{"id":"6ea3e7d6-b73c-4ff2-abb2-44e6601b5988","arxiv_id":"2607.08781","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Property-aligned optimal-transport coupling in flow matching embeds a noise-space scalar that steers logP and QED distributions at inference with no oracle, reward model, or guidance.","lead":"Flow-matching models can be steered by how noise is paired with data at training time, not only by labels or guidance at sampling. A single noise scalar then shifts molecular property distributions without oracles or extra inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The strongest claim is distributional steerability via property-aligned monotone coupling, not pointwise CEM recovery. Proposition 1 already conditions on the coupling-preserving limit and the manuscript reports the realized slack (rho_per << 1) while still obtaining rho_group = 1.000, continuous distribution shifts, and chemically coherent opposite structural programs for logP vs QED. Ablations (Table 3) show OT coupling is necessary; key ablation (Table 4) shows the 1-D summary is not privileged; same-backbone Conditional FM comparison places the method on a distinct Pareto point rather than claiming dominance; epsilon-prediction attenuation is derived analytically. Scope limits (base-model quality, scalar properties, SELFIES) are acknowledged and do not undercut the scoped claim. The reader's weakest assumption is therefore already internalized by the paper; no additional load-bearing concern emerges that would move the verdict from ACCEPT.","tokens_in":25829,"tokens_out":477,"duration_ms":4795,"concrete_test":"Independently re-run the ZINC-250K logP s-sweep from the released checkpoint (or retrain with the published recipe) at n=1000 per s and recompute both rho_group(s, mean y) and the atom-count trajectory; if either rho_group falls below ~0.95 or the opposite-size pattern disappears while validity stays high, the distributional-control claim would need re-scoping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (imperfect coupling preservation, rho_per = 0.57/0.22) is already the paper's own stated residual in Proposition 1 and is treated as a distributional rather than pointwise claim. Empirically the group-mean monotonicity remains perfect (rho_group = 1.000), opposite size programs rule out generic size bias, OT/Dir/unmask ablations eliminate control when removed, the effect replicates on GuacaMol, and the epsilon-prediction negative result plus SNR derivation cleanly delimit where the interface exists. Within the paper's carefully scoped claim of distribution-level control via coupling, I do not find a further load-bearing soft spot that would overturn the central argument.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper reframes the noise–data coupling in flow matching as an alignment interface rather than a training heuristic. Reward Transport builds a property-aligned monotone OT coupling by sorting a scalar noise key s(z)=∥z̄∥₂ against a molecular property y and pairing rank-by-rank; a Direction MLP injects s into a Pre-LayerNorm Transformer. At inference, sweeping or thresholding s steers the generated property distribution with no oracle, reward model, or guidance. Proposition 1 shows that, in the coupling-preserving limit, this recovers one Cross-Entropy Method selection step. On ZINC-250K and GuacaMol the method yields group-mean ρ=1.000 for logP and consistent QED control, with opposite atom-count responses that rule out generic size bias; ablations, a sorting-key study, a same-backbone Conditional FM comparison, and an ε-prediction negative result with SNR analysis delimit the interface.","tokens_in":26036,"tokens_out":667,"duration_ms":6430,"significance":"If the claims hold, the work opens a distinct, distribution-level control channel for flow matching that is complementary to classifier-free guidance and conditional generation and that incurs zero inference overhead. The opposite structural programs for logP versus QED, the component ablations (OT, DirEmb, unmasked MSE), the sorting-key robustness, the GuacaMol replication, and the analytic attenuation of coupling gradients under ε-prediction are concrete, falsifiable contributions. Public code further strengthens reproducibility. The scoped claim—distributional rather than pointwise control under x̂₁/velocity prediction—is carefully stated and of clear interest to molecular generation and generative modeling more broadly.","major_comments":[],"minor_comments":[{"comment":"In §4.1 and Proposition 1 the realized per-molecule ρ_per (0.57 logP, 0.22 QED) is correctly treated as residual slack, but a short explicit sentence in the main text quantifying how far the distributional approximation sits from the ideal CEM truncation would help readers who only skim the proof appendix.","section":null},{"comment":"Table 1 and the GuacaMol tables report validity/uniqueness at 100%/≥99%; the extended FCD/SA audit in Appendix M is valuable but could be cross-referenced more prominently in §5.2 so that the quality–steerability trade-off is visible without leaving the main narrative.","section":null},{"comment":"Notation for the sorting key alternates between s(z), ∥z̄∥₂ and the normalized ŝ; a single consistent definition early in §4 would reduce minor ambiguity when reading Algorithm 1 and the inference paragraph together.","section":null},{"comment":"Appendix J’s SNR derivation is clear; a one-line pointer in §6 to the explicit attenuation factor E_t[α_t] would make the architectural boundary easier to cite without opening the appendix.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is carefully scoped and the empirical package is unusually thorough for a methods paper in this area. I see no load-bearing technical flaw that would justify major revision or rejection; the imperfect ρ_per is already the paper’s own residual and is handled honestly. Fit for a solid ML/methods venue is good."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple and useful. They reframe the noise–data coupling in flow matching as an alignment interface rather than a trajectory-straightening trick, then instantiate it as Reward Transport: sort noise by a 1-D key and molecules by a scalar property, pair by rank, and at inference just dial the same scalar. In the coupling-preserving limit this recovers one CEM truncation step (Prop. 1). No oracle, no reward model, no guidance, no extra compute.\n\nWhat is actually new is the property-valued cost plus the CEM identification, plus the empirical demonstration that the same knob grows molecules for logP and shrinks them for QED. That opposite structural program is the cleanest evidence they have that the coupling is writing target-specific structure rather than a generic size bias. The ablations are load-bearing: remove OT, DirEmb, or unmasked MSE and control collapses; the sorting-key ablation shows any reasonable 1-D projection works; the ε-prediction negative result plus the SNR derivation cleanly marks where the interface is structurally absent. GuacaMol replication and the same-backbone Conditional FM comparison keep the claim honest about being complementary rather than superior.\n\nSoft spots exist but are already scoped by the authors. Per-molecule Spearman is only 0.57/0.22, so Prop. 1 is distributional, not pointwise; they report this and still get perfect group-mean monotonicity. Absolute property levels are base-model bounded, SA fails for topological reasons they explain, and everything is one property at a time on SELFIES. None of that overturns the steerability claim as written. Math is classical 1-D OT plus a stated assumption; data and code look solid; citation pattern is appropriate.\n\nThis is for people who build or use molecular flow models and care about cheap distributional control. It deserves a serious referee. I would engage with it.","headline":"Clean, scoped idea: property-sorted OT coupling turns a noise scalar into a free distributional control knob for flow-matching molecules, with honest residuals and opposite-size evidence that rules out size bias.","tokens_in":26646,"tokens_out":496,"would_cite":true,"duration_ms":10091,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"In flow matching, the noise–data coupling itself can embed property control into the learned flow, so a single noise scalar steers molecular rewards at inference with no oracle or guidance.","keywords":["flow matching","optimal transport coupling","property-controlled generation","molecular generation","Cross-Entropy Method","noise-space alignment","SELFIES","reward transport"],"falsifier":"Retrain under the same recipe and measure per-molecule rank correlation between the noise scalar and generated property: if that correlation stays near zero while group means still appear monotone, or if the same scalar produces identical structural responses for chemically opposite targets, the claimed coupling-to-flow transfer fails.","tokens_in":26719,"feed_emoji":"🧪","tokens_out":613,"duration_ms":4993,"temperature":0.7,"pith_summary":"Flow matching usually treats the rule that pairs noise vectors with data points as a training detail. This paper argues that the pairing can instead be an alignment interface: sort noise by a scalar coordinate and molecules by a target property, then pair them by rank. That monotone coupling writes property structure into the learned flow field. At generation time, choosing the scalar selects a truncated property distribution—recovering one Cross-Entropy Method selection step when rank order is preserved—without any reward model, gradient guidance, or extra compute. On ZINC-250K and GuacaMol the same knob monotonically steers logP and QED, and it grows molecules for logP while shrinking them for QED, showing the structure is property-specific rather than a generic size bias. The result is a distribution-level control channel that is complementary to ordinary conditioning.","feed_headline":"One noise scalar steers molecular properties in flow matching","feed_subtitle":"Property-aligned pairing at training time embeds control; no oracle or guidance needed at generation.","key_machinery":"Reward Transport: the 1-D monotone rearrangement that sorts noise by a scalar key s(z) and molecules by property y, then pairs them rank-by-rank; Proposition 1 shows that, when the learned flow preserves ranks, conditioning on s ≥ τ yields the data distribution truncated to the corresponding property quantile.","core_discovery":"Property-aligned monotone optimal-transport coupling embeds controllable structure into a flow-matching field so that, at inference, varying a single noise-space scalar steers the generated property distribution with no oracle, reward model, gradient guidance, or added computation. In the coupling-preserving limit, thresholding that scalar recovers the Cross-Entropy Method’s truncated reward distribution.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Reward-aligned OT embeds free property control in flow matching","One noise scalar steers molecular rewards at inference","Noise-reward coupling turns a coordinate into a control knob","Vary a single noise coordinate to control molecule properties","OT pairs noise with rewards for oracle-free generation control"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The trained flow must keep enough of the training-time rank order between the noise scalar and the property; if that order collapses, the scalar knob loses its control.","fun_headline_variants_meta":{"raw":{"variants":["Reward-aligned OT embeds free property control in flow matching","One noise scalar steers molecular rewards at inference","Noise-reward coupling turns a coordinate into a control knob","Vary a single noise coordinate to control molecule properties","OT pairs noise with rewards for oracle-free generation control"]},"model":"grok-4.5","effort":"low","cost_usd":0.003866,"raw_usage":{"total_tokens":1231,"prompt_tokens":784,"num_sources_used":0,"completion_tokens":60,"cost_in_usd_ticks":38660000,"prompt_tokens_details":{"text_tokens":784,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":387,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":784,"tokens_out":60,"duration_ms":7000,"temperature":1.0,"reasoning_tokens":387,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T07:32:44.930482+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain under the same recipe and measure per-molecule rank correlation between the noise scalar and generated property: if that correlation stays near zero while group means still appear monotone, or if the same scalar produces identical structural responses for chemically opposite targets, the claimed coupling-to-flow transfer fails.","supporting_citations":[],"review_version":1}