{R}^P$4 on design coordinates and $x \in \mathbb{R}^P$5 on label coordinates, define

$x \in \mathbb{R}^P$6

Proposition 3.1 states that the distribution of $x \in \mathbb{R}^P$7 is equivariant under permutations of parameter and label coordinates. Because the entries of $x \in \mathbb{R}^P$8 are i.i.d., permuting coordinates preserves the target-velocity statistics. By contrast, Proposition 3.2 states that the standard CFM target velocity is not equivariant under coordinate permutations, so the learning problem itself depends on arbitrary coordinate order (Campos et al., 16 Mar 2026).

The paper connects this directly to observed training behavior. On the gas turbine dataset, which is normalized to $x \in \mathbb{R}^P$9, simply reordering the parameter coordinates changes standard CFM’s final round-trip error by nearly one order of magnitude, whereas Diag-CFM’s error is reported as essentially constant across orderings. The same section also attributes improved run-to-run stability to the diagonal construction (Campos et al., 16 Mar 2026).

In the broader Generator Matching framework, “diagonal” has a wider meaning than the zero-anchored inverse-design construction. A generator on $y \in \mathbb{R}^L$0 can be written as

$y \in \mathbb{R}^L$1

and a diagonal structure can be imposed either through coordinate-wise drift $y \in \mathbb{R}^L$2 or through a diagonal diffusion matrix $y \in \mathbb{R}^L$3. The Generator Matching analysis therefore treats diagonality as a structural restriction on the generator family, whereas the Diag-CFM inverse-design paper realizes diagonality through the specific augmented-state pairing $y \in \mathbb{R}^L$4 and the target $y \in \mathbb{R}^L$5 (Patel et al., 2024).

A common misconception is to identify all coordinate-wise linear flows with Diag-CFM. The literature summarized here does not support that equivalence. The inverse-design method is a specific zero-anchored construction; related linear-path methods may be only conceptually aligned with diagonal designs rather than instances of Diag-CFM proper (Campos et al., 16 Mar 2026).

4. Bidirectional flow, abstention, and intrinsic uncertainty

Diag-CFM is bidirectional because the same learned ODE is used in both synthesis and analysis. For inverse generation, given a target $y \in \mathbb{R}^L$6, one samples $y \in \mathbb{R}^L$7, initializes

$y \in \mathbb{R}^L$8

integrates

$y \in \mathbb{R}^L$9

and outputs the design

$L < P$0

For forward prediction, given a design $L < P$1, one sets

$L < P$2

integrates backward with sign reversal,

$L < P$3

and reads off

$L < P$4

There is no separate forward and inverse network; both directions share the same $L < P$5 (Campos et al., 16 Mar 2026).

A distinctive aspect of the method is that the zero-anchored architecture exposes two model-intrinsic uncertainty metrics. The first is Zero-Deviation. If synthesis produces

$L < P$6

then

$L < P$7

Since the target final state for label coordinates is exactly zero, $L < P$8 is intended to be small for feasible in-distribution targets and larger for unrealistic or out-of-distribution targets. The second metric is Self-Consistency. After synthesis, one replaces the residual label coordinates by exact zeros, integrates backward from

$L < P$9

obtains

$y = f(x) + \varepsilon,$0

and defines

$y = f(x) + \varepsilon,$1

The paper emphasizes that this differs from a trivial round trip because the nonzero residual $y = f(x) + \varepsilon,$2 is discarded before the backward pass (Campos et al., 16 Mar 2026).

These two metrics are used for three tasks: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets. The reported gains are task-wide rather than anecdotal. Candidate selection improves over random choice by $y = f(x) + \varepsilon,$3 to $y = f(x) + \varepsilon,$4; rejecting the top $y = f(x) + \varepsilon,$5 most uncertain samples reduces mean error by $y = f(x) + \varepsilon,$6 to $y = f(x) + \varepsilon,$7; and Zero-Deviation in particular achieves strong ROC AUC values, including $y = f(x) + \varepsilon,$8 on Unifoil, $y = f(x) + \varepsilon,$9 on gas turbine, and $f : \mathbb{R}^P \to \mathbb{R}^L$0–$f : \mathbb{R}^P \to \mathbb{R}^L$1 on DTLZ (Campos et al., 16 Mar 2026).

5. Benchmarks, metrics, and observed behavior

The Diag-CFM study evaluates three inverse-design benchmarks: a gas turbine combustor dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$2 and $f : \mathbb{R}^P \to \mathbb{R}^L$3, a Unifoil airfoil dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$4 and $f : \mathbb{R}^P \to \mathbb{R}^L$5 plus two physical conditioning scalars, and the analytical DTLZ2 benchmark with $f : \mathbb{R}^P \to \mathbb{R}^L$6 up to $f : \mathbb{R}^P \to \mathbb{R}^L$7 and $f : \mathbb{R}^P \to \mathbb{R}^L$8. The central empirical metric for inverse quality is round-trip error, namely the error incurred by generating a design for a target and then evaluating its achieved performance. Forward MSE is also reported, together with diversity analyses and uncertainty benchmarks (Campos et al., 16 Mar 2026).

On the gas turbine benchmark, Diag-CFM attains forward MSE $f : \mathbb{R}^P \to \mathbb{R}^L$9, compared with $\varepsilon$0 for the INN baseline and $\varepsilon$1 for standard CFM. Its round-trip error is $\varepsilon$2, compared with $\varepsilon$3 for INN and $\varepsilon$4 for CFM, which the paper summarizes as more than a $\varepsilon$5 improvement. On Unifoil, Diag-CFM achieves forward MSE $\varepsilon$6 versus $\varepsilon$7 for INN and $\varepsilon$8 for CFM, and round-trip error $\varepsilon$9 versus $[x-z; -y]$00 and $[x-z; -y]$01, respectively (Campos et al., 16 Mar 2026).

The DTLZ2 results are particularly important for scalability. At $[x-z; -y]$02, the reported round-trip errors are $[x-z; -y]$03 for Diag-CFM, $[x-z; -y]$04 for CFM, and $[x-z; -y]$05 for INN. At $[x-z; -y]$06, they are $[x-z; -y]$07, $[x-z; -y]$08, and $[x-z; -y]$09, respectively. The paper therefore characterizes Diag-CFM as maintaining round-trip error in the $[x-z; -y]$10–$[x-z; -y]$11 range across design dimensions up to $[x-z; -y]$12, while CFM and INN degrade more sharply (Campos et al., 16 Mar 2026).

The implementation used for these results is deliberately simple. The velocity network is an MLP with time $[x-z; -y]$13 concatenated directly to the state, and inference uses explicit Euler with $[x-z; -y]$14 uniform steps on $[x-z; -y]$15. The paper presents this as evidence that the empirical gains arise primarily from the flow construction rather than from specialized architectural devices (Campos et al., 16 Mar 2026).

6. Relation to flow matching, guided variants, and Generator Matching

Flow matching more generally is a continuous-time alternative to diffusion in which a vector field defines an ODE transporting a simple base distribution to a target distribution, and training regresses the learned velocity onto a known conditional velocity along a prescribed probability path (Lipman et al., 2022). The Generator Matching framework sharpens this relation by showing that both flow matching and diffusion can be written as Markov processes with generators $[x-z; -y]$16, with pure flow matching corresponding to a first-order operator and diffusion corresponding to a second-order operator with nonzero diffusion matrix. In that view, the empirical robustness of flow matching is linked to the first-order transport PDE and the absence of a smoothing operator that must later be inverted (Patel et al., 2024).

This contextualization matters for Diag-CFM. The inverse-design method remains a flow-matching model, and therefore inherits the first-order ODE character of standard flow matching. The Generator Matching paper further notes that a diagonal flow matching variant can be seen as imposing additional structure on either $[x-z; -y]$17, for example coordinate-wise decoupled drift, or on the diffusion term $[x-z; -y]$18, for example a diagonal diffusion matrix. This suggests a broader family of diagonal or partially diagonal generator constructions beyond the specific zero-anchored inverse-design formulation (Patel et al., 2024).

A related but distinct development is guided flow matching for channel knowledge map construction. That method uses a deterministic ODE,

$[x-z; -y]$19

with the linear interpolation

$[x-z; -y]$20

and constant target velocity

$[x-z; -y]$21

The paper states that it does not explicitly mention the term “Diagonal Flow Matching,” but it describes the resulting structure as reminiscent of DFM or diagonal flows because each component evolves along a coordinate-wise linear path and the analytic target velocity contains no explicit cross-coordinate coupling (Huang et al., 6 Jan 2026).

Taken together, these papers place Diag-CFM at a precise intersection of ideas. It is not merely “flow matching with a simple path,” nor merely “a model with diagonal diffusion.” It is a specific zero-anchored conditional flow-matching construction for inverse design, whose defining property is the target $[x-z; -y]$22, and whose theoretical and empirical significance lies in permutation-equivariant training targets, stable round-trip inverse design, and architecture-intrinsic abstention signals (Campos et al., 16 Mar 2026).python - <<\"PY\"\nimport urllib.parse, urllib.request, re\nq=\"(Lipman et al., 2022)\"\nurl=\"http://export.arxiv.org/api/query?id_list=\"^^^^46^^^^"}' ／久久 to=functions.shell атәыла{"output":"<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n<feed xmlns=\"http://www.w3.org/2005/Atom\">\n <link href=\"http://arxiv.org/api/query?search_query=&id_list=([2210.02747](/papers/2210.02747))&start=0&max_results=10\" rel=\"self\" type=\"application/atom+xml\"/>\n <title type=\"html\">ArXiv Query: search_query=&id_list=(Lipman et al., 2022)&start=0&max_results=10</title>\n <id>http://arxiv.org/api/q0wyKR29jlG9gH8ob2+vpSgQe1s</id>\n <updated\>2026-07-05T00:00:00-04:00</updated>\n <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>1/opensearch:totalResults\n <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex\n <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>10/opensearch:itemsPerPage\n","metadata":{"exit_code":0,"duration_seconds":0.47}} Diagonal Flow Matching (Diag-CFM) is a variant of conditional flow matching for generative inverse design that defines a continuous-time, invertible flow in an augmented space $[z; y] \leftrightarrow [x; 0]$ and trains a neural velocity field against the diagonal target $[x-z; -y]$. In the formulation introduced for inverse design, the method uses a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, with the stated goals of removing accidental dependence on coordinate ordering and scaling while preserving bidirectional generation and prediction (Campos et al., 16 Mar 2026). Within the broader Generator Matching view, diagonal structure refers more generally to imposing coordinate-wise restrictions on drift $u_t$ or diagonal diffusion $a_t$, which provides the main conceptual bridge between Diag-CFM and the wider diffusion/flow matching landscape (Patel et al., 2024).

1. Inverse-design setting and the limitation of standard conditional flow matching

Diag-CFM is formulated for parametric engineering design problems with design parameters $x \in \mathbb{R}^P$ and performance labels $y \in \mathbb{R}^L$, typically with $L < P$. The forward relation is written as

$y = f(x) + \varepsilon,$

where $f : \mathbb{R}^P \to \mathbb{R}^L$ is a simulator or surrogate and $\varepsilon$ is noise. The feasible label set is

$[x-z; -y]$0

and, for a target $[x-z; -y]$1 and tolerance $[x-z; -y]$2, the success set is

$[x-z; -y]$3

The objective is to sample diverse designs $[x-z; -y]$4 that satisfy $[x-z; -y]$5, ideally covering the possibly multi-modal success set rather than returning only a single optimizer output (Campos et al., 16 Mar 2026).

Standard conditional flow matching for this setting works in an augmented $[x-z; -y]$6-dimensional space. It uses a latent $[x-z; -y]$7, forms

$[x-z; -y]$8

and linearly interpolates

$[x-z; -y]$9

The corresponding target velocity is

$u_t$0

The paper identifies the central weakness of this construction: different coordinates are matched against different types of quantities, namely “design vs noise” in some coordinates and “design vs label” in others. Because the loss is component-wise, the approximation difficulty depends on the arbitrary decomposition into the first $u_t$1 and last $u_t$2 coordinates. This introduces sensitivity to coordinate ordering and scaling, and the paper reports that standard CFM can be unstable and significantly less accurate than Diag-CFM across all three benchmarks considered (Campos et al., 16 Mar 2026).

2. Zero anchoring and the diagonal construction

Diag-CFM changes the source–target pairing rather than the basic flow-matching recipe. It operates in an augmented space of dimension $u_t$3, samples noise $u_t$4, and defines

$u_t$5

The linear path is

$u_t$6

with target velocity

$u_t$7

This is the sense in which the method is “diagonal”: each design coordinate is paired only with a noise coordinate, and each label coordinate is paired only with zero (Campos et al., 16 Mar 2026).

Construction	Standard CFM	Diag-CFM
Source state	$u_t$8	$u_t$9
Target state	$a_t$0	$a_t$1
Target velocity	$a_t$2	$a_t$3

The training objective is standard flow-matching regression applied to this new state design: $a_t$4 The velocity field is a neural network

$a_t$5

typically implemented in the paper as an MLP. At convergence, integrating

$a_t$6

from $a_t$7 to $a_t$8 approximately transports the distribution of $a_t$9 to the empirical distribution of $x \in \mathbb{R}^P$0 (Campos et al., 16 Mar 2026).

The practical consequence of zero anchoring is that labels are no longer compared directly against a subset of design coordinates. The paper states that this resolves the strong sensitivity of inverse-design CFM to arbitrary ordering and scaling, while still supporting bidirectional use through the same ODE (Campos et al., 16 Mar 2026).

3. Permutation equivariance, robustness, and what “diagonal” means

The main theoretical claim for Diag-CFM is permutation robustness at the level of the target-velocity distribution. Let

$x \in \mathbb{R}^P$1

where $x \in \mathbb{R}^P$2 and $x \in \mathbb{R}^P$3. For permutation matrices $x \in \mathbb{R}^P$4 on design coordinates and $x \in \mathbb{R}^P$5 on label coordinates, define

$x \in \mathbb{R}^P$6

Proposition 3.1 states that the distribution of $x \in \mathbb{R}^P$7 is equivariant under permutations of parameter and label coordinates. Because the entries of $x \in \mathbb{R}^P$8 are i.i.d., permuting coordinates preserves the target-velocity statistics. By contrast, Proposition 3.2 states that the standard CFM target velocity is not equivariant under coordinate permutations, so the learning problem itself depends on arbitrary coordinate order (Campos et al., 16 Mar 2026).

The paper connects this directly to observed training behavior. On the gas turbine dataset, which is normalized to $x \in \mathbb{R}^P$9, simply reordering the parameter coordinates changes standard CFM’s final round-trip error by nearly one order of magnitude, whereas Diag-CFM’s error is reported as essentially constant across orderings. The same section also attributes improved run-to-run stability to the diagonal construction (Campos et al., 16 Mar 2026).

In the broader Generator Matching framework, “diagonal” has a wider meaning than the zero-anchored inverse-design construction. A generator on $y \in \mathbb{R}^L$0 can be written as

$y \in \mathbb{R}^L$1

and a diagonal structure can be imposed either through coordinate-wise drift $y \in \mathbb{R}^L$2 or through a diagonal diffusion matrix $y \in \mathbb{R}^L$3. The Generator Matching analysis therefore treats diagonality as a structural restriction on the generator family, whereas the Diag-CFM inverse-design paper realizes diagonality through the specific augmented-state pairing $y \in \mathbb{R}^L$4 and the target $y \in \mathbb{R}^L$5 (Patel et al., 2024).

A common misconception is to identify all coordinate-wise linear flows with Diag-CFM. The literature summarized here does not support that equivalence. The inverse-design method is a specific zero-anchored construction; related linear-path methods may be only conceptually aligned with diagonal designs rather than instances of Diag-CFM proper (Campos et al., 16 Mar 2026).

4. Bidirectional flow, abstention, and intrinsic uncertainty

Diag-CFM is bidirectional because the same learned ODE is used in both synthesis and analysis. For inverse generation, given a target $y \in \mathbb{R}^L$6, one samples $y \in \mathbb{R}^L$7, initializes

$y \in \mathbb{R}^L$8

integrates

$y \in \mathbb{R}^L$9

and outputs the design

$L < P$0

For forward prediction, given a design $L < P$1, one sets

$L < P$2

integrates backward with sign reversal,

$L < P$3

and reads off

$L < P$4

There is no separate forward and inverse network; both directions share the same $L < P$5 (Campos et al., 16 Mar 2026).

A distinctive aspect of the method is that the zero-anchored architecture exposes two model-intrinsic uncertainty metrics. The first is Zero-Deviation. If synthesis produces

$L < P$6

then

$L < P$7

Since the target final state for label coordinates is exactly zero, $L < P$8 is intended to be small for feasible in-distribution targets and larger for unrealistic or out-of-distribution targets. The second metric is Self-Consistency. After synthesis, one replaces the residual label coordinates by exact zeros, integrates backward from

$L < P$9

obtains

$y = f(x) + \varepsilon,$0

and defines

$y = f(x) + \varepsilon,$1

The paper emphasizes that this differs from a trivial round trip because the nonzero residual $y = f(x) + \varepsilon,$2 is discarded before the backward pass (Campos et al., 16 Mar 2026).

These two metrics are used for three tasks: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets. The reported gains are task-wide rather than anecdotal. Candidate selection improves over random choice by $y = f(x) + \varepsilon,$3 to $y = f(x) + \varepsilon,$4; rejecting the top $y = f(x) + \varepsilon,$5 most uncertain samples reduces mean error by $y = f(x) + \varepsilon,$6 to $y = f(x) + \varepsilon,$7; and Zero-Deviation in particular achieves strong ROC AUC values, including $y = f(x) + \varepsilon,$8 on Unifoil, $y = f(x) + \varepsilon,$9 on gas turbine, and $f : \mathbb{R}^P \to \mathbb{R}^L$0–$f : \mathbb{R}^P \to \mathbb{R}^L$1 on DTLZ (Campos et al., 16 Mar 2026).

5. Benchmarks, metrics, and observed behavior

The Diag-CFM study evaluates three inverse-design benchmarks: a gas turbine combustor dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$2 and $f : \mathbb{R}^P \to \mathbb{R}^L$3, a Unifoil airfoil dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$4 and $f : \mathbb{R}^P \to \mathbb{R}^L$5 plus two physical conditioning scalars, and the analytical DTLZ2 benchmark with $f : \mathbb{R}^P \to \mathbb{R}^L$6 up to $f : \mathbb{R}^P \to \mathbb{R}^L$7 and $f : \mathbb{R}^P \to \mathbb{R}^L$8. The central empirical metric for inverse quality is round-trip error, namely the error incurred by generating a design for a target and then evaluating its achieved performance. Forward MSE is also reported, together with diversity analyses and uncertainty benchmarks (Campos et al., 16 Mar 2026).

On the gas turbine benchmark, Diag-CFM attains forward MSE $f : \mathbb{R}^P \to \mathbb{R}^L$9, compared with $\varepsilon$0 for the INN baseline and $\varepsilon$1 for standard CFM. Its round-trip error is $\varepsilon$2, compared with $\varepsilon$3 for INN and $\varepsilon$4 for CFM, which the paper summarizes as more than a $\varepsilon$5 improvement. On Unifoil, Diag-CFM achieves forward MSE $\varepsilon$6 versus $\varepsilon$7 for INN and $\varepsilon$8 for CFM, and round-trip error $\varepsilon$9 versus $[x-z; -y]$00 and $[x-z; -y]$01, respectively (Campos et al., 16 Mar 2026).

The DTLZ2 results are particularly important for scalability. At $[x-z; -y]$02, the reported round-trip errors are $[x-z; -y]$03 for Diag-CFM, $[x-z; -y]$04 for CFM, and $[x-z; -y]$05 for INN. At $[x-z; -y]$06, they are $[x-z; -y]$07, $[x-z; -y]$08, and $[x-z; -y]$09, respectively. The paper therefore characterizes Diag-CFM as maintaining round-trip error in the $[x-z; -y]$10–$[x-z; -y]$11 range across design dimensions up to $[x-z; -y]$12, while CFM and INN degrade more sharply (Campos et al., 16 Mar 2026).

The implementation used for these results is deliberately simple. The velocity network is an MLP with time $[x-z; -y]$13 concatenated directly to the state, and inference uses explicit Euler with $[x-z; -y]$14 uniform steps on $[x-z; -y]$15. The paper presents this as evidence that the empirical gains arise primarily from the flow construction rather than from specialized architectural devices (Campos et al., 16 Mar 2026).

6. Relation to flow matching, guided variants, and Generator Matching

Flow matching more generally is a continuous-time alternative to diffusion in which a vector field defines an ODE transporting a simple base distribution to a target distribution, and training regresses the learned velocity onto a known conditional velocity along a prescribed probability path (Lipman et al., 2022). The Generator Matching framework sharpens this relation by showing that both flow matching and diffusion can be written as Markov processes with generators $[x-z; -y]$16, with pure flow matching corresponding to a first-order operator and diffusion corresponding to a second-order operator with nonzero diffusion matrix. In that view, the empirical robustness of flow matching is linked to the first-order transport PDE and the absence of a smoothing operator that must later be inverted (Patel et al., 2024).

This contextualization matters for Diag-CFM. The inverse-design method remains a flow-matching model, and therefore inherits the first-order ODE character of standard flow matching. The Generator Matching paper further notes that a diagonal flow matching variant can be seen as imposing additional structure on either $[x-z; -y]$17, for example coordinate-wise decoupled drift, or on the diffusion term $[x-z; -y]$18, for example a diagonal diffusion matrix. This suggests a broader family of diagonal or partially diagonal generator constructions beyond the specific zero-anchored inverse-design formulation (Patel et al., 2024).

A related but distinct development is guided flow matching for channel knowledge map construction. That method uses a deterministic ODE,

$[x-z; -y]$19

with the linear interpolation

$[x-z; -y]$20

and constant target velocity

$[x-z; -y]$21

The paper states that it does not explicitly mention the term “Diagonal Flow Matching,” but it describes the resulting structure as reminiscent of DFM or diagonal flows because each component evolves along a coordinate-wise linear path and the analytic target velocity contains no explicit cross-coordinate coupling (Huang et al., 6 Jan 2026).

Taken together, these papers place Diag-CFM at a precise intersection of ideas. It is not merely “flow matching with a simple path,” nor merely “a model with diagonal diffusion.” It is a specific zero-anchored conditional flow-matching construction for inverse design, whose defining property is the target $[x-z; -y]$22, and whose theoretical and empirical significance lies in permutation-equivariant training targets, stable round-trip inverse design, and architecture-intrinsic abstention signals (Campos et al., 16 Mar 2026)."}' ／久久 to=functions.shell атәыла{"output":"<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n<feed xmlns=\"http://www.w3.org/2005/Atom\">\n <link href=\"http://arxiv.org/api/query?search_query=&id_list=([2210.02747](/papers/2210.02747))&start=0&max_results=10\" rel=\"self\" type=\"application/atom+xml\"/>\n <title type=\"html\">ArXiv Query: search_query=&id_list=(Lipman et al., 2022)&start=0&max_results=10</title>\n <id>http://arxiv.org/api/q0wyKR29jlG9gH8ob2+vpSgQe1s</id>\n <updated\>2026-07-05T00:00:00-04:00</updated>\n <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>1/opensearch:totalResults\n <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex\n <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>10/opensearch:itemsPerPage\n","metadata":{"exit_code":0,"duration_seconds":0.47}} Diagonal Flow Matching (Diag-CFM) is a variant of conditional flow matching for generative inverse design that defines a continuous-time, invertible flow in an augmented space $[z; y] \leftrightarrow [x; 0]$ and trains a neural velocity field against the diagonal target $[x-z; -y]$. In the formulation introduced for inverse design, the method uses a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, with the stated goals of removing accidental dependence on coordinate ordering and scaling while preserving bidirectional generation and prediction (Campos et al., 16 Mar 2026). Within the broader Generator Matching view, diagonal structure refers more generally to imposing coordinate-wise restrictions on drift $u_t$ or diagonal diffusion $a_t$, which provides the main conceptual bridge between Diag-CFM and the wider diffusion/flow matching landscape (Patel et al., 2024).

1. Inverse-design setting and the limitation of standard conditional flow matching

Diag-CFM is formulated for parametric engineering design problems with design parameters $x \in \mathbb{R}^P$ and performance labels $y \in \mathbb{R}^L$, typically with $L < P$. The forward relation is written as

$y = f(x) + \varepsilon,$

where $f : \mathbb{R}^P \to \mathbb{R}^L$ is a simulator or surrogate and $\varepsilon$ is noise. The feasible label set is

$[x-z; -y]$0

and, for a target $[x-z; -y]$1 and tolerance $[x-z; -y]$2, the success set is

$[x-z; -y]$3

The objective is to sample diverse designs $[x-z; -y]$4 that satisfy $[x-z; -y]$5, ideally covering the possibly multi-modal success set rather than returning only a single optimizer output (Campos et al., 16 Mar 2026).

Standard conditional flow matching for this setting works in an augmented $[x-z; -y]$6-dimensional space. It uses a latent $[x-z; -y]$7, forms

$[x-z; -y]$8

and linearly interpolates

$[x-z; -y]$9

The corresponding target velocity is

$u_t$0

The paper identifies the central weakness of this construction: different coordinates are matched against different types of quantities, namely “design vs noise” in some coordinates and “design vs label” in others. Because the loss is component-wise, the approximation difficulty depends on the arbitrary decomposition into the first $u_t$1 and last $u_t$2 coordinates. This introduces sensitivity to coordinate ordering and scaling, and the paper reports that standard CFM can be unstable and significantly less accurate than Diag-CFM across all three benchmarks considered (Campos et al., 16 Mar 2026).

2. Zero anchoring and the diagonal construction

Diag-CFM changes the source–target pairing rather than the basic flow-matching recipe. It operates in an augmented space of dimension $u_t$3, samples noise $u_t$4, and defines

$u_t$5

The linear path is

$u_t$6

with target velocity

$u_t$7

This is the sense in which the method is “diagonal”: each design coordinate is paired only with a noise coordinate, and each label coordinate is paired only with zero (Campos et al., 16 Mar 2026).

Construction	Standard CFM	Diag-CFM
Source state	$u_t$8	$u_t$9
Target state	$a_t$0	$a_t$1
Target velocity	$a_t$2	$a_t$3

The training objective is standard flow-matching regression applied to this new state design: $a_t$4 The velocity field is a neural network

$a_t$5

typically implemented in the paper as an MLP. At convergence, integrating

$a_t$6

from $a_t$7 to $a_t$8 approximately transports the distribution of $a_t$9 to the empirical distribution of $x \in \mathbb{R}^P$0 (Campos et al., 16 Mar 2026).

The practical consequence of zero anchoring is that labels are no longer compared directly against a subset of design coordinates. The paper states that this resolves the strong sensitivity of inverse-design CFM to arbitrary ordering and scaling, while still supporting bidirectional use through the same ODE (Campos et al., 16 Mar 2026).

3. Permutation equivariance, robustness, and what “diagonal” means

The main theoretical claim for Diag-CFM is permutation robustness at the level of the target-velocity distribution. Let

$x \in \mathbb{R}^P$1

where $x \in \mathbb{R}^P$2 and $x \in \mathbb{R}^P$3. For permutation matrices $x \in \mathbb{R}^P$4 on design coordinates and $x \in \mathbb{R}^P$5 on label coordinates, define

$x \in \mathbb{R}^P$6

Proposition 3.1 states that the distribution of $x \in \mathbb{R}^P$7 is equivariant under permutations of parameter and label coordinates. Because the entries of $x \in \mathbb{R}^P$8 are i.i.d., permuting coordinates preserves the target-velocity statistics. By contrast, Proposition 3.2 states that the standard CFM target velocity is not equivariant under coordinate permutations, so the learning problem itself depends on arbitrary coordinate order (Campos et al., 16 Mar 2026).

The paper connects this directly to observed training behavior. On the gas turbine dataset, which is normalized to $x \in \mathbb{R}^P$9, simply reordering the parameter coordinates changes standard CFM’s final round-trip error by nearly one order of magnitude, whereas Diag-CFM’s error is reported as essentially constant across orderings. The same section also attributes improved run-to-run stability to the diagonal construction (Campos et al., 16 Mar 2026).

In the broader Generator Matching framework, “diagonal” has a wider meaning than the zero-anchored inverse-design construction. A generator on $y \in \mathbb{R}^L$0 can be written as

$y \in \mathbb{R}^L$1

and a diagonal structure can be imposed either through coordinate-wise drift $y \in \mathbb{R}^L$2 or through a diagonal diffusion matrix $y \in \mathbb{R}^L$3. The Generator Matching analysis therefore treats diagonality as a structural restriction on the generator family, whereas the Diag-CFM inverse-design paper realizes diagonality through the specific augmented-state pairing $y \in \mathbb{R}^L$4 and the target $y \in \mathbb{R}^L$5 (Patel et al., 2024).

A common misconception is to identify all coordinate-wise linear flows with Diag-CFM. The literature summarized here does not support that equivalence. The inverse-design method is a specific zero-anchored construction; related linear-path methods may be only conceptually aligned with diagonal designs rather than instances of Diag-CFM proper (Campos et al., 16 Mar 2026).

4. Bidirectional flow, abstention, and intrinsic uncertainty

Diag-CFM is bidirectional because the same learned ODE is used in both synthesis and analysis. For inverse generation, given a target $y \in \mathbb{R}^L$6, one samples $y \in \mathbb{R}^L$7, initializes

$y \in \mathbb{R}^L$8

integrates

$y \in \mathbb{R}^L$9

and outputs the design

$L < P$0

For forward prediction, given a design $L < P$1, one sets

$L < P$2

integrates backward with sign reversal,

$L < P$3

and reads off

$L < P$4

There is no separate forward and inverse network; both directions share the same $L < P$5 (Campos et al., 16 Mar 2026).

A distinctive aspect of the method is that the zero-anchored architecture exposes two model-intrinsic uncertainty metrics. The first is Zero-Deviation. If synthesis produces

$L < P$6

then

$L < P$7

Since the target final state for label coordinates is exactly zero, $L < P$8 is intended to be small for feasible in-distribution targets and larger for unrealistic or out-of-distribution targets. The second metric is Self-Consistency. After synthesis, one replaces the residual label coordinates by exact zeros, integrates backward from

$L < P$9

obtains

$y = f(x) + \varepsilon,$0

and defines

$y = f(x) + \varepsilon,$1

The paper emphasizes that this differs from a trivial round trip because the nonzero residual $y = f(x) + \varepsilon,$2 is discarded before the backward pass (Campos et al., 16 Mar 2026).

These two metrics are used for three tasks: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets. The reported gains are task-wide rather than anecdotal. Candidate selection improves over random choice by $y = f(x) + \varepsilon,$3 to $y = f(x) + \varepsilon,$4; rejecting the top $y = f(x) + \varepsilon,$5 most uncertain samples reduces mean error by $y = f(x) + \varepsilon,$6 to $y = f(x) + \varepsilon,$7; and Zero-Deviation in particular achieves strong ROC AUC values, including $y = f(x) + \varepsilon,$8 on Unifoil, $y = f(x) + \varepsilon,$9 on gas turbine, and $f : \mathbb{R}^P \to \mathbb{R}^L$0–$f : \mathbb{R}^P \to \mathbb{R}^L$1 on DTLZ (Campos et al., 16 Mar 2026).

5. Benchmarks, metrics, and observed behavior

The Diag-CFM study evaluates three inverse-design benchmarks: a gas turbine combustor dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$2 and $f : \mathbb{R}^P \to \mathbb{R}^L$3, a Unifoil airfoil dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$4 and $f : \mathbb{R}^P \to \mathbb{R}^L$5 plus two physical conditioning scalars, and the analytical DTLZ2 benchmark with $f : \mathbb{R}^P \to \mathbb{R}^L$6 up to $f : \mathbb{R}^P \to \mathbb{R}^L$7 and $f : \mathbb{R}^P \to \mathbb{R}^L$8. The central empirical metric for inverse quality is round-trip error, namely the error incurred by generating a design for a target and then evaluating its achieved performance. Forward MSE is also reported, together with diversity analyses and uncertainty benchmarks (Campos et al., 16 Mar 2026).

On the gas turbine benchmark, Diag-CFM attains forward MSE $f : \mathbb{R}^P \to \mathbb{R}^L$9, compared with $\varepsilon$0 for the INN baseline and $\varepsilon$1 for standard CFM. Its round-trip error is $\varepsilon$2, compared with $\varepsilon$3 for INN and $\varepsilon$4 for CFM, which the paper summarizes as more than a $\varepsilon$5 improvement. On Unifoil, Diag-CFM achieves forward MSE $\varepsilon$6 versus $\varepsilon$7 for INN and $\varepsilon$8 for CFM, and round-trip error $\varepsilon$9 versus $[x-z; -y]$00 and $[x-z; -y]$01, respectively (Campos et al., 16 Mar 2026).

The DTLZ2 results are particularly important for scalability. At $[x-z; -y]$02, the reported round-trip errors are $[x-z; -y]$03 for Diag-CFM, $[x-z; -y]$04 for CFM, and $[x-z; -y]$05 for INN. At $[x-z; -y]$06, they are $[x-z; -y]$07, $[x-z; -y]$08, and $[x-z; -y]$09, respectively. The paper therefore characterizes Diag-CFM as maintaining round-trip error in the $[x-z; -y]$10–$[x-z; -y]$11 range across design dimensions up to $[x-z; -y]$12, while CFM and INN degrade more sharply (Campos et al., 16 Mar 2026).

The implementation used for these results is deliberately simple. The velocity network is an MLP with time $[x-z; -y]$13 concatenated directly to the state, and inference uses explicit Euler with $[x-z; -y]$14 uniform steps on $[x-z; -y]$15. The paper presents this as evidence that the empirical gains arise primarily from the flow construction rather than from specialized architectural devices (Campos et al., 16 Mar 2026).

6. Relation to flow matching, guided variants, and Generator Matching

Flow matching more generally is a continuous-time alternative to diffusion in which a vector field defines an ODE transporting a simple base distribution to a target distribution, and training regresses the learned velocity onto a known conditional velocity along a prescribed probability path (Lipman et al., 2022). The Generator Matching framework sharpens this relation by showing that both flow matching and diffusion can be written as Markov processes with generators $[x-z; -y]$16, with pure flow matching corresponding to a first-order operator and diffusion corresponding to a second-order operator with nonzero diffusion matrix. In that view, the empirical robustness of flow matching is linked to the first-order transport PDE and the absence of a smoothing operator that must later be inverted (Patel et al., 2024).

This contextualization matters for Diag-CFM. The inverse-design method remains a flow-matching model, and therefore inherits the first-order ODE character of standard flow matching. The Generator Matching paper further notes that a diagonal flow matching variant can be seen as imposing additional structure on either $[x-z; -y]$17, for example coordinate-wise decoupled drift, or on the diffusion term $[x-z; -y]$18, for example a diagonal diffusion matrix. This suggests a broader family of diagonal or partially diagonal generator constructions beyond the specific zero-anchored inverse-design formulation (Patel et al., 2024).

A related but distinct development is guided flow matching for channel knowledge map construction. That method uses a deterministic ODE,

$[x-z; -y]$19

with the linear interpolation

$[x-z; -y]$20

and constant target velocity

$[x-z; -y]$21

The paper states that it does not explicitly mention the term “Diagonal Flow Matching,” but it describes the resulting structure as reminiscent of DFM or diagonal flows because each component evolves along a coordinate-wise linear path and the analytic target velocity contains no explicit cross-coordinate coupling (Huang et al., 6 Jan 2026).

Taken together, these papers place Diag-CFM at a precise intersection of ideas. It is not merely “flow matching with a simple path,” nor merely “a model with diagonal diffusion.” It is a specific zero-anchored conditional flow-matching construction for inverse design, whose defining property is the target $[x-z; -y]$22, and whose theoretical and empirical significance lies in permutation-equivariant training targets, stable round-trip inverse design, and architecture-intrinsic abstention signals (Campos et al., 16 Mar 2026).python - <<\"PY\"\nimport urllib.parse, urllib.request, re\nq=\"(Lipman et al., 2022)\"\nurl=\"http://export.arxiv.org/api/query?id_list=\"+q\nprint(urllib.request.urlopen(url).read().decode()[:800])\nPY"}' ／久久 to=functions.shell атәыла{"output":"<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n<feed xmlns=\"http://www.w3.org/2005/Atom\">\n <link href=\"http://arxiv.org/api/query?search_query=&id_list=([2210.02747](/papers/2210.02747))&start=0&max_results=10\" rel=\"self\" type=\"application/atom+xml\"/>\n <title type=\"html\">ArXiv Query: search_query=&id_list=(Lipman et al., 2022)&start=0&max_results=10</title>\n <id>http://arxiv.org/api/q0wyKR29jlG9gH8ob2+vpSgQe1s</id>\n <updated\>2026-07-05T00:00:00-04:00</updated>\n <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>1/opensearch:totalResults\n <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex\n <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>10/opensearch:itemsPerPage\n","metadata":{"exit_code":0,"duration_seconds":0.47}} Diagonal Flow Matching (Diag-CFM) is a variant of conditional flow matching for generative inverse design that defines a continuous-time, invertible flow in an augmented space $[z; y] \leftrightarrow [x; 0]$ and trains a neural velocity field against the diagonal target $[x-z; -y]$. In the formulation introduced for inverse design, the method uses a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, with the stated goals of removing accidental dependence on coordinate ordering and scaling while preserving bidirectional generation and prediction (Campos et al., 16 Mar 2026). Within the broader Generator Matching view, diagonal structure refers more generally to imposing coordinate-wise restrictions on drift $u_t$ or diagonal diffusion $a_t$, which provides the main conceptual bridge between Diag-CFM and the wider diffusion/flow matching landscape (Patel et al., 2024).

1. Inverse-design setting and the limitation of standard conditional flow matching

Diag-CFM is formulated for parametric engineering design problems with design parameters $x \in \mathbb{R}^P$ and performance labels $y \in \mathbb{R}^L$, typically with $L < P$. The forward relation is written as

$y = f(x) + \varepsilon,$

where $f : \mathbb{R}^P \to \mathbb{R}^L$ is a simulator or surrogate and $\varepsilon$ is noise. The feasible label set is

$[x-z; -y]$0

and, for a target $[x-z; -y]$1 and tolerance $[x-z; -y]$2, the success set is

$[x-z; -y]$3

The objective is to sample diverse designs $[x-z; -y]$4 that satisfy $[x-z; -y]$5, ideally covering the possibly multi-modal success set rather than returning only a single optimizer output (Campos et al., 16 Mar 2026).

Standard conditional flow matching for this setting works in an augmented $[x-z; -y]$6-dimensional space. It uses a latent $[x-z; -y]$7, forms

$[x-z; -y]$8

and linearly interpolates

$[x-z; -y]$9

The corresponding target velocity is

$u_t$0

The paper identifies the central weakness of this construction: different coordinates are matched against different types of quantities, namely “design vs noise” in some coordinates and “design vs label” in others. Because the loss is component-wise, the approximation difficulty depends on the arbitrary decomposition into the first $u_t$1 and last $u_t$2 coordinates. This introduces sensitivity to coordinate ordering and scaling, and the paper reports that standard CFM can be unstable and significantly less accurate than Diag-CFM across all three benchmarks considered (Campos et al., 16 Mar 2026).

2. Zero anchoring and the diagonal construction

Diag-CFM changes the source–target pairing rather than the basic flow-matching recipe. It operates in an augmented space of dimension $u_t$3, samples noise $u_t$4, and defines

$u_t$5

The linear path is

$u_t$6

with target velocity

$u_t$7

This is the sense in which the method is “diagonal”: each design coordinate is paired only with a noise coordinate, and each label coordinate is paired only with zero (Campos et al., 16 Mar 2026).

Construction	Standard CFM	Diag-CFM
Source state	$u_t$8	$u_t$9
Target state	$a_t$0	$a_t$1
Target velocity	$a_t$2	$a_t$3

The training objective is standard flow-matching regression applied to this new state design: $a_t$4 The velocity field is a neural network

$a_t$5

typically implemented in the paper as an MLP. At convergence, integrating

$a_t$6

from $a_t$7 to $a_t$8 approximately transports the distribution of $a_t$9 to the empirical distribution of $x \in \mathbb{R}^P$0 (Campos et al., 16 Mar 2026).

The practical consequence of zero anchoring is that labels are no longer compared directly against a subset of design coordinates. The paper states that this resolves the strong sensitivity of inverse-design CFM to arbitrary ordering and scaling, while still supporting bidirectional use through the same ODE (Campos et al., 16 Mar 2026).

3. Permutation equivariance, robustness, and what “diagonal” means

The main theoretical claim for Diag-CFM is permutation robustness at the level of the target-velocity distribution. Let

$x \in \mathbb{R}^P$1

where $x \in \mathbb{R}^P$2 and $x \in \mathbb{R}^P$3. For permutation matrices $x \in \mathbb{R}^P$4 on design coordinates and $x \in \mathbb{R}^P$5 on label coordinates, define

$x \in \mathbb{R}^P$6

Proposition 3.1 states that the distribution of $x \in \mathbb{R}^P$7 is equivariant under permutations of parameter and label coordinates. Because the entries of $x \in \mathbb{R}^P$8 are i.i.d., permuting coordinates preserves the target-velocity statistics. By contrast, Proposition 3.2 states that the standard CFM target velocity is not equivariant under coordinate permutations, so the learning problem itself depends on arbitrary coordinate order (Campos et al., 16 Mar 2026).

The paper connects this directly to observed training behavior. On the gas turbine dataset, which is normalized to $x \in \mathbb{R}^P$9, simply reordering the parameter coordinates changes standard CFM’s final round-trip error by nearly one order of magnitude, whereas Diag-CFM’s error is reported as essentially constant across orderings. The same section also attributes improved run-to-run stability to the diagonal construction (Campos et al., 16 Mar 2026).

In the broader Generator Matching framework, “diagonal” has a wider meaning than the zero-anchored inverse-design construction. A generator on $y \in \mathbb{R}^L$0 can be written as

$y \in \mathbb{R}^L$1

and a diagonal structure can be imposed either through coordinate-wise drift $y \in \mathbb{R}^L$2 or through a diagonal diffusion matrix $y \in \mathbb{R}^L$3. The Generator Matching analysis therefore treats diagonality as a structural restriction on the generator family, whereas the Diag-CFM inverse-design paper realizes diagonality through the specific augmented-state pairing $y \in \mathbb{R}^L$4 and the target $y \in \mathbb{R}^L$5 (Patel et al., 2024).

A common misconception is to identify all coordinate-wise linear flows with Diag-CFM. The literature summarized here does not support that equivalence. The inverse-design method is a specific zero-anchored construction; related linear-path methods may be only conceptually aligned with diagonal designs rather than instances of Diag-CFM proper (Campos et al., 16 Mar 2026).

4. Bidirectional flow, abstention, and intrinsic uncertainty

Diag-CFM is bidirectional because the same learned ODE is used in both synthesis and analysis. For inverse generation, given a target $y \in \mathbb{R}^L$6, one samples $y \in \mathbb{R}^L$7, initializes

$y \in \mathbb{R}^L$8

integrates

$y \in \mathbb{R}^L$9

and outputs the design

$L < P$0

For forward prediction, given a design $L < P$1, one sets

$L < P$2

integrates backward with sign reversal,

$L < P$3

and reads off

$L < P$4

There is no separate forward and inverse network; both directions share the same $L < P$5 (Campos et al., 16 Mar 2026).

A distinctive aspect of the method is that the zero-anchored architecture exposes two model-intrinsic uncertainty metrics. The first is Zero-Deviation. If synthesis produces

$L < P$6

then

$L < P$7

Since the target final state for label coordinates is exactly zero, $L < P$8 is intended to be small for feasible in-distribution targets and larger for unrealistic or out-of-distribution targets. The second metric is Self-Consistency. After synthesis, one replaces the residual label coordinates by exact zeros, integrates backward from

$L < P$9

obtains

$y = f(x) + \varepsilon,$0

and defines

$y = f(x) + \varepsilon,$1

The paper emphasizes that this differs from a trivial round trip because the nonzero residual $y = f(x) + \varepsilon,$2 is discarded before the backward pass (Campos et al., 16 Mar 2026).

These two metrics are used for three tasks: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets. The reported gains are task-wide rather than anecdotal. Candidate selection improves over random choice by $y = f(x) + \varepsilon,$3 to $y = f(x) + \varepsilon,$4; rejecting the top $y = f(x) + \varepsilon,$5 most uncertain samples reduces mean error by $y = f(x) + \varepsilon,$6 to $y = f(x) + \varepsilon,$7; and Zero-Deviation in particular achieves strong ROC AUC values, including $y = f(x) + \varepsilon,$8 on Unifoil, $y = f(x) + \varepsilon,$9 on gas turbine, and $f : \mathbb{R}^P \to \mathbb{R}^L$0–$f : \mathbb{R}^P \to \mathbb{R}^L$1 on DTLZ (Campos et al., 16 Mar 2026).

5. Benchmarks, metrics, and observed behavior

The Diag-CFM study evaluates three inverse-design benchmarks: a gas turbine combustor dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$2 and $f : \mathbb{R}^P \to \mathbb{R}^L$3, a Unifoil airfoil dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$4 and $f : \mathbb{R}^P \to \mathbb{R}^L$5 plus two physical conditioning scalars, and the analytical DTLZ2 benchmark with $f : \mathbb{R}^P \to \mathbb{R}^L$6 up to $f : \mathbb{R}^P \to \mathbb{R}^L$7 and $f : \mathbb{R}^P \to \mathbb{R}^L$8. The central empirical metric for inverse quality is round-trip error, namely the error incurred by generating a design for a target and then evaluating its achieved performance. Forward MSE is also reported, together with diversity analyses and uncertainty benchmarks (Campos et al., 16 Mar 2026).

On the gas turbine benchmark, Diag-CFM attains forward MSE $f : \mathbb{R}^P \to \mathbb{R}^L$9, compared with $\varepsilon$0 for the INN baseline and $\varepsilon$1 for standard CFM. Its round-trip error is $\varepsilon$2, compared with $\varepsilon$3 for INN and $\varepsilon$4 for CFM, which the paper summarizes as more than a $\varepsilon$5 improvement. On Unifoil, Diag-CFM achieves forward MSE $\varepsilon$6 versus $\varepsilon$7 for INN and $\varepsilon$8 for CFM, and round-trip error $\varepsilon$9 versus $[x-z; -y]$00 and $[x-z; -y]$01, respectively (Campos et al., 16 Mar 2026).

The DTLZ2 results are particularly important for scalability. At $[x-z; -y]$02, the reported round-trip errors are $[x-z; -y]$03 for Diag-CFM, $[x-z; -y]$04 for CFM, and $[x-z; -y]$05 for INN. At $[x-z; -y]$06, they are $[x-z; -y]$07, $[x-z; -y]$08, and $[x-z; -y]$09, respectively. The paper therefore characterizes Diag-CFM as maintaining round-trip error in the $[x-z; -y]$10–$[x-z; -y]$11 range across design dimensions up to $[x-z; -y]$12, while CFM and INN degrade more sharply (Campos et al., 16 Mar 2026).

The implementation used for these results is deliberately simple. The velocity network is an MLP with time $[x-z; -y]$13 concatenated directly to the state, and inference uses explicit Euler with $[x-z; -y]$14 uniform steps on $[x-z; -y]$15. The paper presents this as evidence that the empirical gains arise primarily from the flow construction rather than from specialized architectural devices (Campos et al., 16 Mar 2026).

6. Relation to flow matching, guided variants, and Generator Matching

Flow matching more generally is a continuous-time alternative to diffusion in which a vector field defines an ODE transporting a simple base distribution to a target distribution, and training regresses the learned velocity onto a known conditional velocity along a prescribed probability path (Lipman et al., 2022). The Generator Matching framework sharpens this relation by showing that both flow matching and diffusion can be written as Markov processes with generators $[x-z; -y]$16, with pure flow matching corresponding to a first-order operator and diffusion corresponding to a second-order operator with nonzero diffusion matrix. In that view, the empirical robustness of flow matching is linked to the first-order transport PDE and the absence of a smoothing operator that must later be inverted (Patel et al., 2024).

This contextualization matters for Diag-CFM. The inverse-design method remains a flow-matching model, and therefore inherits the first-order ODE character of standard flow matching. The Generator Matching paper further notes that a diagonal flow matching variant can be seen as imposing additional structure on either $[x-z; -y]$17, for example coordinate-wise decoupled drift, or on the diffusion term $[x-z; -y]$18, for example a diagonal diffusion matrix. This suggests a broader family of diagonal or partially diagonal generator constructions beyond the specific zero-anchored inverse-design formulation (Patel et al., 2024).

A related but distinct development is guided flow matching for channel knowledge map construction. That method uses a deterministic ODE,

$[x-z; -y]$19

with the linear interpolation

$[x-z; -y]$20

and constant target velocity

$[x-z; -y]$21

The paper states that it does not explicitly mention the term “Diagonal Flow Matching,” but it describes the resulting structure as reminiscent of DFM or diagonal flows because each component evolves along a coordinate-wise linear path and the analytic target velocity contains no explicit cross-coordinate coupling (Huang et al., 6 Jan 2026).

Taken together, these papers place Diag-CFM at a precise intersection of ideas. It is not merely “flow matching with a simple path,” nor merely “a model with diagonal diffusion.” It is a specific zero-anchored conditional flow-matching construction for inverse design, whose defining property is the target $[x-z; -y]$22, and whose theoretical and empirical significance lies in permutation-equivariant training targets, stable round-trip inverse design, and architecture-intrinsic abstention signals (Campos et al., 16 Mar 2026)."}' ／久久 to=functions.shell атәыла{"output":"<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n<feed xmlns=\"http://www.w3.org/2005/Atom\">\n <link href=\"http://arxiv.org/api/query?search_query=&id_list=([2210.02747](/papers/2210.02747))&start=0&max_results=10\" rel=\"self\" type=\"application/atom+xml\"/>\n <title type=\"html\">ArXiv Query: search_query=&id_list=(Lipman et al., 2022)&start=0&max_results=10</title>\n <id>http://arxiv.org/api/q0wyKR29jlG9gH8ob2+vpSgQe1s</id>\n <updated\>2026-07-05T00:00:00-04:00</updated>\n <opensearch:totalResults xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>1/opensearch:totalResults\n <opensearch:startIndex xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>0/opensearch:startIndex\n <opensearch:itemsPerPage xmlns:opensearch=\"http://a9.com/-/spec/opensearch/1.1/\"\>10/opensearch:itemsPerPage\n","metadata":{"exit_code":0,"duration_seconds":0.47}} Diagonal Flow Matching (Diag-CFM) is a variant of conditional flow matching for generative inverse design that defines a continuous-time, invertible flow in an augmented space $[z; y] \leftrightarrow [x; 0]$ and trains a neural velocity field against the diagonal target $[x-z; -y]$. In the formulation introduced for inverse design, the method uses a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, with the stated goals of removing accidental dependence on coordinate ordering and scaling while preserving bidirectional generation and prediction (Campos et al., 16 Mar 2026). Within the broader Generator Matching view, diagonal structure refers more generally to imposing coordinate-wise restrictions on drift $u_t$ or diagonal diffusion $a_t$, which provides the main conceptual bridge between Diag-CFM and the wider diffusion/flow matching landscape (Patel et al., 2024).

1. Inverse-design setting and the limitation of standard conditional flow matching

Diag-CFM is formulated for parametric engineering design problems with design parameters $x \in \mathbb{R}^P$ and performance labels $y \in \mathbb{R}^L$, typically with $L < P$. The forward relation is written as

$y = f(x) + \varepsilon,$

where $f : \mathbb{R}^P \to \mathbb{R}^L$ is a simulator or surrogate and $\varepsilon$ is noise. The feasible label set is

$[x-z; -y]$0

and, for a target $[x-z; -y]$1 and tolerance $[x-z; -y]$2, the success set is

$[x-z; -y]$3

The objective is to sample diverse designs $[x-z; -y]$4 that satisfy $[x-z; -y]$5, ideally covering the possibly multi-modal success set rather than returning only a single optimizer output (Campos et al., 16 Mar 2026).

Standard conditional flow matching for this setting works in an augmented $[x-z; -y]$6-dimensional space. It uses a latent $[x-z; -y]$7, forms

$[x-z; -y]$8

and linearly interpolates

$[x-z; -y]$9

The corresponding target velocity is

$u_t$0

The paper identifies the central weakness of this construction: different coordinates are matched against different types of quantities, namely “design vs noise” in some coordinates and “design vs label” in others. Because the loss is component-wise, the approximation difficulty depends on the arbitrary decomposition into the first $u_t$1 and last $u_t$2 coordinates. This introduces sensitivity to coordinate ordering and scaling, and the paper reports that standard CFM can be unstable and significantly less accurate than Diag-CFM across all three benchmarks considered (Campos et al., 16 Mar 2026).

2. Zero anchoring and the diagonal construction

Diag-CFM changes the source–target pairing rather than the basic flow-matching recipe. It operates in an augmented space of dimension $u_t$3, samples noise $u_t$4, and defines

$u_t$5

The linear path is

$u_t$6

with target velocity

$u_t$7

This is the sense in which the method is “diagonal”: each design coordinate is paired only with a noise coordinate, and each label coordinate is paired only with zero (Campos et al., 16 Mar 2026).

Construction	Standard CFM	Diag-CFM
Source state	$u_t$8	$u_t$9
Target state	$a_t$0	$a_t$1
Target velocity	$a_t$2	$a_t$3

The training objective is standard flow-matching regression applied to this new state design: $a_t$4 The velocity field is a neural network

$a_t$5

typically implemented in the paper as an MLP. At convergence, integrating

$a_t$6

from $a_t$7 to $a_t$8 approximately transports the distribution of $a_t$9 to the empirical distribution of $x \in \mathbb{R}^P$0 (Campos et al., 16 Mar 2026).

The practical consequence of zero anchoring is that labels are no longer compared directly against a subset of design coordinates. The paper states that this resolves the strong sensitivity of inverse-design CFM to arbitrary ordering and scaling, while still supporting bidirectional use through the same ODE (Campos et al., 16 Mar 2026).

3. Permutation equivariance, robustness, and what “diagonal” means

The main theoretical claim for Diag-CFM is permutation robustness at the level of the target-velocity distribution. Let

$x \in \mathbb{R}^P$1

where $x \in \mathbb{R}^P$2 and $x \in \mathbb{R}^P$3. For permutation matrices $x \in \mathbb{R}^P$4 on design coordinates and $x \in \mathbb{R}^P$5 on label coordinates, define

$x \in \mathbb{R}^P$6

Proposition 3.1 states that the distribution of $x \in \mathbb{R}^P$7 is equivariant under permutations of parameter and label coordinates. Because the entries of $x \in \mathbb{R}^P$8 are i.i.d., permuting coordinates preserves the target-velocity statistics. By contrast, Proposition 3.2 states that the standard CFM target velocity is not equivariant under coordinate permutations, so the learning problem itself depends on arbitrary coordinate order (Campos et al., 16 Mar 2026).

The paper connects this directly to observed training behavior. On the gas turbine dataset, which is normalized to $x \in \mathbb{R}^P$9, simply reordering the parameter coordinates changes standard CFM’s final round-trip error by nearly one order of magnitude, whereas Diag-CFM’s error is reported as essentially constant across orderings. The same section also attributes improved run-to-run stability to the diagonal construction (Campos et al., 16 Mar 2026).

In the broader Generator Matching framework, “diagonal” has a wider meaning than the zero-anchored inverse-design construction. A generator on $y \in \mathbb{R}^L$0 can be written as

$y \in \mathbb{R}^L$1

and a diagonal structure can be imposed either through coordinate-wise drift $y \in \mathbb{R}^L$2 or through a diagonal diffusion matrix $y \in \mathbb{R}^L$3. The Generator Matching analysis therefore treats diagonality as a structural restriction on the generator family, whereas the Diag-CFM inverse-design paper realizes diagonality through the specific augmented-state pairing $y \in \mathbb{R}^L$4 and the target $y \in \mathbb{R}^L$5 (Patel et al., 2024).

A common misconception is to identify all coordinate-wise linear flows with Diag-CFM. The literature summarized here does not support that equivalence. The inverse-design method is a specific zero-anchored construction; related linear-path methods may be only conceptually aligned with diagonal designs rather than instances of Diag-CFM proper (Campos et al., 16 Mar 2026).

4. Bidirectional flow, abstention, and intrinsic uncertainty

Diag-CFM is bidirectional because the same learned ODE is used in both synthesis and analysis. For inverse generation, given a target $y \in \mathbb{R}^L$6, one samples $y \in \mathbb{R}^L$7, initializes

$y \in \mathbb{R}^L$8

integrates

$y \in \mathbb{R}^L$9

and outputs the design

$L < P$0

For forward prediction, given a design $L < P$1, one sets

$L < P$2

integrates backward with sign reversal,

$L < P$3

and reads off

$L < P$4

There is no separate forward and inverse network; both directions share the same $L < P$5 (Campos et al., 16 Mar 2026).

A distinctive aspect of the method is that the zero-anchored architecture exposes two model-intrinsic uncertainty metrics. The first is Zero-Deviation. If synthesis produces

$L < P$6

then

$L < P$7

Since the target final state for label coordinates is exactly zero, $L < P$8 is intended to be small for feasible in-distribution targets and larger for unrealistic or out-of-distribution targets. The second metric is Self-Consistency. After synthesis, one replaces the residual label coordinates by exact zeros, integrates backward from

$L < P$9

obtains

$y = f(x) + \varepsilon,$0

and defines

$y = f(x) + \varepsilon,$1

The paper emphasizes that this differs from a trivial round trip because the nonzero residual $y = f(x) + \varepsilon,$2 is discarded before the backward pass (Campos et al., 16 Mar 2026).

These two metrics are used for three tasks: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets. The reported gains are task-wide rather than anecdotal. Candidate selection improves over random choice by $y = f(x) + \varepsilon,$3 to $y = f(x) + \varepsilon,$4; rejecting the top $y = f(x) + \varepsilon,$5 most uncertain samples reduces mean error by $y = f(x) + \varepsilon,$6 to $y = f(x) + \varepsilon,$7; and Zero-Deviation in particular achieves strong ROC AUC values, including $y = f(x) + \varepsilon,$8 on Unifoil, $y = f(x) + \varepsilon,$9 on gas turbine, and $f : \mathbb{R}^P \to \mathbb{R}^L$0–$f : \mathbb{R}^P \to \mathbb{R}^L$1 on DTLZ (Campos et al., 16 Mar 2026).

5. Benchmarks, metrics, and observed behavior

The Diag-CFM study evaluates three inverse-design benchmarks: a gas turbine combustor dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$2 and $f : \mathbb{R}^P \to \mathbb{R}^L$3, a Unifoil airfoil dataset with $f : \mathbb{R}^P \to \mathbb{R}^L$4 and $f : \mathbb{R}^P \to \mathbb{R}^L$5 plus two physical conditioning scalars, and the analytical DTLZ2 benchmark with $f : \mathbb{R}^P \to \mathbb{R}^L$6 up to $f : \mathbb{R}^P \to \mathbb{R}^L$7 and $f : \mathbb{R}^P \to \mathbb{R}^L$8. The central empirical metric for inverse quality is round-trip error, namely the error incurred by generating a design for a target and then evaluating its achieved performance. Forward MSE is also reported, together with diversity analyses and uncertainty benchmarks (Campos et al., 16 Mar 2026).

On the gas turbine benchmark, Diag-CFM attains forward MSE $f : \mathbb{R}^P \to \mathbb{R}^L$9, compared with $\varepsilon$0 for the INN baseline and $\varepsilon$1 for standard CFM. Its round-trip error is $\varepsilon$2, compared with $\varepsilon$3 for INN and $\varepsilon$4 for CFM, which the paper summarizes as more than a $\varepsilon$5 improvement. On Unifoil, Diag-CFM achieves forward MSE $\varepsilon$6 versus $\varepsilon$7 for INN and $\varepsilon$8 for CFM, and round-trip error $\varepsilon$9 versus $[x-z; -y]$00 and $[x-z; -y]$01, respectively (Campos et al., 16 Mar 2026).

The DTLZ2 results are particularly important for scalability. At $[x-z; -y]$02, the reported round-trip errors are $[x-z; -y]$03 for Diag-CFM, $[x-z; -y]$04 for CFM, and $[x-z; -y]$05 for INN. At $[x-z; -y]$06, they are $[x-z; -y]$07, $[x-z; -y]$08, and $[x-z; -y]$09, respectively. The paper therefore characterizes Diag-CFM as maintaining round-trip error in the $[x-z; -y]$10–$[x-z; -y]$11 range across design dimensions up to $[x-z; -y]$12, while CFM and INN degrade more sharply (Campos et al., 16 Mar 2026).

The implementation used for these results is deliberately simple. The velocity network is an MLP with time $[x-z; -y]$13 concatenated directly to the state, and inference uses explicit Euler with $[x-z; -y]$14 uniform steps on $[x-z; -y]$15. The paper presents this as evidence that the empirical gains arise primarily from the flow construction rather than from specialized architectural devices (Campos et al., 16 Mar 2026).

6. Relation to flow matching, guided variants, and Generator Matching

Flow matching more generally is a continuous-time alternative to diffusion in which a vector field defines an ODE transporting a simple base distribution to a target distribution, and training regresses the learned velocity onto a known conditional velocity along a prescribed probability path (Lipman et al., 2022). The Generator Matching framework sharpens this relation by showing that both flow matching and diffusion can be written as Markov processes with generators $[x-z; -y]$16, with pure flow matching corresponding to a first-order operator and diffusion corresponding to a second-order operator with nonzero diffusion matrix. In that view, the empirical robustness of flow matching is linked to the first-order transport PDE and the absence of a smoothing operator that must later be inverted (Patel et al., 2024).

This contextualization matters for Diag-CFM. The inverse-design method remains a flow-matching model, and therefore inherits the first-order ODE character of standard flow matching. The Generator Matching paper further notes that a diagonal flow matching variant can be seen as imposing additional structure on either $[x-z; -y]$17, for example coordinate-wise decoupled drift, or on the diffusion term $[x-z; -y]$18, for example a diagonal diffusion matrix. This suggests a broader family of diagonal or partially diagonal generator constructions beyond the specific zero-anchored inverse-design formulation (Patel et al., 2024).

A related but distinct development is guided flow matching for channel knowledge map construction. That method uses a deterministic ODE,

$[x-z; -y]$19

with the linear interpolation

$[x-z; -y]$20

and constant target velocity

$[x-z; -y]$21

The paper states that it does not explicitly mention the term “Diagonal Flow Matching,” but it describes the resulting structure as reminiscent of DFM or diagonal flows because each component evolves along a coordinate-wise linear path and the analytic target velocity contains no explicit cross-coordinate coupling (Huang et al., 6 Jan 2026).

Taken together, these papers place Diag-CFM at a precise intersection of ideas. It is not merely “flow matching with a simple path,” nor merely “a model with diagonal diffusion.” It is a specific zero-anchored conditional flow-matching construction for inverse design, whose defining property is the target $[x-z; -y]$22, and whose theoretical and empirical significance lies in permutation-equivariant training targets, stable round-trip inverse design, and architecture-intrinsic abstention signals (Campos et al., 16 Mar 2026).