{"id":"de76bc37-2bfc-4ece-9d21-f21eb98c4e86","arxiv_id":"2505.01382","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For arbitrary data distributions, classifier-free guidance provably decreases the expected reciprocal classifier probability along the reverse diffusion process.","lead":"Guidance in diffusion models is the standard trick for steering generated images toward a target class. This paper proves that for any data distribution, the average inverse classifier probability falls as guidance strength grows, a rigorous explanation of why guidance raises classifier confidence on average.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's infinitesimal comparison rests on unproved fourth-order exponential bounds (21a)-(21c); Section 4.2 only proves the first-gradient case, leaving the O(δ) remainder control unverified.","rationale":"The reader's weakest assumption identifies exactly the load-bearing gap: the proof of Theorem 1 relies on the unproven exponential bounds (21a)-(21c) to justify the O(δ) remainder control in the Itô expansions (14)-(16). My independent reading confirms this is the least secure part of the argument. The core martingale calculation via Lemma 2 is elegant and, under a positive-time reading, the algebra checks out: for τ>0 the classifier probabilities are positive, and the reverse-time martingale relation follows from the forward Bayes computation. The remaining issue is not the identity itself but the analytic regularity needed to pass from Itô expansions to the infinitesimal comparison. Since Section 4.2 only proves the first-gradient case and the omitted higher-order and mixed time-space derivatives are essential to equations (30)-(33), the theorem is currently conditional on a proof that is sketched rather than supplied. This is fixable: the bounds are plausible for general distributions because Gaussian convolution provides lower bounds on the densities, and derivative tails can be controlled by Gaussian decay. I also checked the stated Lemma 2 at τ=0 and found a boundary defect for distributions with disjoint class supports, but the theorem's proof only needs the relation at positive times, so this does not change the verdict. The numerical experiments are not decisive for the exact theorem quantity, but the claimed sign and the GMM/ImageNet behavior are consistent with the result. Overall, the reader's CONDITIONAL verdict is appropriate: the central claim is likely correct but not yet fully proven as written.","tokens_in":14973,"tokens_out":20200,"duration_ms":217086,"concrete_test":"Derive explicit formulas for ∇^2 pc|X1-t(c|·)^{-1} and ∂_t∇^2 pc|X1-t(c|·)^{-1} as Gaussian-convolution integrals for a heavy-tailed class-conditional example, e.g., X0|c ~ Cauchy(0,1) with X0 a balanced mixture of Cauchy and Gaussian classes. Then check whether the claimed bound exp(C_{t,2,R}(1+||y||^2)) holds for large ||y|| and whether the conditional exponential moment E[exp(C||Y_r||^2)|Y_t=y] is finite with the constants C forced by (23a). If the remainder terms in (28)-(33) are not O(δ), recompute (16) and test whether the leading negative term still dominates; if not, Theorem 1's local decrease fails in that example.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central proof of Theorem 1 compares the unguided and guided reverse SDE expansions (14) and (15) and takes their difference to obtain the negative infinitesimal drift (16). Both expansions are only valid up to O(δ), and the proof explicitly requires the exponential growth bounds (21a)-(21c) on the reciprocal classifier probability and its derivatives up to fourth order, plus mixed time-space derivatives, to control the remainder terms via Itô's formula in Section 4.3. Section 4.2 proves only the first-gradient bound (23b)-(23c) and states that the other bounds 'can be derived using similar techniques'. If any of these higher-order or mixed derivative bounds fail, or if the constants C_{t,k,R} are large enough that the conditional exponential moments in (30)-(31) diverge under the reverse process, then the O(δ) remainders in (14)-(16) are not controlled, and the claimed pointwise decrease in (11) is not established. This is load-bearing because the theorem's entire conclusion is the signed infinitesimal comparison; a sign-reversing or unbounded remainder would invalidate it. A secondary statement-level flaw is that Lemma 2 allows τ=0, where the equality can fail when p(c|X0) vanishes on a positive-probability set; however, the theorem's proof only invokes the relation at positive times, so this is repairable rather than fatal.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies classifier-free guidance (CFG) in continuous-time diffusion models and claims a provable improvement for general data distributions. The central object is the reciprocal classifier probability φ_t(y) = p_{c|X_{1-t}}(c|y)^{-1}. Theorem 1 states that along the guided reverse SDE, E[φ_{t+dt}(Y^w_{t+dt}) | Y^w_t] - φ_t(Y^w_t) equals a nonpositive infinitesimal quantity proportional to w/t times the squared norm of the difference between conditional and unconditional scores, with an additional O(dt) term. The proof relies on Lemma 2, which asserts a martingale property for the reciprocal classifier probability under the unguided reverse process, and on Itô-Taylor expansions whose remainders are controlled by exponential growth bounds on φ and its derivatives. The paper also includes GMM and ImageNet experiments showing that guidance improves the average reciprocal classifier probability even though it does not improve every sample, and an appendix analyzes discrete-time approximation and score-estimation errors.","tokens_in":15237,"tokens_out":14798,"duration_ms":146997,"significance":"If the main theorem is established rigorously, this is a meaningful step beyond the existing case studies of isotropic Gaussian mixtures and one-dimensional interval-supported distributions: it identifies a scalar metric—the average reciprocal classifier probability—that guidance provably improves under general data distributions, while correctly predicting that uniform per-sample improvement fails. The underlying mechanism is attractive and elegant: the reciprocal classifier probability is a martingale along the class-conditioned reverse process, and the CFG extra drift is exactly aligned with the negative gradient of that martingale. The numerical experiments are consistent with the qualitative claim, and the paper is candid about the fact that individual samples may degrade. However, the theorem as currently written depends on a misstated lemma and on unproved high-order derivative bounds; both are repairable in principle, but until they are fixed the central claim is not fully established.","major_comments":[{"comment":"Lemma 2 as stated in Eq. (13a) is false: for a fixed x, E_{X_τ|X_t=x}[p(c|X_τ)^{-1}] is not equal to p(c|X_t=x)^{-1} in general. The proof in §4.1 implicitly conditions on the class label as well, since the first line uses the density p_{X_τ|X_t,c}(x_τ|x,c) rather than p_{X_τ|X_t}(x_τ|x). The identity that the proof actually establishes is the class-conditional martingale property, which is the one needed for Eq. (14) when Y_t is the class-conditioned reverse process. The lemma should be restated with the conditioning on c made explicit, and the range should avoid τ=t or should require ε>0 so that p(c|X_0) is not allowed to vanish on a positive-probability set. As written, the statement is false and the proof does not prove the stated claim.","section":"Lemma 2 and §4.1"},{"comment":"The proof of Theorem 1 uses the O(δ) expansions (14)–(16), whose remainder terms are controlled by the exponential growth bounds (21a)–(21c) on p_{c|X_{1-t}}^{-1} and its derivatives up to fourth order, including mixed time-space derivatives. Section 4.2 proves only the first-gradient estimate (23b)–(23c) and states that the remaining bounds \"can be derived using similar techniques\". Moreover, (30)–(31) require finiteness of conditional exponential moments of ||Y_r||^2 under the reverse process, which is asserted but not shown. Because the entire signed infinitesimal comparison in Eq. (11) depends on these remainders being o(δ), the missing derivations are load-bearing; they must either be supplied in full or the theorem must be stated under explicit assumptions that guarantee them.","section":"§4.2–4.3, bounds (21a)–(21c)"},{"comment":"The proof establishes the conditional infinitesimal inequality (11), but the abstract and Section 3 conclude that guidance decreases the average reciprocal classifier probability over the whole generation process. This integrated conclusion requires an exchange of the infinitesimal limit with expectation over the law of Y^w_t and a proof that the right-hand side of (12) is finite on (δ,1). The paper should state the integrated result explicitly as a corollary: for any t>δ, E[φ_t(Y^w_t)] ≤ E[φ_δ(Y^w_δ)] = E[φ_t(Y_t)], with the difference equal to the integral of the nonnegative term in (11). Without this statement, the paper's main advertised conclusion is not formally derived.","section":"Theorem 1, Eq. (11) vs. Eq. (12)"}],"minor_comments":[{"comment":"In the first line of (16), the subscript \"1−τ\" in \"∇ log p_{X_{1−τ}}(Y^w_t)\" should be \"1−t\"; as printed, τ is undefined in that expression.","section":"Eq. (16)"},{"comment":"Equation (8) states \"{Y_t} d = {X_{1−t}}\" when Y_δ∼p_{X_{1−δ}|c}; this is only correct if X_t is understood to be the forward process conditioned on class c. Please state the conditional law explicitly to avoid confusion with the unconditional forward process.","section":"Lemma 1, Eq. (8)"},{"comment":"The left-hand sides of (21a)–(21c) are tensors; the inequalities should use the appropriate tensor norm. In addition, the dependence of C_{t,k,R} on t should be made explicit, in particular how it behaves as t→1 or t→0.","section":"§4.2, inequalities (21a)–(21c)"},{"comment":"The displayed inequality in (40) appears to miss a factor of 2: the standard coupling argument gives Eμ[f]−Eν[f] ≤ Eμ[(f−1)1_{f>τ}]+(τ−1)TV, and with TV≤P(f>τ) this yields 2Eμ[(f−1)1_{f>τ}], not the expression shown. This does not affect Theorem 1, but the claimed practical transfer should be corrected.","section":"Appendix A, Eq. (40)"},{"comment":"The sentence before Assumption 2 says \"which is stated in the following lemma\"; it should say \"assumption\" rather than \"lemma\".","section":"Assumption 2"}],"recommendation":"major_revision","confidential_remarks":"The core idea is sound and worth publishing after the requested fixes. The false statement of Lemma 2 is concerning but clearly repairable by reformulating the martingale identity with explicit conditioning on the class label. The missing high-order derivative bounds in Section 4.2 are technical but essential; the authors should be encouraged to supply full derivations rather than a heuristic \"similar techniques\" statement. I would not recommend rejection on the basis of the current gaps, provided the authors address these points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know two things about arXiv:2505.01382. First, it is the first paper I have seen that attempts a general-distribution analysis of classifier-free guidance, going beyond the Gaussian-mixture and one-dimensional interval cases in Wu et al., Chidambaram et al., and Bradley and Nakkiran. Second, the main theorem is not fully proved as written, because the proof leans on a set of exponential growth bounds for derivatives of the reciprocal classifier probability that the paper asserts but does not prove.\n\nThe genuine contribution is Lemma 2: the reciprocal classifier probability is a martingale along the forward process. That is a clean observation, and the authors use it to derive an infinitesimal comparison between guided and unguided reverse SDEs. The calculation in Section 4.3 is plausible, and the discrete-time robustness analysis in Appendix A is a useful bonus.\n\nThe soft spots are real but not fatal in spirit. The stress-test note is right: equations (21a)-(21c) require exponential bounds on derivatives up to fourth order, plus mixed time-space derivatives, to control the O(δ) remainders in (14)-(16). Section 4.2 only proves the first-gradient case and waves off the rest as \"similar techniques.\" That is load-bearing, because if any of those bounds fail, the signed infinitesimal comparison in (11) is not established. I suspect the bounds are true under mild conditions on the data distribution, but they need to be written out before the theorem is credible. Also, Lemma 2 as stated allows τ=0, where it can fail for a symmetric Gaussian mixture when p(c|X0) vanishes on a positive-probability set. The proof only uses positive times, so this is a repairable statement-level flaw.\n\nOn circularity: yes, the reciprocal classifier probability is chosen specifically because its gradient is proportional to the CFG drift. The authors admit this. That makes the result a theorem about a technical metric, not about Inception Score or FID. Still, it is a legitimate first step, and the martingale fact is independent of the metric choice.\n\nMy verdict: this deserves a serious referee, but the paper should go back for major revision. The missing derivative bounds and the τ=0 issue need fixing before the main theorem can be accepted. I would bring it to a reading group as a maybe, mainly to discuss the martingale identity and the general strategy. I would not cite it in my own work until the gap is closed.\n\nRecommendation: engage with it, but treat the main theorem as conditional.","headline":"First general-distribution analysis of classifier-free guidance, but the main theorem rests on unproved higher-order derivative bounds and a metric chosen to fit the proof.","tokens_in":15745,"tokens_out":2557,"would_cite":false,"duration_ms":24490,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","60J60","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Guidance provably lifts the average classifier confidence of diffusion-generated samples, for general data distributions.","keywords":["diffusion models","classifier-free guidance","score-based generative modeling","reverse stochastic differential equation","classifier probability","sample quality","Itô calculus","discretization error"],"falsifier":"Simulate the continuous-time guided reverse SDE with known scores for a simple non-Gaussian data distribution, such as a two-component Gaussian mixture with unequal covariance, using many noise seeds, and estimate $E[p_{c|X_1}(c|Y^w_1)^{-1}]$ for $w=0$ and several $w>0$. Theorem 1 predicts a strict decrease and a drop growing with the integrated squared score gap; seeing no decrease, or a non-monotone pattern in $w$, would falsify the average-improvement claim. A more direct check estimates both sides of the infinitesimal identity (11) at a fixed $t$ by Monte Carlo over a small time step.","tokens_in":14737,"feed_emoji":"🎯","tokens_out":11636,"duration_ms":119885,"temperature":0.7,"pith_summary":"This paper tries to prove a general reason why adding guidance to a diffusion model raises sample quality, for essentially arbitrary data distributions rather than the one-dimensional or Gaussian cases studied before. The quality measure is the reciprocal of the classifier probability: guidance should make the average of $p_{c|X_{1-t}}(c|y)^{-1}$ decrease along the reverse diffusion, which means samples with high classifier confidence become more common on average. The core result is an exact infinitesimal identity: at each time $t$, the expected drop in this reciprocal equals $(w/t)\\, p_{c|X_{1-t}}(c|Y^w_t)^{-1}\\|\\nabla\\log p_{X_{1-t}|c}(Y^w_t|c)-\\nabla\\log p_{X_{1-t}}(Y^w_t)\\|^2\\,dt$, where $w$ is the guidance strength. A companion discrete-time bound shows that the conclusion survives discretization and bounded score-estimation error, so the proof is aimed at the setting practitioners actually use.","feed_headline":"Guidance provably lifts average classifier confidence in diffusion models","feed_subtitle":"New theorem: the reciprocal classifier probability falls along the guided reverse diffusion, for general data distributions.","key_machinery":"The load-bearing object is the reciprocal classifier probability $φ_t(y)=p_{c|X_{1-t}}(c|y)^{-1}$. Lemma 2 shows that $φ_t$ is a martingale through the diffusion bridge: its conditional expectation at an earlier noisy time equals its value at a later noisy time, because the forward transition kernels are Gaussian and the classifier probability can be written through Bayes' rule as a ratio of unconditional and conditional densities. This martingale property makes the unguided reverse SDE have zero expected drift in $φ_t$. The proof then compares the unguided drift with the drift of the guided SDE using Itô's formula; the guidance term adds exactly the drift $\\nabla\\log p_{X_{1-t}|c}-\\nabla\\log p_{X_{1-t}}$, which by identity (17) is the negative gradient direction of $φ_t$ up to the factor $φ_t$. Hence the extra drift produces a strictly negative expected increment whose size is the squared norm of the score difference, scaled by $w/t$ and by $φ_t$.","core_discovery":"The paper's central claim, Theorem 1, is that guidance provably reduces the average reciprocal classifier probability along the guided reverse SDE. Concretely, for any $δ<t<1$ and any positive guidance weight $w$, conditional on the current point $Y^w_t$, the one-step-ahead expectation of $φ_{t+dt}(Y^w_{t+dt})$ is smaller than $φ_t(Y^w_t)$ by the non-negative quantity in (11). Since $φ$ is a decreasing function of the classifier probability, this says the guided process shifts probability mass toward samples the classifier finds more likely, on average, for every data distribution satisfying the paper's regularity bounds. The paper also proves Theorem 2, that the discrete-time classifier-free sampler with Lipschitz score estimates and mean-squared score error $ε^2_{\\mathrm{score}}$ is within $O((1+w^2)L^2 d\\log^3N/N + (1+w^4)L^2 R^2\\log^4N/N^2 + (1+w^2)ε^2_{\\mathrm{score}}\\log N)$ in KL divergence of the continuous guided process, so the average-improvement statement carries over to practical implementations.","pith_inferences":["A natural extension, not proved here, is the same argument for $E[\\log p_{c|X}(c|Y)]$, the actual Inception Score weight; convexity of $-\\log$ suggests the martingale mechanism could transfer, but the paper stops at the reciprocal.","The identity (11) implies an adaptive guidance schedule could set $w$ locally proportional to the score gap or to $φ_t$ itself, concentrating guidance where it helps most; the paper only treats the constant-$w$ case.","The martingale property of $φ_t$ may hold for any bridge with Gaussian transition kernels, which might make it a reusable tool for analyzing other guided generative processes beyond diffusion guidance.","Because the drop is inversely proportional to $t$, early-time guidance dominates; a testable prediction is that truncating guidance after intermediate times should retain most of the average improvement while preserving diversity, a consequence the authors do not draw explicitly."],"forward_implications":["For every distribution satisfying the technical bounds, increasing the guidance weight $w$ monotonically decreases the expected reciprocal classifier probability along the reverse process, so average sample confidence improves.","The improvement is not uniform: individual samples can lose classifier confidence even while the population average improves, matching the GMM and ImageNet experiments in the paper.","Guidance has the largest per-time effect where $1/t$ is large, early in the reverse process, and where the conditional and unconditional score functions differ most, so low-confidence regions receive the strongest correction.","Discretization and score-estimation errors only perturb the result by an explicit KL term, so with enough reverse steps and accurate scores, the discrete classifier-free sampler inherits the average-improvement guarantee.","The metric $E[p_{c|X_1}(c|Y^w_1)^{-1}]$ responds to guidance in the same direction as Inception Score-style measures, giving a theoretical explanation for why classifiers see better samples after guidance."],"supporting_citations":[{"why":"Establishes the continuous-time SDE formulation and the forward-reverse time reversal (Lemma 1) on which the martingale computation is built.","marker":"Song et al. (2021b)"},{"why":"Introduces classifier-free guidance and the score-difference form used as the drift in the guided reverse process.","marker":"Ho and Salimans (2021)"},{"why":"Provides the empirical phenomenon of classifier guidance improving sample quality that the theorem is meant to explain.","marker":"Dhariwal and Nichol (2021)"},{"why":"Defines the Inception Score, the classifier-probability-based quality metric that motivates the reciprocal-probability objective.","marker":"Salimans et al. (2016)"},{"why":"Supplies the KL-discretization and score-error analysis technique adapted in Theorem 2.","marker":"Chen et al. (2022)"}],"fun_headline_variants":["Guidance provably shifts samples toward classifier-favored regions","Theorem: guidance boosts average classifier confidence for any data","Guidance provably improves sample quality for general distributions","Average classifier confidence rises under guided diffusion, proven","New proof: guidance lowers reciprocal classifier probability on average"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof of the main infinitesimal identity assumes the reciprocal classifier probability and its derivatives up to fourth order grow at most exponentially in $\\|y\\|_2^2$ with constants depending on $t$, and Section 4.2 proves only the first-derivative bound, stating that the remaining bounds can be derived using similar techniques; if those higher-order growth bounds fail, the Itô-formula comparison that delivers the drop is not justified.","fun_headline_variants_meta":{"raw":{"variants":["Guidance provably shifts samples toward classifier-favored regions","Theorem: guidance boosts average classifier confidence for any data","Guidance provably improves sample quality for general distributions","Average classifier confidence rises under guided diffusion, proven","New proof: guidance lowers reciprocal classifier probability on average"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000148,"raw_usage":{"total_tokens":1177,"prompt_tokens":920,"completion_tokens":257,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":536,"tokens_out":257,"duration_ms":2682,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:21:24.654528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the continuous-time guided reverse SDE with known scores for a simple non-Gaussian data distribution, such as a two-component Gaussian mixture with unequal covariance, using many noise seeds, and estimate $E[p_{c|X_1}(c|Y^w_1)^{-1}]$ for $w=0$ and several $w>0$. Theorem 1 predicts a strict decrease and a drop growing with the integrated squared score gap; seeing no decrease, or a non-monotone pattern in $w$, would falsify the average-improvement claim. A more direct check estimates both sides of the infinitesimal identity (11) at a fixed $t$ by Monte Carlo over a small time step.","supporting_citations":[{"cited_title":"and Nichol, A","cited_arxiv_id":null,"evidence_quote":"Provides the empirical phenomenon of classifier guidance improving sample quality that the theorem is meant to explain."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Inception Score, the classifier-probability-based quality metric that motivates the reciprocal-probability objective."}],"review_version":1}