{"id":"d71908f5-5716-4423-9488-7c5211c71744","arxiv_id":"2509.01543","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Feynman-Kac particle steering, previously diffusion-only, is derived for conditional flow matching and used to generate chirality-correct chemical transition states.","lead":"Flow matching can be steered at inference time with the same Feynman-Kac particle-resampling trick used for diffusion models, tilting generated samples toward an energy or reward function. This gives chemists a way to force transition-state predictions to have the correct chirality, cutting RMSE on the RDB7 benchmark from 0.527 to 0.369 angstroms.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 13 guarantees convergence for potentials that telescope to exp(-λU(x1)); Algorithm 1's one-shot difference schedule tilts by U of an Euler endpoint, so the implemented target is U(one-shot), not U(x1), unless paths are straight.","rationale":"All other weaknesses are secondary: the high-dimensional mode-isolation result uses a coordinate-separable potential D = wΣ max(0,-xi), which factorizes and makes the task easier than a generic high-dimensional tilt; the GFM comparison is self-described as under-tuned and lacks diversity/ESS metrics; and the RDB7 residual TCE (5.8%) is close to the ground-truth inconsistency rate (6.3%), so the 'reliably' claim should be read against data noise. These are valid caveats but they affect interpretation of specific experiments, not the core method. The one-shot endpoint mismatch, by contrast, is the exact bridge between the theorem and every implemented result; without it, Algorithm 1 is not the object to which Eq. 13 applies. The paper itself flags the straightness dependence, which is honest but not a bound; the concern is internal to the argument, not a disagreement with consensus. Because the reader's CONDITIONAL verdict already rests on this gap, the stress-test does not move the verdict: it supplies a concrete test that would either discharge the condition or force a correction to the algorithm (e.g., final endpoint weight). Credit: the code release and explicit limitations discussion are genuine mitigating evidence, and the concern is not that the method is useless but that its advertised target is not established.","tokens_in":14479,"tokens_out":7578,"duration_ms":85630,"concrete_test":"On a synthetic flow with known non-straight trajectories (e.g., a quadratic or rotating velocity field interpolating two Gaussians), train a CFM and run two FK variants with the same resampling times and potential U: (a) Algorithm 1 as written (one-shot difference weights); (b) the same but with a final weight exp(-λU(x1))/exp(-λU(y_last)) applied at the last step so the product telescopes to exp(-λU(x1)) exactly. Compare the empirical x1 samples from both against a reference sample from p1 exp(-λU) (obtained by exact/rejection or a large SMC run), using sliced W2 and a thresholded U-tail probability. If (a) and (b) differ by more than the Monte Carlo uncertainty, the one-shot schedule changes the stated target; repeat under increasing path curvature and with σ>0 to quantify the bias. If they do not differ, the straightness heuristic is sufficient for this test and the concern is dischar","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing bridge is Section II C: Eq. 13 states that FK particle systems converge to p1(x)exp(-λU(x)) only when the product of intermediate potentials telescopes to exp(-λU(x1)). Algorithm 1 (lines 10-12) instead evaluates U at the Euler one-shot endpoint y_i = x_i + (1-t_i) vθ(x_i,t_i), and the difference schedule makes the product along any lineage equal exp(-λU(y_last)), where y_last is the one-shot endpoint at the final resampling step, not the actual sample x1. Thus the asymptotics of the implemented scheme are those of a Feynman-Kac measure tilted by U∘(one-shot map), not by U. The two agree only if the one-shot step is exact: the flow is straight (for deterministic ODE inference), or, with SDE injection, if the omitted score correction and integrated Brownian noise do not affect the endpoint. Neither condition is asserted or bounded. The paper's only warrant is the informal sentence 'depending on the straightness of the paths...' and the remark that multi-step estimates can be used when the flow is strongly non-linear; no sensitivity analysis, error bound, or corrected final weight is given. Because the chirality demonstration and the synthetic benchmarks are all produced with this one-shot schedule, the central claim 'steer CFM to p1 exp(-λU)' is not directly supported by the stated theorem. A simple fix would be a final correction exp(-λU(x1))/exp(-λU(y_last)), which would make Eq. 13 hold exactly; its absence is what makes the concern concrete. Also worth noting: the experiments are said to use a harmonic-sum schedule, whereas Algorithm 1 is written for a difference schedule, so the exact weight recursion used in the reported numbers is under-specified. Favorable context: code is released and the discussion names score-model and OOD limitations, so the issue is a missing correctness argument, not a hidden failure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Feynman-Kac (FK) steering for conditional flow matching (CFM), extending a framework previously developed for diffusion models. The authors prove that a CFM velocity field can be augmented with an isotropic diffusion term and an analytic score correction so that the resulting SDE preserves the flow-matching marginals (Theorem 1, Proposition 2, Corollary 3). They then define a sequential Monte Carlo resampling scheme with energy-based potentials, targeting the tilted distribution p1(x) exp(-λU(x)). The method is evaluated on synthetic tasks, including high-dimensional mode isolation and comparisons with guidance flow matching, and on a chemistry application: steering GoFlow transition-state predictions with a chiral-volume potential. The reported chirality results are strong: persistent chirality error drops from 9.0% to 0.7% and RMSE from 0.527 to 0.369 Å. The authors claim this is the first generative model producing stereochemically correct transition states from 2D graph information alone.","tokens_in":14749,"tokens_out":11295,"duration_ms":138491,"significance":"If the result holds, FK-CFM is a useful inference-time steering method that fills a real gap in the flow-matching literature. The central derivation is clean and standard, and several strengths deserve explicit credit: the code is open-source, the synthetic benchmarks use externally defined tilted targets, and the chirality metrics are external to the steering potential. The high-dimensional mode-isolation experiment directly addresses a known weakness of importance sampling and of some guidance methods. The paper is honest about limitations, including the reliance on a score model for stochastic injection and the heuristic nature of intermediate reward estimation. The main caveat is that the formal convergence statement is not fully connected to the exact implementation described in the text, which is fixable but needs attention.","major_comments":[{"comment":"As printed, the pseudocode computes vs = vθ(xs,t) once before the loop and never recomputes it after propagation or resampling. This means both the SDE propagation and the one-shot endpoint ys use the velocity at t=0 for all later steps. This does not implement the method described in the text, which states that vθ_ti(x_i) is evaluated at the current state and reused for the next ODE/SDE step. The algorithm must recompute vs at the current state/time at each iteration, and the text should state whether the one-shot endpoint uses the pre- or post-propagation velocity. Without this correction, the pseudocode is not a valid CFM steering rule except for exactly straight or constant-velocity flows.","section":"Section II C, Algorithm 1, lines 3-10"},{"comment":"The paper says the harmonic-sum schedule was used throughout the experiments, but Algorithm 1 implements a difference schedule without a final G_n term. The convergence condition in Eq. (13) is for potentials whose product telescopes to exp(-λU(x1)). For the harmonic-sum schedule, the final correction G_n = exp(-λU(x1)) / (∏G_i) (or an equivalent statement) must be given explicitly, together with precise definitions of L, ℓ, and H_L. I do not regard the one-shot endpoint as causing asymptotic bias: in the difference schedule the reward differences telescope to U(x1), and the final step evaluates the actual terminal point, so the one-shot approximation affects resampling efficiency rather than the limiting target. The missing piece is a formal specification of the potentials actually used in each experiment and a verification that Eq. (13) holds for that schedule. This is needed to connec","section":"Section II C, Eq. (13) and harmonic-sum schedule"}],"minor_comments":[{"comment":"The right-hand side of Eq. (13) appears to use x_i, but it should be the terminal variable, e.g. exp(-λU(x_n)) or exp(-λU(x1)). Please correct this to avoid ambiguity.","section":"Eq. (13)"},{"comment":"The text says 'we evaluated FKS' but the method is called FK elsewhere. Also, the caption 'G CFM' is unclear; it should be 'GFM' or 'MC-CFM' to match the notation in Section III B.","section":"Section III B and Table I"},{"comment":"For the TS-generation experiment, λ and the diffusion variance schedule are given, but the number of FK particles and the resampling cadence are not stated in the appendix. Please include these reproducibility details.","section":"Appendix A.3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern raised in the reader's report—that the one-shot endpoint biases the asymptotic target—does not, in my reading, land: the difference schedule telescopes the reward differences to U(x1), so the product condition in Eq. (13) is satisfied by construction; the one-shot approximation only affects resampling variance and particle diversity. However, the mismatch between Algorithm 1 and the harmonic-sum schedule described in the text is real and affects the paper's formal-to-implementation bridge. The authors should clarify the pseudocode and state the exact potentials used in the experiments. I would be comfortable with acceptance after these corrections are made."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. The paper does something genuinely new: it extends Feynman-Kac steering, previously a diffusion-only method, to conditional flow matching, and it shows a real application—chirality-correct transition-state generation on RDB7—with externally validated gains. The second thing is the main soft spot, and the stress-test note has it right: the convergence guarantee and the implemented algorithm don't quite meet.\n\nThe derivation is clean. Theorem 1 and Proposition 2 are standard SDE results assembled correctly, Corollary 3 follows, and Algorithm 1 is a proper particle filter with multinomial resampling. The TS numbers are the most convincing part: RMSE drops from 0.527 to 0.369 Å, persistent chirality error from 9.0% to 0.7%, and the residual thresholded error (5.8%) is below the ground-truth inconsistency rate they found in RDB7 itself (6.3%). That is a genuinely good empirical result, with code released and limitations (score-model requirement, OOD risk) named in the text.\n\nWhere it is soft, in proportion. The gap between Eq. 13 and Algorithm 1 is real: the theorem needs potentials whose product telescopes to exp(-lambda U(x1)), while the one-shot difference schedule makes the product exp(-lambda U(y_last)) for an Euler endpoint. The straightness sentence is acknowledged, but there is no bound on the bias and no sensitivity analysis. A final correction weight exp(-lambda(U(x1)-U(y_last))) would close it exactly, and its absence is what makes the concern concrete. I read this as a missing correctness argument, not a hidden failure—the empirical results behave as expected, which suggests the bias is small in practice, but it should be quantified. Related: Algorithm 1 is written for a difference schedule while the experiments use a harmonic-sum schedule described only in prose, so the exact weight recursion behind the reported numbers is under-specified. Minor: the high-dimensional escape-from-curse claim leans on a coordinate-separable potential; the non-separable indicator shows the usual degradation, and that caveat isn't stated as a limitation. And the GFM comparison is openly under-tuned, so read the W2 advantage as indicative, not settled.\n\nWho this is for: anyone working on inference-time steering of generative models, and anyone doing TS prediction. It deserves a serious referee. If I were handling it, I'd ask for the final correction weight or a bias analysis, a spelled-out weight recursion, and the separability caveat in the main text. I'd bring it to a reading group and I'd cite it.","headline":"FK steering for flow matching is genuinely new and the chirality application is strong; the real soft spot is the gap between the telescoping-potential theorem and the one-shot reward actually implemented.","tokens_in":15503,"tokens_out":4461,"would_cite":true,"duration_ms":48891,"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":"A new inference-time steering method, FK-CFM, applies Feynman-Kac resampling to conditional flow matching, letting users tilt a trained model's output toward any energy potential without retraining.","keywords":["conditional flow matching","Feynman-Kac steering","inference-time steering","energy-tilted posterior","particle filtering","chirality","transition state prediction","generative models"],"falsifier":"Construct a CFM model with a deliberately curved path (e.g., a large-angle circular interpolation or a highly nonlinear schedule), run FK-CFM with a known potential U, and compare the empirical endpoint distribution against the true tilted distribution p1 exp(-lambda U)/Z over many seeds; a systematic deviation growing with curvature and lambda would confirm the one-shot bias. Alternatively, measure the gap between the true final reward U(x1) and the one-shot estimate U(x + (1-t)v(x,t)) along steered trajectories—a large gap indicates the steering target is misspecified.","tokens_in":14189,"feed_emoji":"⚗️","tokens_out":7500,"duration_ms":72619,"temperature":0.7,"pith_summary":"The paper claims that Feynman-Kac steering, a particle-resampling technique previously used for diffusion models, can be transferred to conditional flow matching. It proves that adding a diffusion term and a score correction to a flow's ODE preserves the marginals while enabling resampling without particle collapse, and that if path potentials telescope to exp(-lambda U(x1)), the particle cloud converges to the tilted distribution. The one-shot Euler endpoint reward makes steering cheap because it reuses the velocity already computed for the next integration step. On synthetic benchmarks, FK-CFM isolates modes in high dimensions where importance sampling fails, and in the applied setting it steers GoFlow to generate transition states with correct chirality from 2D graphs alone, cutting persistent chirality errors from 9.0% to 0.7%. If correct, this is the first extension of Feynman-Kac steering to flow matching and the first generative model producing stereochemically correct transition states from graph input.","feed_headline":"Chirality errors fall to 0.7% with steered flow matching","feed_subtitle":"New inference-time method tilts conditional flow models toward energy targets—no retraining needed.","key_machinery":"The central object is the stochastic representation of a CFM path: given a flow field v_t generating path p_t with a Gaussian prior, the SDE dY_t = (v_t + (sigma^2/2) grad log p_t) dt + sigma(t) I dB_t has the same marginals (Theorem 1 / Corollary 3). This lets the FK particle filter resample without collapsing all trajectories to a single point. The second load-bearing identity is the product condition (Eq. 13): the intermediate potentials G_i must multiply along the path to exp(-lambda U(x1)); Algorithm 1 approximates this with a one-shot Euler reward and a difference schedule, using the velocity already computed for the next step so steering adds no extra forward passes.","core_discovery":"FK-CFM is an inference-time steering method that tilts a conditional flow matching model's output distribution toward low values of a user-specified potential U(x). The method simulates a cloud of particles, computes at each integration step a one-shot estimate of the final reward—y = x + (1-t)v(x,t)—resamples particles according to a difference schedule of these rewards, and injects Gaussian noise with a score correction so the marginals are preserved. The paper proves convergence: when the product of intermediate potentials equals exp(-lambda U(x1)), the resampled cloud converges to p1(x) exp(-lambda U(x)) as the particle count goes to infinity. The authors demonstrate this on synthetic ti","pith_inferences":["The one-shot endpoint bias suggests a direct test: on a deliberately curved CFM path (e.g., a nonlinear interpolation schedule), FK-CFM's effective target should drift away from U(x1) toward U(x + (1-t)v) as curvature grows; measuring this drift would quantify the bias.","The chiral-volume potential is additive, so the same resampling machinery could combine it with other geometric rewards (steric clash, bond-length constraints) to enforce multiple physical constraints at once.","The high-dimensional mode-isolation result hints that FK-CFM could serve as a general constrained-sampling tool for scientific generative models such as protein conformers or materials, where retraining is expensive and constraints are often non-differentiable—though the paper does not test this.","A systematic study of particle counts, resampling schedules, and potential tempering would be needed to turn FK-CFM from a demonstration into a predictive engineering tool; the paper leaves the variance behavior of the estimator unanalyzed."],"forward_implications":["FK-CFM steers any CFM model at inference time, without retraining or backpropagation, needing only velocity evaluations and optionally a score model.","The one-shot reward reuses the next step's velocity, making steering computationally nearly free per step; higher-order and multi-step reward estimates trade extra evaluations for accuracy on curved flows.","With continuous potentials, FK-CFM isolates rare modes in high dimensions with small particle counts, where importance sampling degenerates.","Steering GoFlow with the chiral-volume potential removes the wrong-chirality failure mode in transition-state prediction, producing correct stereochemistry from 2D graphs—a task the authors say no prior generative model solved.","The framework extends to non-Gaussian priors by training a score model concurrently (e.g., via SF²M), enabling stochastic steering beyond the Gaussian assumption."],"supporting_citations":[{"why":"Supplies the Feynman-Kac steering framework (reward tilting, resampling schedules) that the paper extends to flow matching.","marker":"[16]"},{"why":"Defines conditional flow matching, the generative class being steered.","marker":"[1]"},{"why":"Theorem 1 source: adding diffusion and score correction to the flow ODE preserves the marginals, enabling stochastic resampling without collapse.","marker":"[35]"},{"why":"Proposition 2 source: expresses the score grad log p_t in terms of the velocity field for Gaussian priors, used in the SDE correction.","marker":"[11]"},{"why":"Provides GoFlow, the flow matching transition-state model that FK-CFM steers for chirality.","marker":"[22]"},{"why":"Supplies score and flow matching (SF²M), used to train the score model for the non-Gaussian prior in the high-dimensional benchmark.","marker":"[37]"},{"why":"Concurrent guidance flow matching, the baseline compared against on synthetic tasks and the work whose reported difficulty FK-CFM addresses.","marker":"[33]"},{"why":"Prior application of FK steering (on diffusion) to chirality and biomolecular modeling that motivates the chiral-volume potential design.","marker":"[31]"},{"why":"Supplies minibatch optimal transport CFM (OT-CFM), an additional base model in the guidance comparison.","marker":"[38]"}],"fun_headline_variants":["Steering flow matching with Feynman-Kac to hit energy targets","Feynman-Kac steering tames conditional flow matching for energy targets","Flow matching now can be steered to low-energy samples without retraining","Chirality errors drop to 0.7% using Feynman-Kac steered flow","First Feynman-Kac steering for flow matching hits chiral targets"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The steering guarantee holds only if the path potentials telescope exactly to exp(-lambda U(finalsample)); the implementation instead evaluates the potential U at a one-shot Euler endpoint y = x + (1-t)v(x,t), so for flows with strong curvature the resampled cloud converges to the distribution tilted by U(one-shot estimate), not by U, and no bound on that bias is provided.","fun_headline_variants_meta":{"raw":{"variants":["Steering flow matching with Feynman-Kac to hit energy targets","Feynman-Kac steering tames conditional flow matching for energy targets","Flow matching now can be steered to low-energy samples without retraining","Chirality errors drop to 0.7% using Feynman-Kac steered flow","First Feynman-Kac steering for flow matching hits chiral targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000799,"raw_usage":{"total_tokens":3348,"prompt_tokens":737,"completion_tokens":2611,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":2510}},"tokens_in":481,"tokens_out":2611,"duration_ms":19849,"temperature":1.0,"reasoning_tokens":2510,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:27:31.597897+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a CFM model with a deliberately curved path (e.g., a large-angle circular interpolation or a highly nonlinear schedule), run FK-CFM with a known potential U, and compare the empirical endpoint distribution against the true tilted distribution p1 exp(-lambda U)/Z over many seeds; a systematic deviation growing with curvature and lambda would confirm the one-shot bias. Alternatively, measure the gap between the true final reward U(x1) and the one-shot estimate U(x + (1-t)v(x,t)) along steered trajectories—a large gap indicates the steering target is misspecified.","supporting_citations":[{"cited_title":"Accelerat- ing Biomolecular Modeling with AtomWorks and RF3,","cited_arxiv_id":null,"evidence_quote":"Supplies score and flow matching (SF²M), used to train the score model for the non-Gaussian prior in the high-dimensional benchmark."},{"cited_title":"Ramakrishnan, P","cited_arxiv_id":null,"evidence_quote":"Concurrent guidance flow matching, the baseline compared against on synthetic tasks and the work whose reported difficulty FK-CFM addresses."},{"cited_title":"Boltz-1: Democratizing Biomolecular Interaction Mod- eling,","cited_arxiv_id":null,"evidence_quote":"Prior application of FK steering (on diffusion) to chirality and biomolecular modeling that motivates the chiral-volume potential design."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies minibatch optimal transport CFM (OT-CFM), an additional base model in the guidance comparison."}],"review_version":1}