{"id":"a065d0b5-fab5-467c-8a6c-cd3c7196e092","arxiv_id":"2512.10339","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Intermediate densities in ratio-of-densities diffusion steering can become non-normalizable when experts use different noise schedules; a positivity criterion C(t)>0 detects this, and a time-varying bump on one exponent (ACE) restores a valid path.","lead":"The paper identifies a failure mode in inference-time steering of generative diffusion models: when separately trained experts are combined with different noise schedules, intermediate probability densities can stop being normalizable even though the start and end are valid. It proposes a test for this collapse and a time-varying exponent correction (ACE) that provably maintains a valid path, with gains in drug-design and compositional image benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.2's one-bump protocol is proven only for a scalar criterion, not the coordinate-wise C_k(t) of Eq. 4; with disjoint expert coordinate sets, a single bump may leave a collapsing coordinate unrepaired.","rationale":"The paper's central contribution is a diagnosis (Theorem 2.1) and a correction (ACE, Theorems 2.2 and 2.3). The path-existence criterion in Theorem 2.1 appears mathematically sound for compactly supported targets: the Gaussian-convolution envelopes in Proposition B.2 have the correct quadratic coefficients, and the coordinate-wise sum C_k(t) controls integrability. The weighted Feynman–Kac sampler in Theorem A.1 is internally consistent for both the ODE and the score-adjusted SDE, since the score term cancels the diffusion in the Fokker–Planck equation. The empirical results are suggestive but not fully rigorous (no error bars in Tables 3–4, small test sets). The reader's concern about learned approximate scores is a genuine practical limitation, and the paper itself acknowledges in §4 that large B linearly amplifies network approximation errors. However, that concern applies to any theoretical sampler and does not threaten the internal validity of the theorems. The most load-bearing issue is internal: the constructive proof of the bump protocol does not match the stated coordinate-wise criterion. Theorem 2.2 is stated for arbitrary coordinate sets I_i, but Theorem B.2's proof reduces to a single scalar S(t) and shows one bump on one expert suffices. With disjoint expert supports, a single bump can leave a collapsing coordinate untouched. This directly undercuts the headline claim that ACE guarantees a valid path for arbitrary heterogeneous experts. The fix is straightforward—require that the bumped expert covers each problematic coordinate, or add bumps to multiple experts—but as written the guarantee is not established. Therefore the paper should remain under conditional acceptance, with the proof or the claim clarified.","tokens_in":50756,"tokens_out":18451,"duration_ms":178286,"concrete_test":"Construct a 2D example satisfying Theorem 2.2's hypotheses: four experts with I_1={1}, I_2={2}, I_3={1}, I_4={2}, schedules taken from Table C.1, and boundary exponents such that C_1(t) and C_2(t) are positive at t=0 and as t→1 but negative in a middle interval. Apply the paper's one-bump construction (choose j maximizing γ_j(t_min)/(α_j(t_min))², add B·t(1−t)). Evaluate C_1(t) and C_2(t) on a fine grid; the coordinate not covered by j should remain negative. If it does, the proof of Theorem B.2 is invalid as stated for heterogeneous coordinate supports.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central guarantee is that ACE restores path existence for arbitrary heterogeneous experts. The key step is Theorem 2.2: adding a bump B·t(1−t) to one exponent γ_j makes the coordinate-wise criterion C_k(t)=Σ_{i:k∈I_i} γ_i(t)/(α_i^(t))² positive for every coordinate k and all t, preserving boundary exponents. However, the constructive proof in Appendix B.2 (Theorem B.2) analyzes only the scalar S(t)=Σ_i γ_i(t)/(α_i^(t))², not the C_k(t) of Eq. 4. If the coordinate where C_k(t) dips below zero is not in I_j (k∉I_j), the bump leaves that C_k(t) unchanged. The theorem's setup explicitly allows arbitrary coordinate subsets I_i (§2.1), so a single-expert bump is not guaranteed to fix all collapsing coordinates. This is not merely a gap in a side lemma: the abstract and §2.4 claim ACE guarantees a valid probability path whenever endpoints are valid, and the proposed protocol is precisely the one-bump construction. The existential form of Theorem 2.2 may still be salvageable (e.g., by bumping experts that cover each coordinate), but the stated constructive guarantee does not follow from the supplied proof. This makes the headline theoretical claim overbroad as written.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies inference-time steering of diffusion/flow models through heterogeneous ratio-of-densities paths p_t^* ∝ ∏_i (q_t^{(i)})^{γ_i(t)}. It identifies Marginal Path Collapse, where an intermediate density becomes non-normalizable despite valid endpoints; gives a coordinate-wise path-existence criterion C_k(t) for compactly supported targets (Theorem 2.1); proposes ACE, a bump-function correction of the exponents (Theorem 2.2) together with a weighted Feynman–Kac SDE/ODE sampler with time-varying exponents (Theorem 2.3); and reports experiments on a synthetic checkerboard, flexible-pose scaffold decoration, and compositional image generation.","tokens_in":51097,"tokens_out":10769,"duration_ms":115672,"significance":"The Gaussian-envelope proof of Theorem 2.1 is a genuine conceptual contribution: it converts an informal instability into a simple, checkable condition, and the converse direction is valid. The extension of Feynman–Kac steering to time-dependent exponents in Theorem A.1 is useful and appears correctly derived. The empirical evaluation is broad, and the paper is candid in Appendix F about several limitations. However, the headline claim that ACE ``guarantees a valid probability path whenever endpoints are valid'' is stronger than what the supplied proofs establish: the constructive one-bump theorem is proved only for a scalar criterion, not for the coordinate-wise C_k(t), and the sampling guarantee assumes exact expert score/velocity fields. These gaps are fixable, but they are load-bearing for the paper's central theoretical promise.","major_comments":[{"comment":"The constructive proof of the one-bump protocol is given for the scalar S(t)=Σ_i γ_i(t)/(α_t^(i))^2, not for the coordinate-wise C_k(t)=Σ_{i:k∈I_i} γ_i(t)/(α_t^(i))^2 in Eq. (4). Adding a bump Bt(1−t) to one expert j changes only the coordinates k∈I_j; if a coordinate k* has C_{k*}(t)<0 and k*∉I_j, that C_{k*} is unchanged. Since §2.1 explicitly permits arbitrary coordinate subsets I_i, the stated conclusion of Theorem 2.2 (``there always exists ... choosing one index j ... satisfies the path existence criterion for all t'') does not follow from the supplied proof. The existential claim may be salvageable by bumping a set of experts whose coordinate sets cover all coordinates where C_k dips below zero, but the current statement is overbroad. This gap is load-bearing because the abstract and §2.4 claim a guaranteed valid path for arbitrary heterogeneous experts.","section":"§2.4, Appendix B.3 (Theorem B.2)"},{"comment":"Theorem 2.3's sampling guarantee assumes exact velocity and score fields v_t^(i), s_t^(i) generating each q_t^(i). For pretrained networks these fields are learned approximations, and the paper itself notes in §4 that a large bump B ``linearly amplifies inherent network approximation errors.'' Therefore the abstract's ``guarantees a valid probability path whenever endpoints are valid'' should be scoped to the exact mathematical path, not to the path realized with approximate scores. This is the standard idealization in the guidance literature and is not fatal, but the unconditional wording in the abstract and §2.4 should be corrected.","section":"Appendix A.1 (Theorem A.1) and §4"}],"minor_comments":[{"comment":"The phrase ``whenever the endpoints are valid'' is not equivalent to Theorem 2.2's hypothesis, which requires C_k(0)>0 and lim_{t→1−}C_k(t)>0 for every coordinate. In particular, integrability of h_1 for compactly supported q_1^(i) can hold even when the limiting coefficient is non-positive. The wording should be aligned with the actual theorem assumptions.","section":"Abstract and §1 vs §2.4, Theorem 2.2"},{"comment":"The sentence ``ACE differs from NR and FKC only by its adaptive exponent correction (with B=30 for all experiments)'' conflicts with Table 2, which sweeps B=10,20,30,40,50,100, and with Figure 6. Presumably B=30 is used in the molecular and main synthetic experiments; please make this explicit.","section":"§3.1"},{"comment":"The stochastic term is written as σ_t \\tilde{s}_t·ξ_j (1/√Δt), which looks dimensionally inconsistent until one multiplies by Δt. A brief note that the displayed quantity is the per-unit-time form of the Itô increment would improve readability.","section":"Algorithm 1, line 8"},{"comment":"The table reports CLIP and Local CLIP, but the caption does not define them. Please add definitions or a reference so that the metric is unambiguous.","section":"Table E.10"},{"comment":"The limitations listed in Appendix F (extension to arbitrary transport, hybrid continuous-categorical domains, and error propagation in composed models) are appropriate and welcome. The main text's ``complete solution'' language should be tempered to acknowledge these scopes.","section":"Appendix F"}],"recommendation":"major_revision","confidential_remarks":"The main obstacle is the coordinate-wise gap in the proof of Theorem 2.2: the constructive bump argument proves only a scalar criterion, while the paper's headline guarantee is coordinate-wise. If the authors fix this by either proving an existential multi-bump construction, adding a coverage condition, or restricting the guarantee accordingly, the paper would be a solid contribution. The exact-expert assumption should also be stated prominently in the main text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick take: the paper identifies a real failure mode in ratio-of-densities steering—Marginal Path Collapse—and gives a genuinely useful criterion for compactly supported targets. The correction algorithm, ACE, is a sensible extension of Feynman–Kac correctors to time-varying exponents. But the headline guarantee is stronger than the proof supports: Theorem 2.2's constructive bump proof is proven only for a scalar criterion, not the coordinate-wise C_k(t) of Eq. 4.\n\nWhat's actually new: Theorem 2.1's condition C_k(t) = sum_{i:k in I_i} gamma_i(t)/(alpha_i^(t))^2 > 0 is a clean, computable test, and the Gaussian-envelope argument in the appendix is coherent. The converse direction (a negative coefficient forces non-integrability) looks valid. The paper also extends FKC to time-dependent exponents and derives a weighted SDE/ODE sampler, with the FKC constant-exponent algorithm as a special case. That is a real extension, not a repackaging. The synthetic checker benchmark and COCO-MIG results support the practical value of shaping C(t), and the scaffold-decorating application is well-motivated.\n\nSoft spots, in proportion:\n\nFirst, the stress-test concern is correct. Theorem B.2, which is supposed to prove Theorem 2.2, analyzes S(t)=sum_i gamma_i(t)/(alpha_i^(t))^2, not min_k C_k(t). If the coordinate where C_k dips below zero is not in I_j for the bumped expert, the bump leaves that coordinate unchanged. The main text's statement that \"there always exists a positive constant B such that choosing one index j and adding Bt(1−t) ... satisfies the path existence criterion\" is therefore not established in full generality. The existential claim may be salvageable by bumping multiple experts or by covering each coordinate, but the supplied proof doesn't deliver it. This is a load-bearing gap because the abstract and Section 2.4 promise a guaranteed valid probability path for arbitrary heterogeneous experts.\n\nSecond, the guarantees assume exact C^1 velocity and score fields. The paper itself notes that large B amplifies network approximation errors, so the \"provable\" path-existence guarantee does not directly transfer to pretrained models used in the molecular experiments. On those experiments: no code or data shipped, and the molecular evaluation is small with no error bars. The abstract also omits the compact-support and exact-expert assumptions, so it overstates the guarantee.\n\nThird, B is a hand-set hyperparameter; the sensitivity analysis is honest and B=30 is well supported, so this is minor.\n\nWho should read it: anyone composing pretrained diffusion/flow models for modular drug design or compositional generation; the criterion is a useful diagnostic even if the correction proof needs tightening. It deserves a serious referee, but with a request to fix or weaken Theorem 2.2 and to release code/data.","headline":"A genuinely useful diagnosis of path collapse in heterogeneous diffusion steering, with a correction that is promising but whose headline guarantee overreaches its coordinate-wise proof.","tokens_in":51599,"tokens_out":3392,"would_cite":true,"duration_ms":36997,"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":"Composing diffusion models by ratio-of-densities can silently produce non-normalizable intermediate paths; the paper derives a criterion that predicts this collapse and a time-varying-exponent correction that provably prevents it.","keywords":["diffusion steering","ratio-of-densities","marginal path collapse","path existence criterion","Feynman–Kac correctors","time-varying exponents","noise schedules","molecular generation"],"falsifier":"For a compactly supported target with exact (analytic) expert densities, pick schedules and exponents so that C_k(t*) < 0 at some interior time; the theorem predicts the integral of the composed density at t* diverges. Numerically computing that integral and finding it finite would refute the criterion.","tokens_in":50679,"feed_emoji":"🧪","tokens_out":6404,"duration_ms":65688,"temperature":0.7,"pith_summary":"The paper identifies a silent failure in inference-time steering of diffusion and flow models: when several pretrained experts are combined as a product (or ratio) of their densities with fixed weights, the intermediate probability distributions along the denoising path can stop being normalizable even though the start and end points are perfectly valid. The authors prove that for compactly supported data this collapse is decided by a single coordinate-wise formula built from the noise schedules and the weights, and they introduce ACE, which replaces the fixed weights by time-varying ones (a positive 'bump' in the middle of the path) so the condition stays satisfied without changing the endpoints. They also supply a particle-filter sampler that follows the corrected path by reweighting and resampling particles as it goes. If correct, this converts ratio-of-densities steering from an unstable heuristic into a guaranteed procedure for combining heterogeneous pretrained models, with demonstrated gains in molecular scaffold decoration and compositional image generation.","feed_headline":"One formula predicts when diffusion steering collapses","feed_subtitle":"Composing pretrained models by ratio-of-densities can create non-normalizable paths; adaptive exponents fix it.","key_machinery":"The load-bearing object is the coordinate-wise path-existence criterion C_k(t) = Σ_{i:k∈I_i} γ_i(t)/(α_t^(i))², which converts the question 'is the composed intermediate density normalizable?' into a sign check on schedules and exponents. The correction is the bump-function protocol: adding B·t(1−t) to one positive-exponent schedule lifts C_k(t) above zero on (0,1) without altering the endpoint distributions. The sampling engine is a weighted Feynman–Kac SDE/ODE whose weight update includes the new time-derivative terms γ̇_i(t) log q_t^(i), together with resampling on effective sample size; setting the exponents constant recovers the earlier Feynman–Kac corrector as a special case.","core_discovery":"For compactly supported targets, a ratio-of-densities composition h_t = ∏ (q_t^(i))^{γ_i(t)} stays integrable exactly when C_k(t) := Σ_{i:k∈I_i} γ_i(t)/(α_t^(i))² > 0 for every coordinate k (given h_1 integrable). If some C_k(t) < 0, Marginal Path Collapse occurs: an intermediate density is non-normalizable, so the sampler follows a different, unintended path. ACE fixes this by adding a bump B t(1−t) to one exponent, preserving the endpoints while restoring C_k(t)>0, and provides a weighted Feynman–Kac SDE/ODE whose marginal is the corrected path. Because the criterion depends only on schedules and exponents, path validity can be certified before any network evaluation.","pith_inferences":["The criterion also suggests a quantitative safety margin: the minimum positive value of C(t) could serve as an a priori 'stability score' for a composition, guiding how much bump is needed before running any model.","Because the proof of the criterion relies on compact support and exact score fields, a testable extension is to measure how empirical sample validity degrades as the learned scores' error grows, and whether that degradation tracks the dip in C(t).","The paper notes that the bump parameter B linearly amplifies network approximation errors; this implies an optimal-B trade-off curve (peak near B=30 in their experiments) that could be predicted from the score-error profile of each expert, rather than tuned per task.","Extending ACE to mixed continuous-categorical molecular spaces would need a discrete analogue of the Gaussian-envelope argument; the paper explicitly leaves this open, so the current guarantee does not cover models that evolve atom types as categorical variables."],"forward_implications":["Inference-time steering of heterogeneous pretrained models (e.g., de-novo, conformer, and pocket-conditioned molecular experts) can be run at high guidance scales without the generated molecules fragmenting into invalid structures.","Practitioners can pre-screen any composition: compute C(t) on the sampler's time grid from schedules and exponents alone; if it stays positive, the path is guaranteed valid, so costly failures are caught before sampling.","Because ACE keeps the endpoint densities fixed and only reshapes intermediate marginals, it can replace constant-exponent ratio-of-densities recipes (classifier-free guidance, product-of-experts, contrastive decoding) whenever experts use mismatched schedules.","Even when no collapse occurs, increasing C(t) via a small bump tightens the intermediate distributions' tails, which empirically improves attribute accuracy in compositional image generation.","The weighted sampler with resampling removes out-of-distribution trajectories from the batch, which the no-resampling heuristic leaves in place."],"fun_headline_variants":["A formula predicts diffusion steering collapse","Adaptive exponents stop generative path collapse","Collapse criterion for diffusion steering: new test","ACE: fix collapsed diffusion paths with one tweak","Certify diffusion path validity with one formula"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The guarantee assumes each expert's velocity and score fields are the exact generators of its density path; real pretrained models supply learned approximations, and the paper itself notes that the correction parameter B linearly amplifies their errors.","fun_headline_variants_meta":{"raw":{"variants":["A formula predicts diffusion steering collapse","Adaptive exponents stop generative path collapse","Collapse criterion for diffusion steering: new test","ACE: fix collapsed diffusion paths with one tweak","Certify diffusion path validity with one formula"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1378,"prompt_tokens":782,"completion_tokens":596,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":544}},"tokens_in":526,"tokens_out":596,"duration_ms":6440,"temperature":1.0,"reasoning_tokens":544,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T17:11:33.934073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a compactly supported target with exact (analytic) expert densities, pick schedules and exponents so that C_k(t*) < 0 at some interior time; the theorem predicts the integral of the composed density at t* diverges. Numerically computing that integral and finding it finite would refute the criterion.","supporting_citations":[],"review_version":1}