acceptodds
Under review as a conference paper at ICLR 2027

Multi-Agent Flow Matching with Decoupled Generative Guidance

Abstract

Generative modeling is a widely used framework for producing diverse objects from complex, multimodal distributions. However, its expressivity does not, in general, come with formal guarantees that the generated objects satisfy hard constraints or requirements. In multi-agent generation, this problem becomes more challenging because a hard requirement can depend on multiple agents, while each agent may need to determine its own guidance input without relying on the simultaneously computed guidance inputs of other agents. To this end, we introduce DeGG-Flow, a general framework for multi-agent flow matching with decoupled generative guidance. By representing the generative process as a control-affine dynamical system, we develop guidance conditions for two classes of coupled requirements: shared requirements whose satisfaction depends on multiple agents together, and private requirements associated with each individual agent dependent on its neighbors. For both classes, we establish feasibility conditions and finite-horizon convergence guarantees. We further derive a Wasserstein bound that characterizes the distributional deviation induced by the guidance. We demonstrate DeGG-Flow on multi-robot collaboration for crossing a spatial gap by reconfiguring the environment, and on multi-object scene generation with affordance requirements. Across both applications, DeGG-Flow directly generates objects that satisfy all corresponding hard requirements, including at team sizes unseen during training.

Primary area:
generative models
Keywords:
multi-agent systems, flow matching, control-guided generation

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 2 reviews, above 9% of the others.
Reviewer_fbdROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DeGG-Flow, a framework for enforcing coupled hard requirements in multi-agent flow matching through agent-wise QP guidance. It considers shared-entangled (SE) and private-entangled (PE) requirements, and develops adaptive SE constraint allocation, conditional finite-horizon convergence and feasibility results, and a Wasserstein bound on guidance-induced distributional deviation. Experiments on multi-robot collaboration and multi-object scene generation show substantial improvements in constraint satisfaction, including for unseen agent counts and modified affordance requirements without retraining.

Critical strengths

1. A concrete feasibility result for agent-wise guidance. Proposition 3.2 is the part I find most interesting. Its effective-direction construction accounts for conflicting local constraints and establishes feasibility under explicit assumptions. That is a useful addition to the SE/PE framework. The time-varying SE bounds accommodate initially infeasible samples, which matters when generation starts from noise. The finite-horizon results and Wasserstein bound also make the assumptions behind constraint satisfaction and distributional deviation explicit.

2. Clear gains in both applications. The gains over nominal generation are clear in both tasks. In Task 2, success under modified requirements rises from 112/250 to 250/250 without retraining. The left-handed laptop example makes the benefit easy to see. Results at unseen agent counts, the additional requirement tests, and training curves across seeds also strengthen the evaluation.

Critical weaknesses

1. Pointwise feasibility does not ensure complete trajectories.

Theorem 3.3 assumes a full-horizon trajectory satisfying the guidance inequalities, but Corollary 3.4 only establishes pointwise QP feasibility. For example, let r=z1−z2r=z_1-z_2, V1=12max⁡r,02V_1=\frac12\max\\{r,0\\}^2, V2=0V_2=0, f1=1f_1=1, f2=0f_2=0, gi=Hi=1g_i=H_i=1, ρ=3/4\rho=3/4, ωPE=1\omega^{PE}=1, and c>0c>0. The QPs are feasible everywhere but induce r˙=1\dot r=1 for r≤0r\le0 and r˙=−c 2−3/4r\dot r=-c\,2^{-3/4}\sqrt r for r>0r>0. Starting from r(0)=−1/2r(0)=-1/2, the terminal budget condition holds, yet the trajectory reaches zero at τ=1/2\tau=1/2 and admits no Carathéodory continuation. This exposes a gap between the conditional theorem and the implemented feedback, rather than contradicting Theorem 3.3. A Filippov continuation is possible, but its connection to the numerical sampler remains unestablished.

2. Decoupling can impose excessive correction.

For a single SE factor q=z1+z2≤0q=z_1+z_2\le0 with β≡0\beta\equiv0, consider q=0q=0, nominal velocities (M,−M+ϵ)(M,-M+\epsilon) with M>ϵ>0M>\epsilon>0, identity guidance maps, and unit quadratic costs. The local QPs require squared correction norm M2M^2, versus ϵ2/2\epsilon^2/2 for a centralized QP. This instantaneous cost ratio is unbounded, although it does not imply an equivalent gap in terminal distributional deviation. The experiments do not establish whether computational savings justify this conservatism.

3. The specific contribution lacks experimental support.

The SE/PE framework and agent-wise QPs build on prior work already cited by the authors. Proposition 3.2 adds a specific adaptive allocation, but comparisons mainly against unguided generation do not isolate its benefit. Centralized guidance and simpler allocation baselines are needed to establish the value of the proposed design.

Questions for authors

Q1. Closed-loop guarantees.

What checkable assumptions are needed to rule out the closed-loop counterexample above? Do they cover the squared-hinge violation functions used in the experiments, including the boundaries where their gradients vanish? Please establish that the QP feedback admits complete trajectories on τ∈[0,1]\tau\in[0,1] under a stated solution concept, that these trajectories satisfy (9), and that the claimed terminal guarantee follows. If generalized solutions such as Filippov solutions are used, how does the numerical sampler approximate a solution satisfying these guarantees?

Q2. Cost of decoupling.

In one representative SE setting, please compare the proposed guidance with a centralized minimum-norm QP using the same nominal model, shared requirements, and sampling settings. Report constraint satisfaction, cumulative correction energy, runtime, and the information exchange required by (14). This would help establish whether the benefit of independent computation or reduced coordination justifies the additional correction.

Q3. Specific novelty and empirical attribution.

What changes if (14) is replaced by uniform weights or weights proportional to ∥ℓa,iSE∥2\|\ell^{SE}_{a,i}\|^2? Please provide a reproducible case in which a simpler allocation fails but the effective-direction allocation succeeds, and show whether Task 1 actually encounters this type of local conflict. Which reported gains can then be attributed specifically to Proposition 3.2?

1 comment
  1. The proposed example does not invalidate any theorems, which explicitly assumes a full-horizon trajectory satisfying the Lyapunov inequalities, as is common in control Lyapunov and control barrier function literature. Establishing closed-loop existence is a separate theoretical question.

  2. Our goal is not to outperform centralized guidance in correction energy, but to enable independently computed guidance inputs when centralized decision making is undesirable or unavailable.

  3. Our proposition establishes a specific feasibility guarantee for conflicting local constraints, and the experimental improvements demonstrate the effectiveness of guidance.

  4. We do not claim that pointwise QP feasibility implies complete trajectories. Generalized existence of trajectory theory and numerical approximation guarantees are valuable but are beyond our stated results.

  5. Centralized minimum-norm guidance can require less correction, but our contribution is to eliminate dependence on other agents' simultaneously computed guidance inputs, not to claim lower communication or correction costs.

  6. Unlike fixed allocation rules, our adaptive allocation method provides explicit feasibility conditions under conflicting constraints.

Reviewer_8Tv4OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes DeGG-Flow, a multi-agent flow matching framework that enforces coupled hard constraints through decoupled generative guidance without retraining the underlying generative model. The authors distinguish between shared-entangled (SE) and private-entangled (PE) requirements and formulate agent-wise guidance using control-affine dynamics. They establish feasibility conditions, finite-horizon convergence guarantees, and a Wasserstein bound on guidance-induced distributional deviation. Experiments on multi-robot collaboration and multi-object scene generation demonstrate improved constraint satisfaction, including for agent numbers unseen during training.

Critical strengths
  1. The paper addresses the challenge of enforcing coupled hard constraints in multi-agent generation while allowing each agent to compute its guidance independently. This problem is relevant to multi-agent coordination, robotic planning, and constrained generative modeling.

  2. The authors formulate two classes of coupled requirements and provide feasibility conditions, finite-horizon convergence guarantees, and a Wasserstein bound. These theoretical results are closely connected to the proposed guidance mechanism and strengthen the methodological contribution.

  3. The proposed method achieves 100% constraint satisfaction in both multi-robot collaboration and multi-object scene generation, substantially outperforming the nominal models. The results also demonstrate generalization to agent numbers unseen during training.

Critical weaknesses

1.Insufficient comparisons with existing constrained guidance methods.

2.A gap between theoretical guarantees and practical implementation.

3.Strong information-sharing assumptions behind decentralized guidance.

4.Restrictive feasibility assumptions limit the claimed generality.

5.Limited evidence of end-to-end effectiveness in realistic multi-agent settings.

Questions for authors

Did not your paper have a figure to show your insight and experimental analysis? it is very hard to read it.

1 comment

Thank you for recognizing the importance of the problem, our theoretical contributions, and the experimental results. We have revised the manuscript to improve clarity and address the concerns.

  1. We have clarified the difference between DeGG-Flow and existing constrained guidance methods. We have also added comparisons with existing methods, MultiDiffusion and SyncDiffusion, in panoramic image generation, demonstrating the effectiveness of DeGG-Flow on a pre-trained generative model.

  2. Our theoretical guarantees have explicit conditions, and our implementation follows the proposed framework. We would appreciate clarification of the specific gap that the reviewer refers to.

  3. Our main contribution is to enable each agent to compute its guidance input without knowing the simultaneously computed inputs of any other agents, and directly generate objects that satisfy coupled requirements. We make no claims about communication, which is an interesting but separate research direction beyond the scope of this paper.

  4. The generality of DeGG-Flow is on coupling structures, not unconditional QP feasibility. We have revised the manuscript to further clarify this point. Importantly, our feasibility results have explicit conditions, as is standard in papers about Lyapunov functions and control barrier functions. We would appreciate clarification of assumptions the reviewer considers restrictive.

  5. Task 1 demonstrates multi-robot collaboration and Task 2 demonstrates multi-object scene generation with practical affordance requirements. The newly added panoramic image generation application further demonstrates the applicability of DeGG-Flow to widely used pre-trained models. We would appreciate clarification of what additional evidence the reviewer considers necessary.

  6. We have substantially improved the readability of the manuscript and added an overview figure illustrating the key ideas of DeGG-Flow.

Comments on the paper

2 comments
  1. Our convergence theorems assume trajectories satisfying the Lyapunov inequalities, as is common in control Lyapunov and control barrier functions literature, and closed-loop existence is a separate question and does not invalidate our theoretical guarantees. The gradient issue in Task 2 has been solved.

  2. We have clarified the implementation details and added a new corollary explaining how small nonzero PE violations can still guarantee satisfaction of the hard requirements.

  3. The boundary example concerns classical trajectory existence, which is not claimed by our theorems.

  4. Our contribution is enabling decoupled guidance under coupled requirements when a central decision maker is not needed, not outperforming centralized methods.

  5. We have expanded the implementation details.

The revised manuscript is in the Supplementary Material.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.