{"id":"82b44bfa-6e6c-4036-9505-21485241c44e","arxiv_id":"2502.02150","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A unified derivation of energy guidance for general flow matching yields an asymptotically exact Monte Carlo method, approximate gradient methods, and training losses that recover DPS, LGD, and PiGDM as special cases.","lead":"This paper proposes a general way to steer flow matching generative models toward desired outcomes by adding a guidance vector field, and derives several practical guidance methods from it. A smart generalist should read it because it unifies many existing diffusion and flow guidance techniques under one theory and tests them on images and offline reinforcement learning.","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The P=1 approximation is load-bearing: for dependent couplings it alters the effective source distribution, so Theorem 3.1's exact target is only reached from a modified base distribution, and gMC's unbiasedness claim inherits this gap.","rationale":"The reader identified P=1 as the weakest assumption; I agree. The central theorem itself is a formal identity and is not internally inconsistent; the issue is that the guidance problem requires the modified vector field, started at the original base distribution, to yield the energy-weighted target. With a dependent coupling and P=1, the modified joint p'(z)=π(x0|x1)Z^{-1}p(x1)e^{-J} has a t=0 marginal that is not p0, so the flow started from p0 does not follow p'_t. This affects gMC—advertised as asymptotically exact for arbitrary source distributions—because Eq. (3) is exactly the P=1 version. Table 3 compares ∥v_OT−v_CFM∥, which is necessary but not sufficient: P is multiplied by e^{-J}, so a small deviation in P can be amplified by a strongly varying J. The concrete test directly estimates P and compares the guided distributions with and without it. If the test passes, the concern does not land and the conditional verdict can be upgraded; if it fails, the central claim should be restricted to independent couplings or accompanied by a corrected source distribution. The paper already flags the strong-coupling limitation in Section 5, so a rejection is not warranted; a conditional verdict with the requested P analysis is appropriate. Since this matches the reader's conditional verdict, the verdict is unchanged.","tokens_in":42604,"tokens_out":10826,"duration_ms":109394,"concrete_test":"On the synthetic 2D OT-CFM setup from §4.1 with a dependent coupling, choose a sharply varying J. Estimate π(x0|x1) from the mini-batch OT paired samples (e.g., by kernel density on (x0,x1) pairs). Fix the source by setting π'(x0|x1)=p0(x0), compute P=p0(x0)/π(x0|x1) along guided trajectories, and solve the ODE with Eq. (1) using this P. Compare the terminal distribution in Wasserstein-2 to p(x1)e^{-J}/Z against the P=1 version. If the P-corrected run is materially closer while P deviates from 1, the P=1 approximation is the limiting assumption; if P≈1 along the trajectories, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The theorem is a correct marginalization identity for the path p'_t(x_t)=∫p_t(x_t|z)π'(x0|x1)Z^{-1}p(x1)e^{-J(x1)}dz. But the guidance problem is only solved if the modified vector field, initialized from the original source p0, produces the energy-weighted target. Setting P=1 corresponds to choosing π'=π, so the induced source at t=0 is ∫π(x0|x1)Z^{-1}p(x1)e^{-J(x1)}dx1, which equals p0 only when the coupling is independent. All implemented methods (gMC, gcov, gsim, gϕ) run with P=1; for dependent couplings such as mini-batch OT, Theorem 3.1's exactness therefore does not apply to sampling started from p0. Appendix A.3 acknowledges this but validates P≈1 only by comparing OT-CFM and CFM vector fields (Table 3), not by measuring P or the source mismatch. Moreover, Section 3.2 calls gMC 'unbiased and applicable to any source distribution'; the self-normalized estimator is only asymptotically unbiased even under P=1. The load-bearing gap is not the algebra of Theorem 3.1 but the unverified identification P≈1 in the strong-coupling, strong-J regime that Section 5 itself flags.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified framework for energy-guided sampling with flow matching. Starting from conditional flow matching, Theorem 3.1 expresses the guidance field as a difference of conditional vector fields averaged under the reweighted coupling, with a reverse-coupling ratio P. The paper then sets P≈1 and derives: a training-free Monte Carlo guidance method gMC (Algorithm 1); localized approximations glocal/gcov and gsim that recover DPS, LGD, ΠGDM, and OT-ODE as special cases; and training-based guidance losses gϕ. Experiments cover 2D synthetic flows with non-Gaussian sources, D4RL locomotion planning, and CelebA-HQ inverse problems. Full proofs and an error bound for the localized approximation are provided in the appendices, and code is released.","tokens_in":1740,"tokens_out":1835,"duration_ms":78756,"significance":"If the framework holds as stated, this is a substantial unification: it would justify guidance for non-Gaussian source distributions, dependent couplings, and general conditional paths, and it would place several established diffusion guidance methods as special cases. The paper's strengths are the clean marginalization identity in Theorem 3.1, the explicit recovery of classical methods in Appendices A.4 and A.13, an error bound for glocal in Appendix A.9, extensive appendices, and publicly available code. The central claims are tempered by two substantial caveats: the P=1 approximation changes the source distribution for dependent couplings, and the finite-sample unbiasedness of the self-normalized Monte Carlo estimator is not correct as stated. These caveats affect the interpretation of the OT-CFM experiments and the ``applicable to any source distribution'' claims, but the core identity remains a sound construction for the reweighted-path vector field.","major_comments":[{"comment":"Theorem 3.1 is a correct identity for the path p'_t(x_t)=∫p_t(x_t|z)π'(x0|x1)Z^{-1}p(x1)e^{-J(x1)}dz, but the guidance problem is solved by integrating v'_t from the original source p0 only when p'_0(x0)=p0(x0). Setting P=1 corresponds to π'=π, which gives p'_0(x0)=∫π(x0|x1)Z^{-1}p(x1)e^{-J(x1)}dx1, and this equals p0 only when the coupling is independent. All implemented methods (gMC, gcov, gsim, gϕ) run with P=1, so for mini-batch OT couplings Theorem 3.1 does not by itself establish exactness for sampling started from p0. Appendix A.3 acknowledges this, but Table 3 validates P≈1 only indirectly by comparing the vector fields of OT-CFM and CFM, not by measuring P or the source mismatch, and Section 5 states that the guidance direction is no longer correct under strong coupling and strong J. The abstract and Section 1 should be scoped accordingly, or the paper should directly measure P or the induced source mismatch in the strong-coupling regime.","section":"§3.1, Theorem 3.1; Appendix A.3"},{"comment":"Section 3.2 states that gMC is ``unbiased and applicable to any source distribution.'' Algorithm 1, however, uses a self-normalized estimator: both \\itilde Z_t and \\itilde p_t are Monte Carlo estimates formed from the same samples, and gMC is a ratio of these estimates. A self-normalized importance-sampling estimator is generally biased at finite N and only asymptotically unbiased under standard regularity conditions. This affects the ``asymptotically exact'' claim and the interpretation of the scaling experiment in Figure 4. Please replace ``unbiased'' with ``asymptotically unbiased'' or provide a genuinely unbiased estimator, for example by using independent sample sets for the numerator and denominator or by using a known normalizing constant pt(xt).","section":"§3.2, Algorithm 1"},{"comment":"Section 3.5 claims that the losses ℓVGM, ℓRGM, and ℓMRGM produce the same minimizer as ℓGM and that all provide unbiased gradient estimates, but the proof in Appendix A.7 relies on the assumption in Eq. (51) that vθ(xt,t)=E_{z∼p(z|xt)}[v_{t|z}(xt|z)], i.e., a perfectly learned flow-matching vector field. For a practical pre-trained model with approximation error, the gradients of ℓVGM and ℓRGM differ from that of ℓGM by terms involving vθ−vt. This assumption should be stated in the main text, and the equivalence claims should be qualified or supplemented with an error bound that depends on the flow-matching training error.","section":"§3.5, Appendix A.7"}],"minor_comments":[{"comment":"The sentence introducing Table 3 contains a typo: ``the VF of the OT-CFM (batch size 128) in and uncoupled CFM trained'' should read ``... in and uncoupled CFM trained...'' or similar, and Table 3 is not cited at the point in Section 3.1 where P=1 is first introduced.","section":"Appendix A.3, Table 3"},{"comment":"The notation E_{x1,x0∼p(z)} is confusing because p(z) is a joint coupling π(x0,x1); please write E_{(x0,x1)∼π} or equivalent so that the coupling is explicit.","section":"§3.2, Eq. (3)"},{"comment":"In Algorithm 1 the symbol \\itilde p_t is used both for the estimated marginal density and for the normalized importance weights; a short comment or a distinct variable name would clarify the self-normalization step.","section":"Algorithm 1"},{"comment":"The caption is confusing: ``the guided generation errors (crosses) do not converge to the dashed lines is that they measure the W2 distance of p(x1) and p(x1)e^{−J(x1)}/Z'' mixes quantities. Please state explicitly which two distributions are compared for the dashed lines and which two for the crosses.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the central marginalization identity is correct. The main risk is overclaiming exactness for dependent couplings and for finite-sample Monte Carlo guidance; both issues are fixable by scoping the claims and adding a direct empirical check of P or the source mismatch. I do not see attribution or novelty concerns that would require rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper delivers a real result: a single guidance identity (Theorem 3.1) that covers DPS, LGD, PiGDM, and OT-ODE, and gives a principled route to guidance for non-Gaussian sources and dependent couplings. The derivation is coherent under the stated definitions, and the new MC estimator and training losses are natural consequences, not afterthoughts. The experiments on synthetic, D4RL, and CelebA-HQ give a fair picture of where each method works; the synthetic results showing gMC and gphi matching the energy-weighted ground truth are genuinely convincing. The discussion of prior work, including the overlap with Zhang et al., is honest. The soft spots are in the claims around exactness. The P=1 approximation is load-bearing. For dependent couplings, setting P=1 means the guided flow effectively starts from a modified base distribution, not the original p0. The paper acknowledges this in Section 5, but Appendix A.3 validates P≈1 only by comparing OT-CFM and CFM vector fields, not by measuring P or the induced source mismatch. For strong coupling with rapidly varying J, that is exactly the regime where the approximation fails, and the paper says so. That is an honest limitation, but it undercuts the 'any source, any coupling' framing if not flagged clearly in the abstract and in Section 3.1. Similarly, gMC is described as unbiased, but the self-normalized ratio estimator in Algorithm 1 is only asymptotically unbiased. The paper uses 'asymptotically exact' in the abstract and then 'unbiased' in Section 3.2. That is a minor but real overclaim. The VGM/RGM loss equivalence also requires the learned vector field to be exact, which the main text does not flag; it is buried in the appendix proof. These are fixable with corrected wording, not fatal objections. The image experiments lack error bars; given that table entries are close in places, that matters for the ranking claims. Not a dealbreaker, but worth noting. Bottom line: this is a serious paper by people who know the area. The unified framework is valuable and the central theorem holds up. The exactness claims need to be pulled back to match what is actually shown. I'd send it to referees, and I'd expect them to ask for a more careful treatment of P and the MC estimator. For a reader working on flow matching guidance, it is worth citing now.","headline":"A genuinely unifying framework for flow-matching guidance, with a load-bearing approximation that needs to be stated more carefully before the exactness claims are accepted.","tokens_in":705,"tokens_out":1729,"would_cite":true,"duration_ms":36254,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Flow matching guidance reduces to a single explicit vector-field identity that covers exact, learned, and classical diffusion methods.","keywords":["flow matching","energy guidance","guidance vector field","Monte Carlo estimation","training-based guidance","diffusion guidance","offline reinforcement learning","generative modeling"],"falsifier":"On a low-dimensional dataset with strongly dependent mini-batch OT couplings and a sharply varying energy $J$, guide samples with $g^{MC}$ using many Monte Carlo samples and compare the terminal distribution to $p(x)e^{-J(x)}/Z$; a mismatch beyond sampling error shows that the $P=1$ premise fails for dependent couplings.","tokens_in":42360,"feed_emoji":"🎯","tokens_out":8308,"duration_ms":77544,"temperature":0.7,"pith_summary":"This paper claims that steering a flow-matching generative model—one that pushes a source distribution to a target by following a learned vector field—toward the energy-weighted distribution $p'(x) \\propto p(x) e^{-J(x)}$ reduces to adding one explicit guidance vector field to the learned field. The identity, Theorem 3.1, applies to arbitrary source distributions, couplings, and conditional paths, not only to the Gaussian-source setting where flow matching becomes diffusion. If the identity is correct, many classical diffusion guidance methods are special cases of one formula, and new training-free and training-based guidance methods follow from it. The paper tests the framework on synthetic distributions, image inverse problems, and offline reinforcement learning planning.","feed_headline":"One formula guides any flow-matching model to a target distribution","feed_subtitle":"A single guidance vector-field identity unifies Monte Carlo, learned, and classical diffusion guidance for flow matching.","key_machinery":"The carrying object is the guidance vector field identity itself: a closed-form expression for the additive field that reweights the terminal distribution by the Boltzmann factor. The mechanism that keeps the identity tractable is the reverse coupling ratio $P = \\pi'(x_0|x_1)/\\pi(x_0|x_1)$, which is exactly 1 for independent couplings; every practical method in the paper operates under the $P=1$ approximation. Under the affine Gaussian uncoupled path, conditional vector fields become conditional scores, which is what turns the identity into the gradient guidance formula used in diffusion models.","core_discovery":"At the center is the exact guidance identity of Theorem 3.1: adding $g_t(x_t)$ to the original vector field $v_t(x_t)$ generates the path ending at $p'(x)$, where $g_t(x_t) = \\int \\big(P e^{-J(x_1)}/Z_t(x_t) - 1\\big) v_{t|z}(x_t|z) p(z|x_t) dz$ and $Z_t(x_t) = \\int P e^{-J(x_1)} p(z|x_t) dz$, with $P$ the reverse coupling ratio. Under the uncoupled affine Gaussian path assumption the identity collapses onto the familiar diffusion guidance $\\nabla_{x_t}\\log Z_t(x_t)$ up to a schedule factor. From this single identity the paper derives asymptotically exact Monte Carlo guidance ($g^{MC}$), learned guidance ($g^{\\phi}$), localized Taylor guidance ($g^{local}$, $g^{cov}$), and Gaussian-approximation guidance ($g^{sim}$), each covering previously known diffusion methods as special cases.","pith_inferences":["The framework leaves open the possibility of choosing the reverse coupling deliberately rather than approximating it, which would make exact guidance available for optimal-transport and rectified flows.","Using a learned approximate guidance as the proposal in $g^{MC}$ importance sampling (the paper's $g^{MC\\text{-}IS}$ sketch) is a natural way to make asymptotically exact guidance practical in high dimensions.","If exact coupling-aware guidance becomes practical, flows with straighter paths, such as mini-batch OT or rectified flows, would bring guided sampling closer to one-step accuracy than Gaussian-source diffusion guidance can."],"forward_implications":["Any pre-trained flow matching model with a non-Gaussian source or dependent coupling can be guided without retraining by $g^{MC}$, which is asymptotically exact as the Monte Carlo sample count grows.","Learned guidance $g^{\\phi}$ can be trained with simple conditional losses whose gradients match an intractable marginal loss, giving exact training-based guidance for general flow matching.","Classical diffusion guidance methods—DPS, LGD, $\\Pi$GDM, and OT-ODE—are recovered as special cases, so their validity extends to flow matching paths beyond the uncoupled affine Gaussian setting.","Approximate gradient guidance is reliable when the energy is smooth and the flow time is close to 1, while $g^{MC}$ excels in low-dimensional or frequently conditioned tasks and fails on image inverse problems because of Monte Carlo variance."],"supporting_citations":[{"why":"supplies the conditional flow matching construction and the marginal vector field formula Theorem 3.1 builds on","marker":"Lipman et al., 2023"},{"why":"defines mini-batch optimal transport conditional flow matching, the dependent-coupling setting where the paper sets P=1","marker":"Tong et al., 2024"},{"why":"the diffusion classifier guidance that the framework recovers under the uncoupled affine Gaussian path","marker":"Dhariwal & Nichol, 2021"},{"why":"DPS, the gradient posterior sampler recovered by the gcov-G guidance approximation","marker":"Chung et al., 2023"},{"why":"LGD, covered by the gsim-MC estimate of the normalizing constant","marker":"Song et al., 2023b"},{"why":"ΠGDM, covered by gsim-inv and used as the strongest image inverse-problem baseline","marker":"Song et al., 2023a"},{"why":"contrastive energy guidance baseline in the synthetic experiments and an alternative way to train Zt","marker":"Lu et al., 2023"},{"why":"its training-based flow matching guidance loss coincides with the paper's MRGM loss","marker":"Zhang et al., 2025"},{"why":"provides the x1-parameterization used to estimate x̂1 from the learned vector field under affine paths","marker":"Lipman et al., 2024"}],"fun_headline_variants":["Unified guidance identity for flow matching","General flow matching guidance, exact to approximate","First framework for flow matching guidance","Flow matching guidance unified by one exact identity","One identity guides any flow-matching model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reverse coupling ratio $P$, the ratio between the guided and unguided distributions of the source given a target point, is 1; exact for independent couplings, this is only an approximation for dependent couplings with rapidly varying energy.","fun_headline_variants_meta":{"raw":{"variants":["Unified guidance identity for flow matching","General flow matching guidance, exact to approximate","First framework for flow matching guidance","Flow matching guidance unified by one exact identity","One identity guides any flow-matching model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001057,"raw_usage":{"total_tokens":4440,"prompt_tokens":956,"completion_tokens":3484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":3421}},"tokens_in":572,"tokens_out":3484,"duration_ms":23214,"temperature":1.0,"reasoning_tokens":3421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T13:08:09.965241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a low-dimensional dataset with strongly dependent mini-batch OT couplings and a sharply varying energy $J$, guide samples with $g^{MC}$ using many Monte Carlo samples and compare the terminal distribution to $p(x)e^{-J(x)}/Z$; a mismatch beyond sampling error shows that the $P=1$ premise fails for dependent couplings.","supporting_citations":[{"cited_title":"Improving and generalizing flow-based generative models with minibatch optimal transport","cited_arxiv_id":null,"evidence_quote":"defines mini-batch optimal transport conditional flow matching, the dependent-coupling setting where the paper sets P=1"},{"cited_title":"and Nichol, A","cited_arxiv_id":null,"evidence_quote":"the diffusion classifier guidance that the framework recovers under the uncoupled affine Gaussian path"},{"cited_title":"T., Klasky, M","cited_arxiv_id":null,"evidence_quote":"DPS, the gradient posterior sampler recovered by the gcov-G guidance approximation"},{"cited_title":"Energy-weighted flow matching for offline reinforcement learning","cited_arxiv_id":null,"evidence_quote":"its training-based flow matching guidance loss coincides with the paper's MRGM loss"}],"review_version":1}