{"id":"8f5ec0d9-23c3-40e6-b2a4-71e940d42f3c","arxiv_id":"2501.14216","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"TFG-Flow guides multimodal flow models at inference time by weighted Monte Carlo sampling for discrete atom types and gradient ascent for continuous coordinates, improving targeted molecular generation without extra training.","lead":"This paper introduces TFG-Flow, a way to steer generative flow models that handle both discrete and continuous data toward target properties, without retraining the model. The method is tested on molecular design, where users want molecules with specific quantum properties, structures, or protein-binding behavior.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The continuous guidance loop is a heuristic not shown to implement Theorem 3.2; the claimed marginal preservation and alignment are unproven for the actual algorithm.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: Algorithm 1's continuous guidance is not proven to realize the guided velocity of Theorem 3.2. My stress-test confirms this is the central issue. The discrete-guidance concentration bound in Theorem 3.4 is internally consistent as a consistency bound, though the abstract's 'unbiased sampling' wording is stronger than the ratio-estimator guarantee. The SO(3)-invariance argument in Theorem 3.5 is sound under its stated assumptions. But the paper's headline theoretical claims — preserving unguided flow marginals while aligning with the target predictor — apply to the idealized process of Theorem 3.2, not to the gradient-ascent perturbation actually executed in Algorithm 1. The Gaussian toy test would settle whether the heuristic reproduces the exact guided velocity; if it does not, the empirical results are not explained by the theory, and the paper should either weaken the theoretical claims or replace the continuous update with one that provably estimates E[X1|x_t,c]. This reinforces the reader's CONDITIONAL verdict rather than changing it. I also note the inconsistency between Table 1 and Appendix E for several TFG-Flow MAE values, which further supports caution but is not the load-bearing concern.","tokens_in":33678,"tokens_out":7180,"duration_ms":68800,"concrete_test":"Run a 1D Gaussian toy flow: x_t = (1-t)x_0 + t x_1 with x_0 ~ N(0,1), x_1 ~ N(mu, sigma^2), and predictor f_c(x_1) = exp(-(x_1 - c)^2 / (2 gamma^2)). For t in {0.1, ..., 0.9} and a grid of x_t, compute the exact guided velocity from Theorem 3.2 in closed form and compare it with Algorithm 1's effective update, using the exact E[X1|x_t] in place of the learned network and one or four gradient steps. If the relative L2 error exceeds 10%, or if simulated trajectories from Algorithm 1 do not track the exact Fokker-Planck marginal p_t(x_t|c), then the algorithm is a heuristic rather than an implementation of Theorem 3.2.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The implemented continuous guidance is not the guided flow of Theorem 3.2. Theorem 3.2 defines v_t(x_t|c) = E_{p1|t(x1|x_t,c)}[(x1 - x_t)/(1 - t)]; Algorithm 1 never evaluates this expectation. Instead, lines 15-19 replace x_t by x_t' = ProjectGamma(x_t + rho_t * grad_{x_t} log f(E[X1|x_t,a_t])) and then use the unguided network output E[X1|x_t',a_t] to step the ODE. No statement in the paper shows that p1|t(x1|x_t',a_t) approximates p1|t(x1|x_t,c). For a Gaussian flow with a Gaussian target it is straightforward to see that the Bayes-conditional mean has a different form from the gradient-ascent point; the update moves the input to the network, not the posterior. The actual simulated trajectory is therefore not the process of Theorem 3.1, and the claimed marginal preservation and alignment are unproven for the algorithm that produced Tables 1-4. The empirical gains could come from the discrete guidance, the tuned rho and tau, or a heuristic continuous drift, rather than from the paper's theoretical construction. This is an internal mismatch between theorem and method, not a disagreement with external consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TFG-Flow, a training-free guidance method for multimodal (continuous + discrete) flow matching models, with applications to 3D molecular design. The method constructs a guided flow via a joint measure (Theorem 3.1), derives guided velocity and rate matrices (Theorem 3.2), handles discrete guidance with a Monte Carlo estimator whose sample complexity is logarithmic in the number of atoms and atom types (Theorem 3.4), and handles continuous guidance by a gradient-ascent update on the input coordinates with a claimed SO(3)-invariance guarantee (Theorem 3.5). Experiments on QM9, GEOM-Drug, and CrossDocked2020 report improved property alignment, structural similarity, and binding scores relative to several training-free and training-based baselines.","tokens_in":33910,"tokens_out":11462,"duration_ms":99843,"significance":"If the theoretical claims were fully supported, this would be a valuable extension of training-free guidance to discrete and multimodal data, and the logarithmic sample-complexity result for discrete guidance would be a genuine contribution. The paper also introduces a useful pocket-based drug-design benchmark and provides open-source code. However, the current manuscript contains a load-bearing gap between the ideal guided flow of Theorems 3.1-3.2 and the continuous guidance loop actually implemented, as well as a discrepancy between the discrete estimator analyzed in Theorem 3.4 and the estimator used in Algorithm 1. These issues, together with an inconsistency between the headline tables and the appendix, mean that the paper's central claims are not yet established as stated.","major_comments":[{"comment":"The continuous guidance implemented in Algorithm 1 is not shown to realize the guided flow of Theorem 3.2. Theorem 3.2 (Eq. 9) defines the guided velocity as vt(xt|c) = E_{p1|t(x1|xt,c)}[(x1−xt)/(1−t)], but the algorithm never evaluates this expectation. Instead, it perturbs the input by Xt ← ProjectΓ(Xt + ρt ∇Xt log f(E[X1|Xt,at])) and then reads the unguided network output E[X1|Xt',at] to advance the ODE. No statement in the paper shows that p1|t(·|Xt',at) approximates p1|t(·|Xt,c); for a Gaussian flow the Bayes-conditional mean has a different functional form from this gradient-ascent point. Consequently, the simulated continuous trajectory is not the Fokker-Planck flow of Theorem 3.2, and the claims of marginal preservation and alignment are unsupported for the algorithm that produced Tables 1-4. This is a load-bearing gap: either a proof of the approximation should be supplied, or the continuous loop should be explicitly reframed as a heuristic whose empirical behavior is not covered by Theorems 3.1-3.2.","section":"§3.3, Algorithm 1 (lines 16-19)"},{"comment":"The implemented discrete estimator does not match the estimator analyzed in Theorem 3.4. Eq. (13) requires i.i.d. samples G1|t,k = (X1|t,k, a1|t,k) from the joint posterior p1|t(G1|Gt), but Algorithm 1 samples only the discrete component a1|t,k and reuses the single point estimate X1|t = E[X1|Gt] in every evaluation f(X1|t, a1|t,k). Because the target predictor fc(G1) is nonlinear in X1, E[fc(X1,a1)] ≠ fc(E[X1],a1), so Theorem 3.4's high-probability guarantee does not apply to the implemented estimator. Moreover, the abstract's 'unbiased sampling in guiding discrete variables' is stronger than anything proven: the self-normalized ratio estimator in Eq. (13) is only consistent and high-probability accurate under the assumptions of Theorem 3.4, not unbiased in finite samples. Please either modify the algorithm to sample X1|t,k jointly (if the flow model can provide samples) or provide an analysis of the bias introduced by the point-estimate substitution.","section":"§3.2, Algorithm 1 (lines 5-10), Eq. (13)"},{"comment":"The relationship between the stated goal and the theorem needs to be corrected. Section 3 states that the construction yields p1(G1|c) = pdata(G1|c), but Theorem 3.1 (formal version, App. B.1) only constructs a joint measure satisfying P(c|G1) = fc(G1), from which P(G1|c) ∝ fc(G1)pdata(G1) follows. This equals pdata(G1|c) only when fc is exactly the true conditional pdata(c|G1), which is not guaranteed for an off-the-shelf training-free predictor. Relatedly, the informal Theorem 3.1 in the main text says marginals are preserved for t ∈ [0,1], while the formal statement preserves finite-dimensional marginals only for t < 1; the final-time marginal cannot be preserved and simultaneously satisfy P(c|G1) = fc(G1) with nontrivial fc. The proof of Theorem B.2 also takes expectations with respect to pdata(x1|c), which is not the final marginal of the constructed P. Please align the informal and formal statements and avoid claiming exact conditional sampling.","section":"§3.1 and App. B.1"},{"comment":"The headline results are inconsistent with the appendix. For TFG-Flow on QM9, Table 1 reports Cv = 1.75, α = 2.32, Δε = 804, and εLUMO = 941, whereas Table 8 (App. E.2) reports Cv = 1.48, α = 3.52, Δε = 914, and εLUMO = 998 for the same metric, with no explanation (the tables also differ in validity and novelty columns). The differences are far larger than the reported standard deviations (e.g., α: 0.0523), so either the appendix or the main table is wrong. Because the central quantitative claims (e.g., '+20.3% over TFG') rest on Table 1, the discrepancy must be resolved and the exact hyperparameter configuration for each reported number must be stated.","section":"Table 1 vs. App. E.2 (Table 8)"}],"minor_comments":[{"comment":"In the proof of Theorem 3.5, the sentence 'X 7→ gθ(X, at)X is SO(3)-invariant' should read 'SO(3)-equivariant'; more importantly, the proof does not track the random discrete sample a1|t drawn at line 15 of Algorithm 1, so the invariance argument is incomplete as written.","section":"App. B.5, proof of Theorem 3.5"},{"comment":"The notation for the target predictor is inconsistent: Algorithm 1 line 10 uses f in place of fc from Eq. (13), while line 15 uses fc; the temperature τ introduced in §3.4 also does not appear in the algorithm listing. Please make the notation uniform.","section":"Algorithm 1 and §3.2"},{"comment":"The 'Upper bound' row in Table 3 is not defined as clearly as in the quantum-property tables; please specify whether it is the shuffled-label baseline from EDM and why the method is allowed to exceed it.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"For the editor only: the mismatch between Table 1 and Appendix E.2, if not resolved, would undermine trust in the empirical results; the authors should be asked to provide exact configurations and code outputs. The algorithm-theory gap for continuous guidance also needs an explicit statement of what is heuristic and what is proven before the theoretical claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on TFG-Flow. The genuinely new thing is the discrete guidance: they construct a guided rate matrix via a weighted Monte Carlo estimator that needs only O(log(n|A|)) samples, and they prove a high-probability error bound. That part is solid and useful. The continuous guidance, however, does not do what the theory says. Theorem 3.2 defines the guided velocity as the conditional expectation under p1|t(x1|xt,c); Algorithm 1 instead nudges xt by gradient ascent on log f(E[X1|xt,at]), then calls the unguided network on the perturbed input. That's a reasonable heuristic in the spirit of DPS/FreeDoM, but it is not shown to sample from the guided flow of Theorem 3.1, and the claims about preserving marginals and exact alignment do not follow for the actual algorithm. The abstract's 'unbiased sampling' claim is also stronger than the ratio estimator in Eq. (13), which is consistent rather than unbiased.\n\nThe experiments are broad and show real gains over continuous-only training-free baselines, and the paper is honest about the gap to training-based methods and the weak Tanimoto scores. But there are internal inconsistencies: Table 1 lists α MAE as 2.32 while Appendix E Table 8 reports 3.52; εLUMO 941 vs 998; Δε 804 vs 914. That's the kind of thing that needs a correction pass.\n\nWho should read it? Anyone working on guidance for flow matching or controllable molecular generation. The discrete part is a genuine contribution and the invariance analysis is clean. The continuous part should be treated as heuristic; the paper would be much better if it said so plainly and left the exact-guidance claim for the discrete component.\n\nMy recommendation: send it to review. The core idea is novel and the discrete analysis is worth refereeing. But the authors need to either prove something about the gradient-ascent update, or clearly mark it as an approximation and scale back the theoretical claims. Also fix the table mismatches. A major revision could make this a solid paper.","headline":"Genuinely novel discrete guidance for multimodal flow with a real sample-complexity bound, but the continuous-guidance algorithm is a heuristic that Theorem 3.2 does not actually cover, and the main-table numbers disagree with the appendix.","tokens_in":34499,"tokens_out":2763,"would_cite":true,"duration_ms":26688,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"TFG-Flow claims to construct a guided flow for multimodal generation that preserves the unguided flow's marginals, guides discrete atom types with a logarithmic Monte-Carlo estimator, and steers continuous coordinates equivariantly…","keywords":["training-free guidance","flow matching","multimodal flow","molecular design","discrete guidance","equivariant graph neural network","importance sampling","SO(3)-invariance"],"falsifier":"Run Algorithm 1 on a small known-data distribution where the guided marginal can be computed exactly (e.g., a two-atom Gaussian flow with a logistic classifier) and compare the distance between the algorithm's output distribution and the true p_1(x_1|c). If the distance grows with rho and does not shrink as N_iter increases, the gradient-ascent loop is not implementing the claimed guided flow.","tokens_in":1737,"feed_emoji":"🧪","tokens_out":2005,"duration_ms":75356,"temperature":0.7,"pith_summary":"The paper claims that guidance of a pretrained flow model can be extended from purely continuous data to multimodal data—coordinates plus discrete atom types—without any additional training. It proves that a guided flow exists that keeps the original flow's marginals while aligning the final sample with a target property, and derives the corresponding guided velocity and rate matrix. The key practical claim is that discrete guidance can be done with a self-normalised Monte-Carlo estimator using only a logarithmic number of samples, avoiding the exponential cost of exact expectations over atom-type assignments. Continuous guidance is performed by projected gradient ascent and provably preserves rotation invariance. If these claims hold, training-free guidance becomes applicable to molecular design and other scientific settings where discrete and continuous variables coexist, and the tests on quantum properties, structural similarity, and pocket-based drug design support that picture.","feed_headline":"Guidance without retraining works for mixed discrete-continuous molecular flow","feed_subtitle":"Logarithmic Monte-Carlo discrete guidance plus equivariant gradient ascent beats continuous training-free baselines.","key_machinery":"The central object is the guided multimodal flow p_t(G_t|c): a joint distribution over the flow trajectory {G_t}_{t in [0,1]} and the target condition c, built so that p(c|G_1)=f_c(G_1), the trajectory marginals equal the unguided flow's for t<1, and G_t and c are independent given G_1. The guided velocity v_t(x_t|c)=E_{p_{1|t}(x_1|x_t,c)}[v_{t|1}(x_t|x_1)] and the guided rate matrix R_t(a_t,b|c)=E_{p_{1|t}(a_1|a_t,c)}[R_{t|1}(a_t,b|a_1)] then generate the guided flow through the Fokker-Planck and Kolmogorov equations. The discrete part becomes tractable because the guided rate matrix equals a ratio of unguided posterior expectations, E[f_c(G_1)R_{t|1}]/E[f_c(G_1)], which the algorithm estimates with K i.i.d. samples from the flow model's posterior; Theorem 3.4 bounds the uniform error by O(sqrt((1/K) log(n|A|/delta))). The continuous part is steered by repeating X_t <- Project_Gamma(X_t + rho_t grad_{X_t} log f_c(E[X_1|X_t,a_t])) inside the mean-zero coordinate simplex; with an equivariant graph-neural-network backbone and an invariant predictor, this makes the simulated flow SO(3)-invariant.","core_discovery":"The paper claims that an unconditional multimodal flow model—one generating both continuous coordinates and discrete atom types—can be steered at inference time by an off-the-shelf time-independent predictor, without retraining. It constructs a joint distribution over the flow trajectory and the target condition such that the flow marginals stay exactly those of the unguided model, the condition aligns with the predictor at the clean sample, and the trajectory and condition are independent given the clean sample (Theorem 3.1); from this construction, the guided velocity and guided rate matrix follow as expectations conditioned on the current state and the target (Theorem 3.2). For discrete variables, exact computation of the guided rate matrix costs O(|A|^n) over atom-type assignments, but a self-normalised Monte-Carlo estimator using K = O((1/$epsilon^{2}$) log(n|A|/delta)) samples approximates it uniformly within epsilon with probability at least 1-delta (Theorem 3.4). For continuous coordinates, iterating a projected gradient-ascent step on the predictor's score gives an equivariant guided ODE whose simulated distribution is SO(3)-invariant (Theorem 3.5). On QM9, GEOM-Drug, and CrossDocked2020, this yields lower MAE than continuous training-free baselines on quantum properties, higher Tanimoto similarity to target structures, and the best Vina score among the compared pocket-based drug-design methods.","pith_inferences":["One testable extension: in a synthetic Gaussian flow with a known classifier, the exact guided marginal p_1(x_1|c) proportional to p_data(x_1)f_c(c) is computable, so comparing Algorithm 1's output with that exact distribution would isolate whether the gradient-ascent loop actually implements the ideal guided velocity—a link the paper asserts but does not prove.","If the equivalence between gradient ascent and the guided posterior expectation fails at large guidance strength, the marginals-preservation property would break before alignment, which would show up as falling validity at high rho and tau; the paper's ablation already hints at this trade-off.","The logarithmic discrete estimator could transfer directly to other flow or diffusion settings with factorised conditional rate matrices and off-the-shelf scorers, such as discrete text generation or protein sequence-structure co-design.","The SO(3)-invariance guarantee assumes an equivariant backbone and an invariant predictor; using a non-equivariant predictor on invariant features would still guide the final sample but would not inherit the coordinate-marginal invariance proof."],"forward_implications":["On QM9, TFG-Flow's MAE for six quantum properties is on average 20.3% better than the strongest continuous training-free baseline, and it is competitive with or better than a conditional flow that required labelled training data.","Target-structure guidance raises Tanimoto similarity over unconditional generation by 76.8% on QM9 and 22.4% on GEOM-Drug.","On CrossDocked2020, guided pocket-based design improves Vina score to -7.65, the best among the compared methods, while keeping QED and SA comparable.","Discrete guidance becomes practical: about 16 posterior samples already suffice empirically, consistent with the logarithmic sample-complexity theorem, and the method needs only four hyperparameters (rho, tau, K, N_iter).","Because the construction preserves flow marginals and the conditional-independence structure, the same guided velocity and rate-matrix recipe applies to any multimodal flow model with a factorised conditional flow, not only the molecular setting."],"supporting_citations":[{"why":"Supplies the multimodal flow framework (factorised conditional flow, conditional velocity and rate matrix, flow matching on discrete state spaces) that TFG-Flow builds on and adapts.","marker":"Campbell et al., 2024"},{"why":"Provides the training-free guidance formulation for continuous diffusion that TFG-Flow extends to multimodal flow, and supplies the strongest continuous baseline TFG in the experiments.","marker":"Ye et al., 2024"},{"why":"Defines the rectified-flow linear interpolation between noise and data that underlies the conditional flow construction used throughout the paper.","marker":"Liu et al., 2022"},{"why":"Supplies the EDM benchmark setup, molecule generation architecture, and the Cond-EDM training-based baseline whose MAE numbers are copied for comparison.","marker":"Hoogeboom et al., 2022"},{"why":"Provides the EGNN backbone whose SO(3)-equivariance is the load-bearing assumption for Theorem 3.5.","marker":"Satorras et al., 2021"},{"why":"Supplies the EEGSDE training-based guided baseline and the QM9 benchmark numbers that the paper compares against.","marker":"Bao et al., 2022"}],"fun_headline_variants":["Training-free guidance works for mixed discrete-continuous flow models","Steer molecular flow with discrete and continuous guidance, no retraining","TFG-Flow: unbiased discrete and equivariant continuous guidance for molecular generation","Guidance for flow without retraining: handles both discrete and continuous data"],"cache_read_input_tokens":36608,"weakest_assumption_plain":"The continuous guidance loop assumes that iterating projected gradient ascent on the predictor score produces a sample from the guided posterior used in Theorem 3.2, but the paper never proves this equivalence; if it fails at high guidance strength, the empirical gains are not explained by the theory.","fun_headline_variants_meta":{"raw":{"variants":["Training-free guidance works for mixed discrete-continuous flow models","Steer molecular flow with discrete and continuous guidance, no retraining","TFG-Flow: unbiased discrete and equivariant continuous guidance for molecular generation","Guidance for flow without retraining: handles both discrete and continuous data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000663,"raw_usage":{"total_tokens":3058,"prompt_tokens":1003,"completion_tokens":2055,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":619,"completion_tokens_details":{"reasoning_tokens":1979}},"tokens_in":619,"tokens_out":2055,"duration_ms":11545,"temperature":1.0,"reasoning_tokens":1979,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:14:49.522849+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Algorithm 1 on a small known-data distribution where the guided marginal can be computed exactly (e.g., a two-atom Gaussian flow with a logistic classifier) and compare the distance between the algorithm's output distribution and the true p_1(x_1|c). If the distance grows with rho and does not shrink as N_iter increases, the gradient-ascent loop is not implementing the claimed guided flow.","supporting_citations":[{"cited_title":"Equivariant diffusion for molecule generation in 3d","cited_arxiv_id":null,"evidence_quote":"Supplies the EDM benchmark setup, molecule generation architecture, and the Cond-EDM training-based baseline whose MAE numbers are copied for comparison."}],"review_version":1}