Divergence-Suppressing Couplings for Rectified Flow
Yimeng Min, Carla P. Gomes
Abstract
The promise of Rectified Flow rests on producing self-generated couplings whose trajectories are straight, or nearly so. In practice, trajectories generated by the base flow model can bend and intertwine, and the resulting coupling inherits this distortion. In this paper, we identify that such trajectory entanglement is often associated with regions of nonzero divergence in the learned velocity field, where local expansion or contraction distorts trajectories and steers particles away from their ideal endpoints. We then propose divergence-suppressing couplings for Rectified Flow, an offline correction that attenuate the divergent component of the learned velocity during coupling generation. The correction is paid only once per coupling pair and amortized over training, so deployment runs plain Euler at identical wall-clock cost to standard Rectified Flow. Empirically, this offline modification yields consistent improvements on 2D synthetic benchmarks and on image generation.
AI Impact Assessments
(1 model)Scientific Impact Assessment: Divergence-Suppressing Couplings for Rectified Flow
1. Core Contribution
The paper identifies that coupling distortion in Rectified Flow (RectFlow) is mechanistically linked to regions of nonzero divergence in the learned velocity field, where local expansion or contraction bends trajectories and degrades the quality of self-generated coupling pairs. The proposed remedy, DS-RectFlow, applies a divergence-suppressing correction *only during offline coupling generation*, leaving training objectives, model architectures, and inference procedures entirely unchanged. The key insight is that this is a one-time offline cost amortized over training, so deployment runs plain Euler at identical wall-clock cost to standard RectFlow.
The method works by either (a) in low dimensions, projecting the velocity to reduce its divergent component via a first-order gradient correction, or (b) in high dimensions, performing a zeroth-order search among perturbed candidate states to find nearby positions with lower estimated divergence (via Hutchinson's trace estimator) before taking each Euler step.
2. Methodological Rigor
Strengths in methodology:
Weaknesses in methodology:
3. Potential Impact
The paper addresses a genuine practical bottleneck in RectFlow: the observation that vanilla rectification saturates quickly because coupling quality is limited by the base model's integration quality. If the approach generalizes reliably, it could become a standard preprocessing step in flow-matching pipelines.
However, the empirical evaluation is limited in scope:
The authors suggest applicability to stochastic interpolants, OT-FM, and score-based diffusion, but provide no evidence for these claims.
4. Timeliness & Relevance
The paper is timely. Flow matching and rectified flow are active research areas driving modern generative models (Stable Diffusion 3, Flux). Reducing NFE requirements is a genuine engineering need. The diagnosis that coupling quality—not model capacity or training loss—is the bottleneck is a valuable conceptual contribution that could redirect research effort.
Two concurrent works (FDS by Cha et al. 2026, FDM by Huang et al. 2026) independently identify divergence as important, validating the general direction. The differentiation from these works (offline-only correction, mechanistic framing) is clearly articulated.
5. Strengths & Limitations
Key Strengths:
Notable Limitations:
Summary
This paper makes a clear conceptual contribution by connecting velocity field divergence to coupling distortion in RectFlow and proposing an offline remedy. The mechanism is well-motivated and the experiments show consistent improvements within the tested scope. However, the evaluation is limited to small-scale benchmarks, the method's scalability to modern generative modeling settings is undemonstrated, and the zeroth-order correction strategy lacks theoretical guarantees. The work is a solid incremental contribution to the flow matching literature but falls short of demonstrating transformative impact.
Generated May 19, 2026
Comparison History (18)
Paper 1 addresses the critical and highly timely issue of LLM safety during fine-tuning. Its counterintuitive 'jailbreak to protect' framework is highly innovative and directly applicable to widespread commercial Fine-tuning-as-a-Service platforms. While Paper 2 provides solid methodological improvements to generative modeling, Paper 1's immediate relevance to real-world AI deployment, societal safety, and the broader AI community gives it a higher potential for broad scientific impact.
Paper 2 addresses the critical bottleneck of KV cache memory explosion in Large Language Models for long-context tasks. By dynamically generating Soft Tokens and preserving evicted information, it offers a highly timely and practical solution with broad, immediate real-world applications across NLP and industry. While Paper 1 provides a solid methodological improvement for Rectified Flow generative models, Paper 2's focus on LLM efficiency gives it a higher potential for widespread adoption and immediate scientific impact.
Paper 2 likely has higher impact: it introduces a principled, broadly applicable modification to Rectified Flow grounded in field properties (divergence), with clear computational advantages (offline correction, no deployment overhead) and demonstrated gains on synthetic and image-generation tasks. This targets a timely core generative-modeling paradigm and could influence diffusion/flow training practices across domains. Paper 1 is valuable infrastructure (a large benchmark + metrics) but is narrower (programmatic video/code evaluation) and its impact depends on community adoption and alignment with emerging generation paradigms.
Paper 2 addresses a critical and highly timely issue—privacy and security in LLM agents. By introducing a comprehensive benchmark for privacy-utility trade-offs, it provides a crucial tool for AI safety that will likely drive widespread future research and model improvements. While Paper 1 offers a valuable algorithmic enhancement for generative models, Paper 2's focus on LLM privacy has broader real-world applications, cross-disciplinary relevance, and addresses urgent security vulnerabilities in widely deployed systems.
Paper 2 addresses a fundamental issue in Rectified Flow—a widely adopted generative modeling framework—by identifying trajectory entanglement linked to divergence in velocity fields and proposing a principled, computationally efficient correction. This has broad applicability across generative modeling (images, video, etc.) and offers a clean theoretical insight with practical gains at no deployment cost. Paper 1 contributes a useful Bayesian optimization framework for prompt engineering, but operates in a narrower, more applied niche (system prompt tuning) with less fundamental methodological novelty and more limited cross-field impact.
Paper 2 targets a timely, high-visibility area (diffusion/flow-based generative modeling) and proposes a broadly applicable, deployment-neutral correction that can improve sample quality without increasing inference cost—strong real-world relevance and cross-use in generative modeling. The divergence-based diagnosis is conceptually clean and may generalize to other ODE-based models. Paper 1 is novel in formalizing trust calibration as preference learning, but its impact may be narrower (human-in-the-loop agent governance) and leans on established GP classification/PBO machinery with less clear empirical validation from the abstract.
Paper 1 bridges a critical gap by transitioning machine learning applications from controlled lab settings to noisy, real-world educational environments for a highly vulnerable population. Its capacity to predict challenging behaviors 10 minutes in advance offers profound, immediate real-world clinical and societal applications. While Paper 2 presents a solid technical refinement for generative models, Paper 1 demonstrates higher cross-disciplinary impact, translating foundation models into actionable interventions that directly improve human safety and quality of life.
Paper 2 addresses a fundamental issue in Rectified Flow, a widely-used generative modeling framework, with a theoretically grounded and practical solution (divergence-suppressing couplings). It identifies a specific failure mode (trajectory entanglement linked to divergence) and proposes an elegant fix with no additional deployment cost. This has immediate applicability to the large and active generative modeling community. Paper 1, while useful, introduces a benchmark for a relatively niche subproblem (skill generation for LLM agents) that is still emerging. Benchmarks typically have lower impact than methodological innovations unless they become widely adopted standards.
AGPO addresses a fundamental limitation of RLVR methods—reasoning boundary shrinkage—with a novel asymmetric reinforcement strategy that has both theoretical insight and demonstrated practical impact. It shows state-of-the-art results across five math benchmarks and includes large-scale industrial deployment at JD for search ads relevance, demonstrating real-world applicability. Paper 1 offers a useful but incremental improvement to Rectified Flow via divergence suppression, which is narrower in scope and impact. Paper 2's broader relevance to the rapidly growing LLM reasoning field and its dual academic/industrial validation give it higher potential impact.
Paper 1 addresses a fundamental and broadly relevant problem—reproducibility of LLM-assisted scientific analysis—with a practical, deployable solution (typed mediation) validated over six months of real-world use across multiple instruments. It bridges AI and experimental science with clear real-world applications, reducing analysis from weeks to minutes while guaranteeing deterministic outputs. Paper 2 offers a useful but incremental improvement to Rectified Flow training, a narrower contribution within generative modeling. Paper 1's cross-disciplinary impact, practical deployment evidence, and relevance to the reproducibility crisis give it higher potential impact.