Logic-Guided Vector Fields for Constrained Generative Modeling

Ali Baheri

REVIEW 3 major objections 4 minor 3 cited by

Logic-guided vector fields let flow-matching generative models enforce logical and geometric constraints along the entire generation trajectory, cutting constraint violations by 59–82% in the paper's benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:28 UTC pith:MGILAR3G

load-bearing objection Eq. (2)'s logic loss has no gradient w.r.t. θ, so the claimed LGVF training effects cannot come from the stated method; the inference-time adjustment is a known idea with toy support. the 3 major comments →

arxiv 2602.02009 v2 pith:MGILAR3G submitted 2026-02-02 cs.LG

Logic-Guided Vector Fields for Constrained Generative Modeling

Ali Baheri This is my paper

classification cs.LG

keywords flow matchingconstraint satisfactionlogic-guided vector fieldsdifferentiable constraintsguided samplinggenerative modelingneuro-symbolic methodsobstacle avoidance

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper proposes Logic-Guided Vector Fields (LGVF), a way to make flow-matching generative models respect logical and geometric constraints by acting on the whole generation trajectory, not just the final sample. It adds a time-weighted logic loss during training that penalizes intermediate states violating the constraint, plus a scheduled gradient correction during sampling that steers trajectories out of forbidden regions. The paper reports this two-stage scheme reduces constraint violations by 59–82% across linear, ring, and multi-obstacle benchmarks compared with standard flow matching, while preserving or improving distributional fidelity in two of three cases. If correct, LGVF is a lightweight recipe for injecting safety rules, feasibility regions, or obstacle-avoidance behavior into continuous-time generative models without explicit path planning or rejection sampling.

Core claim

The paper's central claim is that constraint satisfaction in flow matching is best enforced along the flow, not only at the endpoint. Standard flow-matching trains a vector field to interpolate from noise to data, and the learned transport can pass through forbidden regions even when all training data are valid. LGVF augments the flow-matching objective with a logic loss weighted by lambda(t)=lambda_max * t, penalizing violations at every intermediate state, and then at sampling time subtracts eta(t) * grad_x l_logic(x) from the learned vector field after t0=0.3 to actively steer samples into the valid set. Across three 2D constraint families (half-plane, ring, three obstacles) and a high-di

What carries the argument

The central object is the LGVF-augmented flow: a learned vector field v_theta(x,t) trained with objective L_FM + E[lambda(t) * l_logic(x_t)], plus the inference-time adjusted field v_tilde(x,t) = v_theta(x,t) - eta(t) * grad_x l_logic(x), where l_logic is a differentiable relaxation of the logical constraint (zero if and only if satisfied) and eta is a quadratic ramp that starts at t0=0.3. The mechanism carrying the argument is the descent identity d/dt l(x_t) = grad_l^T v_theta - eta * ||grad_l||^2, which shows the adjustment adds a guaranteed non-positive dissipation term, turning constraint reduction into a Lyapunov-style process. The training loss shapes the vector field globally; the in

Load-bearing premise

The load-bearing assumption is that the sampling-time gradient correction is only a gentle nudge: strong enough to remove violations but weak enough that solving the modified ODE leaves the generated distribution essentially unchanged; the paper's formal support for this is a bound on path deviation, not on output distributions.

What would settle it

Measure the Wasserstein-2 distance between samples produced by the base flow and by LGVF+Adjusted on the d=100 linear constraint at the largest reported eta_max; if that distance is large even as violations drop to near zero, the perturbation-preserves-quality assumption fails.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

On the three 2D benchmarks, LGVF with adjustment cuts violations by 59–82% relative to flow matching and achieves the lowest violation rate in every case (0.4%, 1.2%, 0.7%).
On a high-dimensional half-space constraint (d=10 to 100), LGVF+Adjusted keeps violations near zero (0.0–0.4%) while baseline violations grow to 14.8% at d=100.
Training-time logic loss alone helps for convex constraints but can hurt on multi-region non-convex constraints (obstacle violations rose from 1.7% to 2.5%), whereas the inference-time adjustment is uniformly beneficial.
Distributional fidelity (MMD) improves in the linear and ring cases but degrades in the multi-obstacle case, so the method exhibits a satisfaction–fidelity trade-off rather than a free improvement.
The adjustment term is a guaranteed descent direction on the violation function, so violations cannot increase over time whenever the schedule is strong enough to dominate the alignment term.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same training-loss-plus-gradient-correction recipe should transfer to other continuous-time generative models (for example, diffusion ODEs or continuous normalizing flows) as long as the vector field and violation function are differentiable; the paper only demonstrates it on flow matching.
The L2 trajectory-deviation bound is a weak proxy for distributional fidelity; a direct test of the perturbation assumption would measure Wasserstein distance or run a two-sample test between base-flow and adjusted-flow outputs, rather than relying only on 2D MMD values.
The multi-obstacle failure of training-time guidance suggests that summing per-obstacle violations creates spurious critical points; decomposing the logic loss or using a smoothed max instead of a sum might make training-time shaping more robust in non-convex landscapes.
A natural extension is to learn the violation function l_logic itself for implicit constraints—for example, from demonstrations or a learned feasibility classifier—which the paper lists as future work and which would let LGVF apply to constraints that are not manually specified.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Eq. (2)'s logic loss has no gradient w.r.t. θ, so the claimed LGVF training effects cannot come from the stated method; the inference-time adjustment is a known idea with toy support.

read the letter

Two things you should know. First, the training-time logic loss in Eq. (2) is a constant with respect to the model parameters: x_t = (1-t)x0 + t x1, sampled independently of v_θ, so ∂L_logic/∂θ ≡ 0. Algorithm 1's update therefore reduces to plain flow matching. Table 1 and Figure 6 report differences between FM and LGVF training that can't come from the stated objective. That's a load-bearing inconsistency, not a minor typo. The appendix never analyzes a θ-dependent training loss; it only bounds the inference-time adjustment.

Second, what's actually new is modest but real: applying a late-time gradient correction to a flow-matching sampler, with a quadratic ramp and some elementary bounds. That's a reasonable extension of classifier/universal guidance to flows. The trajectory-deviation bound (Prop 4) is correct but only L2 path deviation, not distributional distance. The geometric examples (half-space, obstacle, ring) are nice sanity checks.

The experiments are weak elsewhere: Table 1 has single-run numbers without error bars; the only baseline is vanilla flow matching plus an ablation; none of the constrained-diffusion methods cited in Section 2 are compared; no code is shipped. The conclusion says 'no trade-off,' but the obstacle MMD gets worse (0.40 → 0.82), which the text acknowledges but the conclusion forgets.

The Appendix's Lyapunov view is fine but elementary. Proposition 5 is just a statement that a characteristic kernel separates distributions with different support, not a fidelity result.

Bottom line: the two-stage method, as written, has one inert stage. The empirical evidence for the training-time component must come from an unreported mechanism or a bug. A serious referee would catch this quickly, but I'd rather the editor return the paper to the authors with a request to either fix the loss to actually depend on θ (e.g., penalize violations along model-generated trajectories) or remove the training-time claims entirely. If they fix it, the inference-time part plus an honest comparison might support a short paper. As is, I would not cite or teach it.

Referee Report

3 major / 4 minor

Summary. The paper proposes Logic-Guided Vector Fields (LGVF), a two-stage method for injecting differentiable logical/geometric constraints into flow-matching generative models. The training stage adds a time-weighted logic loss L_logic to the flow-matching objective, supposedly penalizing constraint violations along flow trajectories. The inference stage modifies the learned ODE by subtracting a scheduled gradient of the violation function. The method is evaluated on three 2D constraint geometries (linear half-plane, ring, multi-obstacle) and a high-dimensional half-space constraint, reporting violation-rate reductions of 59–82% relative to plain flow matching, with MMD either improving or degrading depending on the setting. The paper also presents an appendix with Lyapunov-style analyses of the inference-time adjustment and a trajectory-deviation bound.

Significance. The problem of constraining generative models at both training and sampling time is relevant, and the inference-time adjustment idea is a natural extension of classifier guidance to flow matching. The paper includes useful visualizations and several formal statements about the correction dynamics. However, the central method as written is not internally consistent: the training-time logic loss has no dependence on the model parameters, so Algorithm 1 is exactly standard flow matching. The reported differences between LGVF and plain flow matching are therefore impossible under the stated objective. This undermines the primary contribution and the claimed complementarity between training-time and inference-time guidance. The inference-time component alone may be salvageable, but the manuscript's central quantitative claims are not supported by its own equations.

major comments (3)

[Section 3, Eq. (2); Algorithm 1] The logic loss is defined as L_logic(θ) = E_{t,x0,x1}[λ(t)ℓ((1−t)x0 + t x1)]. Since x_t is an interpolation of independently sampled x0 and x1 and does not depend on v_θ, ∂L_logic/∂θ ≡ 0. The update in Algorithm 1 is therefore exactly plain flow matching; the logic loss is a constant in the optimization. The differences reported in Table 1 (e.g., Flow Matching 2.20% vs LGVF 2.00% in Linear; 5.65% vs 3.45% in Ring; 1.70% vs 2.50% in Obstacles) and the ablation in Figure 6(d) cannot arise from the stated objective. The appendix develops theory only for the inference-time adjustment, never for a θ-dependent training loss. This is a load-bearing flaw: the training-time half of the central claim is unsupported, and the experiments must be re-specified and re-run with a properly defined training objective.
[Section 4, Table 1 and Figure 5] The main quantitative claim of 59–82% violation reduction is based on single-run 2,000-sample rates with no confidence intervals in Table 1, and Figure 5 has no error bars. Several key comparisons are between small percentages (1.7% vs 2.5% vs 0.7% in the obstacle case), where Monte Carlo variability could change the qualitative conclusion. The ablation figure (Figure 6) provides error bars only for the linear case. In addition, the only baselines are vanilla flow matching and FM+Adjusted; comparison with projection, rejection sampling, classifier guidance, or a constrained-diffusion baseline would be needed to substantiate the framework-level claims. As it stands, the stated improvements are not statistically grounded.
[Section 3, after Eq. (4); Proposition 4] The paper claims that when η(t) is small relative to ∥v_θ∥, the inference-time adjustment is a perturbation that does not significantly alter the learned distribution. Proposition 4, however, bounds the L2 path deviation between adjusted and unadjusted trajectories; it does not bound any distributional distance. The MMD evidence is limited to two 2D cases and worsens in the obstacle setting (Table 1: MMD 0.40→0.82). Thus the fidelity-preservation part of the central claim is not supported by the provided theory or experiments.

minor comments (4)

[Section 2] The related-work subsections 'Flow-Based Generative Models' and 'Guided Generation' appear twice, nearly verbatim. This is a presentation error that should be fixed.
[Section 4, paragraph after Table 1] The text says 'MMD scores remain comparable or improve with LGVF, indicating that there is no trade-off,' but the Obstacles row in Table 1 shows MMD increasing from 0.40 to 0.82, and the earlier text explicitly describes a satisfaction–fidelity trade-off. This contradiction should be resolved.
[Abstract and Section 1] There are typos: 'toenforcedeclarative' in the abstract and 'guaranty' in Section 1. In the references, a placeholder '?' appears in the rectified-flow citation.
[Appendix A] The appendix title says 'Theoretical Insights for Logic-Guided Vector Fields,' but the theory covers only the inference-time adjustment. This should be stated clearly in the main text so readers do not expect formal support for the training-time loss.

Circularity Check

3 steps flagged

Training-time logic loss is θ-independent, so Eq. (2) cannot produce the reported LGVF gains; the operative inference-time correction is gradient descent on the same violation function used as the evaluation metric, making much of the violation reduction definitional.

specific steps

other [Section 3, Eq. (2) and Algorithm 1]
"LLGVF(θ)=LFM(θ)+E_{t,x0,x1}[λ(t)·ℓ_logic(x_t)] ... x_t=(1−t)x0+t x1 ... Algorithm 1: 6: L_logic ← (1/B)Σ λ(t^(i))·ℓ_logic(x_t^(i)); 7: Update θ to minimize L_FM + L_logic."

Since x0 and x1 are sampled independently of vθ and t is exogenous, L_logic contains no model parameters: ∂L_logic/∂θ ≡ 0. Algorithm 1 is therefore exactly conditional flow matching. The paper nevertheless attributes LGVF-alone violation reductions to this term (e.g., Ring 5.65%→3.45%; Obstacles 1.7%→2.5%; ablation 0.50%→0.20%). Those reductions cannot follow from Eq. (2): the training-time 'guidance' is a constant in the optimization, so the claimed shaping of the learned vector field is not derivable from the paper's own equations.
self definitional [Section 3, Eqs. (3)–(4), Algorithm 2; Section 4 metrics]
"ṽ(xt,t)=vθ(xt,t)−η(t)·∇xℓ_logic(xt) ... We report three metrics ... (1) Violation Rate, the percentage of samples failing to satisfy the constraint; (2) Average Violation, the mean constraint violation magnitude across all samples."

The inference-time adjustment is explicit gradient descent on ℓ_logic, while Violation Rate and Average Violation are computed from that same ℓ_logic. Thus LGVF+Adjusted's violation reductions (82%, 79%, 59%, 96–100%) are the direct effect of the very optimization being evaluated, not an independent empirical confirmation. The genuinely independent content is MMD, which the paper honestly reports worsens in the obstacle setting (0.40→0.82×10−3), so the circularity is partial rather than total.
renaming known result [Abstract/Conclusion vs. Appendix A.4]
"LGVF yields constraint-aware vector fields exhibiting emergent obstacle-avoidance behavior, routing samples around forbidden regions without explicit path planning ... −η∇ℓ(x)=η (x−c)/∥x−c∥ ... the adjustment is a purely radial 'repulsion' term."

Obstacle avoidance is presented as an emergent property learned by LGVF, but Algorithm 2 explicitly injects this radial repulsion at sampling time, and Eq. (2) is θ-independent and so cannot learn it into vθ. The appendix's own derivation identifies the adjustment as a purely radial repulsion term. The 'emergent' behavior is therefore an explicit input renamed as an emergent output, not a discovered consequence of the training procedure.

full rationale

The central derivation chain breaks at the training objective: in Eq. (2), L_logic averages λ(t)ℓ_logic(x_t) over independently sampled x0, x1, and t, with x_t=(1−t)x0+t x1, so it does not depend on vθ. Consequently ∂L_logic/∂θ=0 and Algorithm 1 reduces to plain conditional flow matching. The reported FM-vs-LGVF differences are impossible under the stated method, which removes the training-time half of the claimed two-stage mechanism. At inference time, Eq. (3) explicitly performs gradient descent on ℓ_logic, and Violation Rate/Average Violation are computed from that same ℓ_logic, so improvement on those metrics is in the direction the algorithm is defined to optimize; that part is by construction. The paper is not wholly circular because it also evaluates MMD, reports a satisfaction–fidelity trade-off in the obstacle case, and provides ablation comparisons (FM+Adj vs LGVF+Adj), and the self-citations (Baheri 2025; Baheri and Lindemann 2025; Yifru and Baheri 2024) are background related work rather than load-bearing evidence. Appendix Proposition 4 is a genuine perturbation bound, though it controls L2 trajectory deviation rather than distributional distance; that is an evidentiary gap, not circularity. Overall, the violation-reduction claim is substantially definitional once the inert training-time term is recognized, so a partial-circularity score of 6 is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method requires the user to supply a differentiable violation function that is exactly zero on the feasible set, and it relies on standard flow matching assumptions plus regularity conditions for the theory. Main empirical free parameters are the logic-loss scale, the adjustment strength, and the adjustment start time; their values are reported only as ranges. No invented entities are introduced.

free parameters (5)

lambda_max (logic loss scale) = 10-15 (2D experiments); up to 50 in ablation
Training-time penalty weight; chosen per constraint difficulty; no selection rule reported; violations decrease monotonically with it (Fig. 6a).
eta_max (inference adjustment strength) = 0.5-1.5
Strength of gradient correction; chosen per case; ablation shows both LGVF+Adj and FM+Adj improve with larger eta.
t0 (adjustment start time) = 0.3
Start of late-time correction; ablation shows small sensitivity but t0≤0.3 best.
alpha in lambda(t)=lambda_max*t^alpha = 1
Linear weighting chosen in all experiments.
MMD Gaussian bandwidth sigma = 1.0
Metric choice; sensitivity not reported.

axioms (5)

domain assumption A differentiable violation function ℓ exists with ℓ(x)=0 iff constraint ϕ holds and gradient available a.e.
Section 3 Methodology: 'We assume access to a differentiable relaxation...' Many real constraints may not admit such exact relaxation.
standard math Conditional flow matching objective (Eq. 1) is a valid way to learn the marginal flow
Relies on Lipman et al. 2023; standard.
domain assumption Regularity conditions A1-A3 (Lipschitz drift, smooth violation, bounded gradients)
Appendix A.1; needed for Propositions 2-4 but not verified empirically.
domain assumption Target distribution q has zero mass in infeasible region
Proposition 5 assumes q(S^c)=0; used to interpret MMD improvements; main experiments place Gaussian modes fully in valid region.
standard math Gaussian MMD kernel is characteristic
Standard property; needed for Proposition 5.

pith-pipeline@v1.3.0-alltime-deepseek · 11571 in / 13544 out tokens · 125244 ms · 2026-08-03T05:28:50.895473+00:00 · methodology

0 comments

Cite this review

Pith. "Pith review of Logic-Guided Vector Fields for Constrained Generative Modeling." pith.science (2026). https://pith.science/paper/MGILAR3G

@misc{pith2026260202009,
  author       = {Pith},
  title        = {Pith review of: Logic-Guided Vector Fields for Constrained Generative Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MGILAR3G}},
  note         = {Machine review of arXiv:2602.02009}
}

Share X Bluesky LinkedIn Reddit HN

read the original abstract

Neuro-symbolic systems aim to combine the expressive structure of symbolic logic with the flexibility of neural learning; yet, generative models typically lack mechanisms to enforce declarative constraints at generation time. We propose Logic-Guided Vector Fields (LGVF), a neuro-symbolic framework that injects symbolic knowledge, specified as differentiable relaxations of logical constraints, into flow matching generative models. LGVF couples two complementary mechanisms: (1) a training-time logic loss that penalizes constraint violations along continuous flow trajectories, with weights that emphasize correctness near the target distribution; and (2) an inference-time adjustment that steers sampling using constraint gradients, acting as a lightweight, logic-informed correction to the learned dynamics. We evaluate LGVF on three constrained generation case studies spanning linear, nonlinear, and multi-region feasibility constraints. Across all settings, LGVF reduces constraint violations by 59-82% compared to standard flow matching and achieves the lowest violation rates in each case. In the linear and ring settings, LGVF also improves distributional fidelity as measured by MMD, while in the multi-obstacle setting, we observe a satisfaction-fidelity trade-off, with improved feasibility but increased MMD. Beyond quantitative gains, LGVF yields constraint-aware vector fields exhibiting emergent obstacle-avoidance behavior, routing samples around forbidden regions without explicit path planning.

Figures

Figures reproduced from arXiv: 2602.02009 by Ali Baheri.

**Figure 1.** Figure 1: LGVF learns constraint-aware vector fields. While standard flow matching follows straight-line paths (dashed) that may traverse forbidden regions, LGVF combines flow matching loss LFM with logic guidance Llogic to learn vector fields that route trajectories around constraint violations. The logic loss penalizes violations at intermediate states xt , encouraging the model to discover valid paths from base d… view at source ↗

**Figure 2.** Figure 2: Case Study 1: Linear half-plane constraint (x1 + x2 ≥ 0). (a) Target distribution with two Gaussian modes in the valid region. (b) Flow Matching: 2.2% violations near the diagonal boundary. (c) LGVF: slight improvement to 2.0%. (d) LGVF + Adjusted: 0.4% violations (82% reduction). Red crosses indicate violations; blue dots are valid samples [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Case Study 2: Nonlinear ring constraint (1.5 ≤ ∥x∥ ≤ 2.8). The valid region is the white annulus; inner disk and outer region are forbidden. Flow Matching produces 5.65% violations in both regions. LGVF reduces this to 3.45%, and LGVF + Adjusted achieves 1.20% (79% reduction), with violations nearly eliminated at both boundaries. function is ℓlogic(x) = max(0, rmin − ∥x∥) + max(0, ∥x∥ −rmax) [PITH_FULL_IM… view at source ↗

**Figure 4.** Figure 4: Case Study 3: Multi-obstacle avoidance. Three circular obstacles (pink) block direct paths to target modes (green in panel a). Flow Matching: 1.7% violations. LGVF alone increases violations to 2.5% due to complex non-convex geometry, but LGVF + Adjusted recovers to 0.7% (59% improvement), demonstrating the robustness of inference-time correction. violation rate of 1.7%, with violations distributed among a… view at source ↗

**Figure 5.** Figure 5: Case Study 4: Scaling to high dimensions. Flow Matching violations increase with dimension, while LGVF + Adjusted maintains near-zero violations. Ablation Study. We conduct ablation studies on the linear constraint case to understand hyperparameter sensitivity and component contributions. All experiments report the mean and standard deviation over 3 independent runs ( [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗

**Figure 6.** Figure 6: Ablation studies on the linear constraint. (a) Training-time weight λmax: violations decrease monotonically with stronger guidance. (b) Inference-time strength ηmax: LGVF + Adj consistently outperforms FM + Adj across all settings. (c) Adjustment start time t0: earlier correction is slightly more effective, but performance is robust. (d) Component analysis: combining training and inference guidance achieve… view at source ↗

discussion (0)

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Incantation: Natural Language as the Action Interface for Multi-Entity Video World Models
cs.CV 2026-05 unverdicted novelty 7.0

Incantation is the first video world model to use per-frame natural language conditioning for simultaneous multi-entity control and concept-level cross-entity transfer in interactive video generation.
Incantation: Natural Language as the Action Interface for Multi-Entity Video World Models
cs.CV 2026-05 conditional novelty 7.0

Per-frame natural-language action prompts enable simultaneous multi-entity control and cross-entity action transfer in interactive video world models, outperforming discrete action-index interfaces.
Wasserstein Stability of Contracting Flows: Effective Rates, Euler Self-Correction, and Noise Tightening
eess.SY 2026-07 conditional novelty 4.0

A displacement-weighted effective contraction rate strictly improves worst-case Wasserstein bounds for nonlinear contracting flows, with new peak-time and variance-tightening predictions.

Reference graph

Works this paper leans on

9 extracted references · 8 linked inside Pith · cited by 2 Pith papers

[4]

Jonathan Ho and Tim Salimans

URLhttps://arxiv.org/ abs/1810.01367. Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS Workshop on Deep Generative Models and Downstream Applications,

Pith/arXiv arXiv
[7]

org/abs/2210.02747

URLhttps://arxiv. org/abs/2210.02747. Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations,

Pith/arXiv arXiv
[2004]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer

URLhttps://link.springer.com/book/10.1007/978-1-4757-4145-2. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion models. InConference on Computer Vision and Pattern Recognition,

doi:10.1007/978-1-4757-4145-2
[2019]

URLhttps://arxiv.org/abs/1901.10060. Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary dif- ferential equations. InAdvances in Neural Information Processing Systems (NeurIPS), 2018a. URLhttps://arxiv.org/abs/1806.07366. Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differ- enti...

Pith/arXiv arXiv 1901
[2020]

11 BAHERI Michael Janner, Yilun Du, Joshua B

URLhttps://arxiv.org/ abs/2006.11239. 11 BAHERI Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. InInternational Conference on Machine Learning (ICML),

Pith/arXiv arXiv 2006
[2021]

Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio

URLhttps://arxiv.org/ abs/2011.13456. Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative mod- els with minibatch optimal transport.Transactions on Machine Learning Research,

Pith/arXiv arXiv 2011
[2022]

Yaron Lipman, Ricky T

URLhttps://arxiv.org/abs/2205.09991. Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

Pith/arXiv arXiv
[2023]

Sherrie Christopher, Jens Behrmann, Yaron Lipman, and Ricky TQ Chen

URLhttps://arxiv.org/abs/2303.04137. Sherrie Christopher, Jens Behrmann, Yaron Lipman, and Ricky TQ Chen. Constrained diffusion models via dual training. InInternational Conference on Machine Learning,

Pith/arXiv arXiv
[2025]

Metriplectic conditional flow matching for dissipative dynamics

Ali Baheri and Lars Lindemann. Metriplectic conditional flow matching for dissipative dynamics. arXiv preprint arXiv:2509.19526,

Pith/arXiv arXiv

[1] [4]

Jonathan Ho and Tim Salimans

URLhttps://arxiv.org/ abs/1810.01367. Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS Workshop on Deep Generative Models and Downstream Applications,

Pith/arXiv arXiv

[2] [7]

org/abs/2210.02747

URLhttps://arxiv. org/abs/2210.02747. Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. InInternational Conference on Learning Representations,

Pith/arXiv arXiv

[3] [2004]

Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer

URLhttps://link.springer.com/book/10.1007/978-1-4757-4145-2. Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High- resolution image synthesis with latent diffusion models. InConference on Computer Vision and Pattern Recognition,

doi:10.1007/978-1-4757-4145-2

[4] [2019]

URLhttps://arxiv.org/abs/1901.10060. Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary dif- ferential equations. InAdvances in Neural Information Processing Systems (NeurIPS), 2018a. URLhttps://arxiv.org/abs/1806.07366. Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differ- enti...

Pith/arXiv arXiv 1901

[5] [2020]

11 BAHERI Michael Janner, Yilun Du, Joshua B

URLhttps://arxiv.org/ abs/2006.11239. 11 BAHERI Michael Janner, Yilun Du, Joshua B. Tenenbaum, and Sergey Levine. Planning with diffusion for flexible behavior synthesis. InInternational Conference on Machine Learning (ICML),

Pith/arXiv arXiv 2006

[6] [2021]

Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio

URLhttps://arxiv.org/ abs/2011.13456. Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector- Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative mod- els with minibatch optimal transport.Transactions on Machine Learning Research,

Pith/arXiv arXiv 2011

[7] [2022]

Yaron Lipman, Ricky T

URLhttps://arxiv.org/abs/2205.09991. Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747,

Pith/arXiv arXiv

[8] [2023]

Sherrie Christopher, Jens Behrmann, Yaron Lipman, and Ricky TQ Chen

URLhttps://arxiv.org/abs/2303.04137. Sherrie Christopher, Jens Behrmann, Yaron Lipman, and Ricky TQ Chen. Constrained diffusion models via dual training. InInternational Conference on Machine Learning,

Pith/arXiv arXiv

[9] [2025]

Metriplectic conditional flow matching for dissipative dynamics

Ali Baheri and Lars Lindemann. Metriplectic conditional flow matching for dissipative dynamics. arXiv preprint arXiv:2509.19526,

Pith/arXiv arXiv

Logic-Guided Vector Fields for Constrained Generative Modeling

Core claim §

What carries the argument

Load-bearing premise §

What would settle it §

If this is right

Where Pith is reading between the lines

discussion (0)

Core claim

Load-bearing premise

What would settle it