{"id":"4180c69a-77be-475b-ae4c-ff5b1ba983b8","arxiv_id":"2506.12544","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Constrained Diffusers enforces trajectory constraints on pre-trained diffusion models without retraining by replacing the reverse process with constrained Langevin sampling.","lead":"This paper adds safety constraints to pre-trained diffusion planners by treating the denoising process as constrained Langevin sampling, using projected, primal-dual, and augmented Lagrangian updates so no retraining is needed. Safety is enforced directly on generated trajectories, and an inverse dynamics model turns those trajectories into control actions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DDPM reverse update (Eq. 4) is not equivalent to the SGLD update (Eq. 6) that all three constrained samplers build on; the Tweedie substitution leaves a factor 1/sqrt(1-beta_t) and a beta_t score coefficient, not beta_t/2, so Theorem 4.1 addresses a different process.","rationale":"The reader's weakest assumption correctly identifies the Eq. (4)-Eq. (6) equivalence as the foundation of all three constrained samplers. My read sharpens this: the discrepancy is not only the omitted 1/sqrt(1-beta_t) scaling; the score coefficient is beta_t in the derived update and beta_t/2 in Eq. (6), so the drift is off by a factor of two plus an extra (beta_t/2)x_t term. This makes the theoretical convergence claim (Theorem 4.1) inapplicable to the algorithm as written, independent of the missing appendices. I do not see a reason to move the verdict: the experiments provide partial empirical support, and the concerns can be addressed by including the missing proofs, correcting or justifying the Langevin approximation, and tempering the claims. Hence the conditional verdict stands.","tokens_in":10642,"tokens_out":10942,"duration_ms":130103,"concrete_test":"Run a controlled experiment on a 2D Gaussian mixture with known score: generate 10,000 samples using the exact DDPM update (4) with the true score substituted for the noise estimate, and 10,000 samples using Eq. (6) with the same beta_t schedule, then compute the 2-Wasserstein distance between the two sample sets. If the distance is larger than a few percent of the data standard deviation, the updates are not interchangeable, and the constrained samplers are not solving problem (9) for the pretrained distribution. As a secondary check, verify whether the beta_t schedule satisfies the stochastic-approximation conditions (sum beta_t = infinity, sum beta_t^2 < infinity) used by the cited convergence theorems.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction depends on treating Eq. (6) as the reverse process of the pretrained diffusion model. Substituting the Tweedie score relation (5) into the standard DDPM update (4) yields x_{t-1} = (1/sqrt(1-beta_t))(x_t + beta_t * grad log p_t(x_t)) + sqrt(beta_t) z, not x_{t-1} = x_t + (beta_t/2) * grad log p_t(x_t) + sqrt(beta_t) z. Expanding the prefactor to first order introduces an extra drift (beta_t/2) x_t and leaves the score coefficient as beta_t instead of beta_t/2. Because Eq. (6) is the base update for the projected, primal-dual, and augmented-Lagrangian methods in Eqs. (10), (13), and (16), all three algorithms enforce constraints on a process that is not the model's reverse diffusion. Consequently, the KL(q||p) objective in (9) is not the objective actually being optimized, and Theorem 4.1's convergence statement cannot be transferred from the Langevin literature (e.g., [13,14]) without an additional argument; the cited theory is for stationary, time-homogeneous dynamics, while the reverse diffusion is finite-horizon and time-inhomogeneous. This is a load-bearing technical gap, not merely a missing appendix.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Constrained Diffusers, a framework for enforcing safety and trajectory constraints on pre-trained diffusion planning models without retraining. It rewrites the denoising update as a Langevin sampling step and then applies projected, primal-dual, and augmented-Lagrangian updates (Eqs. (10), (13), (16)) to the reverse diffusion process. Discrete control barrier functions and an inverse dynamics model are introduced for closed-loop safety. Experiments on Maze2D, MuJoCo locomotion, and PyBullet ball running compare the three variants against Diffuser, Conditional Diffuser, and SafeDiffuser. Theorem 4.1 claims almost-sure convergence to a local saddle point / local optimum of the constrained sampling problem (9) as T goes to infinity.","tokens_in":10946,"tokens_out":8462,"duration_ms":96507,"significance":"If the derivations were correct, the framework would be a practically attractive plug-and-play route to constraint satisfaction without retraining: the three algorithms are simple to implement, the DCBF formulation matches trajectory-level safety requirements, and the empirical results show near-zero violations with lower computation time than SafeDiffuser. The paper also deserves credit for evaluating both static and time-varying constraints across several benchmarks. However, the validity of the central 'Langevin reformulation' and of the convergence theorem is currently not established; the missing appendices and the algebraic gap in Section 2.1 mean that the paper's theoretical value cannot be assessed as submitted.","major_comments":[{"comment":"The claimed reformulation of the DDPM update (4) as the Langevin update (6) is algebraically incorrect. Substituting the Tweedie score relation (5) into (4) gives x_{t-1} = (1/sqrt(1-beta_t))(x_t + beta_t grad_{x_t} log p_t(x_t)) + sqrt(beta_t) z, not Eq. (6). Expanding 1/sqrt(1-beta_t) to first order adds a drift (beta_t/2)x_t and leaves the score coefficient as beta_t rather than beta_t/2, so Eq. (6) does not describe the reverse process of the pretrained DDPM. Because Eqs. (10), (13) and (16) all use Eq. (6), the constrained sampling procedure is not sampling from the distribution p(x_0:T) defined by the original diffusion model, and the KL objective in (9) is not the objective actually being optimized. This gap also prevents transferring the convergence guarantees of [12-14] to the proposed updates.","section":"Section 2.1, Eqs. (4)-(6)"},{"comment":"Theorem 4.1 is the central theoretical claim of the paper, but its proof is deferred to Appendix E, and Appendices D and F, which are supposed to contain the analyses for the projected and ALM methods, are not present in the manuscript. The 'mild conditions' of the theorem are not stated, and the object whose convergence is claimed (the sequence of measures induced by the sample path, or the sample path itself?) is never defined. As submitted, the theorem is an assertion rather than a verifiable result, and the same problem affects the convergence statements in Sections 4.1 and 4.3.","section":"Section 4.2, Theorem 4.1 and Section 4.1/4.3"},{"comment":"The dual update in Eq. (14) and the slack/dual updates in Eqs. (17)-(18) require the expectation E_{x~q}[g(x_t)] over the current distribution, but the algorithm maintains only a single trajectory. If this expectation is approximated by the instantaneous value g(x_t), the resulting stochastic approximation has a bias that must be analyzed; if a batch-based estimator is intended, it is not specified. Without an explicit oracle for these expectations, the 'almost surely' convergence statement of Theorem 4.1 cannot be evaluated, and the pseudocode in Algorithm 1 is incomplete.","section":"Section 4.2, Eqs. (13)-(14)"},{"comment":"The reverse diffusion process is finite-horizon and time-inhomogeneous: the score p_t changes with t and the index descends from T to 0. The Langevin convergence results in [12-14] apply to stationary, time-homogeneous dynamics with step sizes tending to zero. The phrase 'as T goes to infinity' does not by itself bridge this gap because no rescaling or time-change of the diffusion schedule is provided. Even if the missing appendix were supplied, Theorem 4.1 would need a separate argument for the time-inhomogeneous finite-horizon setting.","section":"Section 4.2, Theorem 4.1 and Section 2.1"}],"minor_comments":[{"comment":"The line `u_tau = IDM(x0_0, x1_0)` uses fixed indices 0 and 1 rather than the environment time index tau; presumably x_tau^0 and x_{tau+1}^0 are intended. Please correct.","section":"Algorithm 1, line 9"},{"comment":"In Table 2, Projected has a per-step time of 0.9305 s for Hopper and 0.9049 s for Swimmer, which is roughly 300 times larger than Primal-Dual and ALM, yet the discussion claims computational efficiency for the proposed methods; the cause of this discrepancy (presumably the cost of projection) should be discussed explicitly.","section":"Section 6, Table 2"},{"comment":"The index convention in the updates is confusing: the iteration descends from T to 0, but the primal update (13) and dual update (14) both write t-1 on the left while using lambda_t on the right. The direction of iteration and the role of lambda_t should be clarified to make the algorithm unambiguous.","section":"Section 4, Eqs. (13)-(14)"}],"recommendation":"major_revision","confidential_remarks":"The main contribution is close to existing constrained sampling work [12-14]; the paper's added value is the adaptation to diffusion. Given the technical gap in Eq. (6) and the missing appendices, I would not accept in the current form. If the appendices contain rigorous proofs and the Langevin reformulation is corrected or explicitly redefined as a different process, a resubmission would be worth another review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core engineering idea is good, and the experiments mostly back it up: instead of retraining a diffusion planner, you constrain the reverse process by borrowing projected, primal-dual, and augmented-Lagrangian Langevin sampling, and you use discrete CBFs plus an inverse dynamics model to make it work online. On Maze2D, MuJoCo, and PyBullet, the primal-dual and ALM variants reliably reduce constraint violations at roughly the same per-step cost as a plain diffuser, which beats SafeDiffuser's QP-based cost by a large margin. That is a practical result.\n\nWhat is genuinely new is the combination itself. The individual pieces—projected LMC, primal-dual sampling, ALM—are not new, and the paper says so plainly via its citations. The contribution is applying them to a pretrained DDPM and linking the constraints to DCBFs. The experiments support the claim that this works in practice, at least for the tested benchmarks.\n\nThe soft spot is where the stress-test note lands, and it is load-bearing. Substituting Tweedie's formula into the standard DDPM update (4) does not give Eq. (6). You get x_{t-1} = (1/sqrt(1-β_t))(x_t + β_t ∇log p) + sqrt(β_t)z, which, expanded to first order, carries an extra (β_t/2)x_t drift and a score coefficient β_t rather than β_t/2. So Eq. (6) is not the model's reverse diffusion, and all three constrained samplers inherit that mismatch. Theorem 4.1 is then a statement about a different process. The proof is also deferred to a missing appendix, and the cited primal-dual convergence result is for time-homogeneous stationary Langevin dynamics, not a finite-horizon, time-inhomogeneous diffusion. A transfer argument is needed and not supplied.\n\nThere are smaller issues too. The abstract's \"less computation time\" is only true for the primal-dual and ALM variants; the projected method costs about the same as SafeDiffuser per step (0.93s) in the locomotion tasks. And the Maze2D large results show large distances relative to the baseline, so the tradeoff between staying near the data distribution and satisfying constraints deserves a closer look.\n\nWho is this for? People building diffusion-based planners who need a safety plug-in and are willing to treat the theory as suggestive rather than proven. I would not cite it for the convergence claim, but I would cite it as evidence that constrained Langevin sampling can be made to work on DDPMs in practice. I agree with the reader's conditional verdict.\n\nRecommendation: send it to peer review. The flaws are fixable in principle—re-derive or justify the Langevin step, supply the appendices, and temper the claims. A serious referee would be the right way to find out if the gap is patchable.","headline":"A promising plug-and-play safety idea for diffusion planners is undermined by a mis-derived Langevin step and a missing proof, but the experiments and DCBF integration make it worth reviewing.","tokens_in":11474,"tokens_out":2973,"would_cite":false,"duration_ms":33403,"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":"The paper shows that safety constraints can be enforced on pretrained diffusion planners by replacing the reverse denoising updates with constrained Langevin sampling steps—projected, primal-dual, and augmented Lagrangian—with no…","keywords":["diffusion models","constrained sampling","Langevin dynamics","control barrier functions","safe planning","trajectory optimization","inverse dynamics","safety-critical control"],"falsifier":"Run the primal-dual or augmented-Lagrangian sampler on a two-dimensional Gaussian with a known score and a nonconvex constraint such as the outside of a disk, and measure the fraction of final samples that violate the constraint as the diffusion horizon increases; if the violation fraction does not approach zero while the samples stay close to the unconstrained target, the claimed equivalence to constrained Langevin sampling is false.","tokens_in":10408,"feed_emoji":"🛡️","tokens_out":9699,"duration_ms":101113,"temperature":0.7,"pith_summary":"This paper is trying to establish that a pretrained diffusion model for trajectory planning can be made constraint-satisfying at deployment time, with no retraining and no change to the network architecture. The central move is to reinterpret the reverse denoising loop as stochastic gradient Langevin sampling and to insert three constrained-optimization mechanisms—projection, primal-dual updates, and an augmented Lagrangian—so that every denoising step steers the trajectory toward the safe set while still following the learned data distribution. Discrete control barrier functions express safety requirements on the generated trajectory, and an inverse dynamics model converts the safe states into actions for closed-loop control. If the claim holds, safety becomes a plug-in property of the sampler rather than something that must be learned or solved as a separate quadratic program at each step.","feed_headline":"Safe diffusion planning without retraining or QP solvers","feed_subtitle":"Three constrained-sampling updates keep generated trajectories safe at lower compute than per-step safety solvers.","key_machinery":"The load-bearing identity is the reformulation of the DDPM reverse update as Langevin sampling: $x_{t-1} = x_t + \\frac{\\beta_t}{2}\\nabla_{x_t}\\log p(x_t) + \\sqrt{\\beta_t}z$, where the score is recovered from the denoiser as $\\nabla_{x_t}\\log p(x_t) = -\\epsilon_\\theta(x_t,t)/\\sqrt{1-\\bar\\alpha_t}$. This identity turns the pretrained denoiser into a score oracle, onto which the paper grafts three constraint mechanisms: a projection operator $\\Pi_C$ onto the feasible set; a primal-dual step that carries a Lagrange multiplier $\\lambda$ for $\\mathbb{E}_{x\\sim q}[g(x)] \\le 0$; and an augmented-Lagrangian step with slack variable $s$ and penalty $\\rho$. Safety at the trajectory level is expressed by the discrete control barrier function inequality $h(x_{\\tau+1}) \\ge (1-\\alpha)h(x_\\tau)$, and an inverse dynamics model $u_\\tau = \\mathrm{IDM}(x_\\tau, x_{\\tau+1})$ maps the resulting safe state transitions to control actions.","core_discovery":"The central claim is that constrained planning can be solved entirely inside the reverse diffusion process. The paper rewrites the standard DDPM reverse update as a Langevin sampling step whose score is supplied by the neural denoiser, so that adding safety constraints reduces to modifying each denoising update: project onto the constraint set, subtract the constraint gradient weighted by a Lagrange multiplier, or apply an augmented-Lagrangian penalty with a slack variable. Theorem 4.1 states that the primal-dual version converges almost surely to a local saddle point—equivalently a local optimum of the minimum-KL constrained problem—as the diffusion horizon $T$ tends to infinity. The experiments in Maze2D, locomotion, and a moving-obstacle ball-running task are offered as evidence that the finite-horizon versions keep constraint violations near zero at computation times close to the unconstrained baseline and far below the CBF-QP baseline.","pith_inferences":["Editorial inference: the same constrained-Langevin surgery could be applied to any differentiable constraint on any diffusion generative model, not just trajectory planners; image and video editors could enforce geometric or semantic constraints without retraining.","Editorial inference: Theorem 4.1 is asymptotic in the diffusion horizon and assumes the score is the true score; finite-horizon runs with learned scores need empirical safety margins because the theorem does not quantify the gap.","Editorial inference: the method assumes an inverse dynamics model exists and that inputs are unconstrained; systems with actuation limits or underactuation would require folding input constraints into the DCBF condition or adding a feasible-input projection.","Editorial inference: because safety is enforced only during generation, a miss by the learned score or the constraint gradient is uncaught; layering a lightweight control-barrier filter on the executed action would bound worst-case violations."],"forward_implications":["A pretrained trajectory diffuser can serve multiple safety specifications at deployment; only the sampling loop changes.","Constraint satisfaction no longer requires solving a quadratic program at every diffusion step, so safe planning runs at near-unconstrained speed.","Time-varying constraints, such as moving obstacles, can be handled online by recomputing the DCBF condition from the latest observation and resampling.","In the infinite-horizon limit with the correct score, the primal-dual updates converge almost surely to a local saddle point of the constrained sampling problem.","Closed-loop control inherits safety only when the inverse dynamics model accurately maps safe state transitions to actions."],"supporting_citations":[{"why":"Supplies the DDPM forward and reverse update that the constrained sampler modifies.","marker":"[2]"},{"why":"Supplies the trajectory-diffusion planner that the method constrains.","marker":"[3]"},{"why":"Supplies the Conditional Diffuser baseline that struggles with nonconvex constraints.","marker":"[4]"},{"why":"Supplies the score-matching and stochastic gradient Langevin dynamics link used to reinterpret denoising.","marker":"[9]"},{"why":"Supplies the projected Langevin Monte Carlo method that the projection update adapts.","marker":"[12]"},{"why":"Supplies the primal-dual constrained sampling formulation and the local-saddle-point optimality concept behind Theorem 4.1.","marker":"[13]"},{"why":"Supplies an alternative variational gradient view of constrained sampling used as a basis for the dual updates.","marker":"[14]"},{"why":"Supplies the score/noise reparameterization that turns the DDPM update into a Langevin step.","marker":"[17]"},{"why":"Supplies discrete control barrier functions used as trajectory-level safety constraints.","marker":"[19]"},{"why":"Supplies the SafeDiffuser CBF-QP baseline that the method compares against on computation time.","marker":"[24]"}],"fun_headline_variants":["Constrained diffusers plan safely at lower compute than QP","Diffusion planning gains safety without retraining or QP solvers","Constrained Langevin sampling yields safe diffusion plans","Diffusers enforce constraints in denoising steps for safe planning","Embedding constraints into diffusion updates speeds up safe planning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the DDPM reverse update is exactly a Langevin sampling step with the learned score and the same step size, despite dropping the $1/\\sqrt{1-\\beta_t}$ rescaling; if that step-size or score approximation is materially wrong, the safety and convergence claims do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Constrained diffusers plan safely at lower compute than QP","Diffusion planning gains safety without retraining or QP solvers","Constrained Langevin sampling yields safe diffusion plans","Diffusers enforce constraints in denoising steps for safe planning","Embedding constraints into diffusion updates speeds up safe planning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000371,"raw_usage":{"total_tokens":1945,"prompt_tokens":864,"completion_tokens":1081,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":997}},"tokens_in":480,"tokens_out":1081,"duration_ms":12129,"temperature":1.0,"reasoning_tokens":997,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:46:25.492305+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the primal-dual or augmented-Lagrangian sampler on a two-dimensional Gaussian with a known score and a nonconvex constraint such as the outside of a disk, and measure the fraction of final samples that violate the constraint as the diffusion horizon increases; if the violation fraction does not approach zero while the samples stay close to the unconstrained target, the claimed equivalence to constrained Langevin sampling is false.","supporting_citations":[{"cited_title":"Sampling from a log-concave distribution with projected langevin monte carlo","cited_arxiv_id":null,"evidence_quote":"Supplies the projected Langevin Monte Carlo method that the projection update adapts."},{"cited_title":"Constrained sampling with primal-dual langevin monte carlo","cited_arxiv_id":null,"evidence_quote":"Supplies the primal-dual constrained sampling formulation and the local-saddle-point optimality concept behind Theorem 4.1."},{"cited_title":"Sampling with trusthworthy constraints: A variational gradient framework","cited_arxiv_id":null,"evidence_quote":"Supplies an alternative variational gradient view of constrained sampling used as a basis for the dual updates."},{"cited_title":"Discrete control barrier functions for safety-critical control of discrete systems with application to bipedal robot navigation","cited_arxiv_id":null,"evidence_quote":"Supplies discrete control barrier functions used as trajectory-level safety constraints."}],"review_version":1}