{"id":"7d81544e-8435-46fa-ba06-44d713547a0d","arxiv_id":"2607.20674","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A scalable end-to-end training method for neural controllers with embedded control-barrier-function safety filters, demonstrated up to 1200 state dimensions and 400 control dimensions, with convergence guarantees under a nonsmooth-analysis framework.","lead":"This paper trains neural feedback controllers that must obey hard safety constraints by building a quadratic-program safety filter into the training loop, using a splitting scheme and a cheap approximate gradient to scale to problems with 1200 state variables. It reports nonsmooth convergence guarantees and demonstrations on multi-agent systems with up to 100 quadcopters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The descent proof hinges on Assumption 6(iii), a slack-coupling bound admitted to be intractable to verify and proxied only by Figure 4a; a violation at any visited point invalidates Lemmas 10/22 and Theorems 11/12.","rationale":"I agree with the reader's identification of Assumption 6(iii) as the weakest load-bearing point. The convergence theorem is the paper's main theoretical novelty, and its proof is built on a pointwise inequality that is admitted to be practically unverifiable and is supported only by a proxy measurement. This is not an internal contradiction, so rejection is not warranted, but it does mean the central guarantee is conditional in precisely the sense the reader stated. The concrete test would directly measure the assumption rather than a downstream consequence, and would tell us whether the condition holds on the problems where the method is claimed to work. I also considered the discrete-time inter-sample safety gap and the mismatch between the gradient-flow theorem and the Adam algorithm actually used; both are real concerns, but they weaken the practical or scope claims rather than the internal proof structure. Assumption 6(iii) is more fundamental: without it, Lemmas 10, 22, and Theorems 11 and 12 do not go through, so the theoretical contribution itself is at stake. The reader already captured this, so the verdict should remain CONDITIONAL.","tokens_in":38716,"tokens_out":7128,"duration_ms":66517,"concrete_test":"On the 6-agent double-integrator configuration (small enough to enumerate active-set selections), instrument the training loop to compute the LHS of (76) at a random sample of DYS fixed points: use automatic differentiation to extract M_y = ∂Tθ/∂y at the converged fixed point, form K_us, K_ss, M_sθ, M_uθ, compute Ξθ = K_us(I−K_ss)^−1 M_sθ(M_uθ)^T, and compare ||Ξθ||2 against ρθ from (76) using Gθ = M_uθ(M_uθ)^T. Record the maximum of ||Ξθ||2 − ρθ over the run. If the maximum is positive at any point where training continues to descend, Assumption 6(iii) is violated at points in the trajectory and the proof of Lemma 22 cannot be invoked; if it is negative throughout, the assumption is at least satisfied on the sampled run and Figure 4a is corroborated on a directly relevant quantity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical contribution—convergence to a Clarke stationary point—rests on Assumption 6(iii), made precise as (75)–(76), which requires ||Ξθ||2 ≤ ρθ for every Clarke selection M_y ∈ ∂C_y Tθ at every (θ,t,z). This is exactly the inequality used in Lemma 22 to prove pointwise alignment ⟨v,w⟩ ≥ 0; Lemma 10, Theorem 11, and Theorem 12 all inherit it. The paper states that Assumption 6(iii) is 'generally intractable to verify analytically' and points to Figure 4a, but Figure 4a shows JFB/AD gradient alignment—a consequence one would hope to prove—not a measurement of Ξθ or ρθ. The slack block K_us, the inverse (I−K_ss)^−1, and M_sθ can amplify the mismatch independently of the control block M_uθ; none of the contraction, rank, or conditioning assumptions control this term. If the inequality is violated at any point actually visited during training, Lemma 22 can fail and the descent/convergence guarantees are not in force. Thus the proof is conditional on a condition that is essentially as hard to certify as the alignment it is used to establish, and the paper provides no independent certification procedure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a scalable end-to-end training framework for neural feedback controllers with embedded CBF-QP safety filters. The method combines Davis–Yin splitting (DYS) for the forward safety projection with Jacobian-Free Backpropagation (JFB) for the backward pass, and is demonstrated on multi-agent single-integrator, double-integrator, and quadcopter problems with state dimensions up to 1200 and control dimensions up to 400. The theoretical contribution claims convergence of the training update to a Clarke stationary point under nonsmoothness induced by the safety layer, with the main result stated as Theorem 12. Numerical comparisons against AD-unrolled DYS and CVXPY Layers report lower memory usage, competitive costs, and stability in regimes where the baselines diverge.","tokens_in":39060,"tokens_out":5297,"duration_ms":48824,"significance":"If the claims hold, the paper addresses a real bottleneck: prior end-to-end training through differentiable CBF-QP layers has been restricted to small state dimensions, while this work demonstrates scaling to 1200-dimensional problems. The contraction analysis for every Clarke selection of the DYS operator (Corollary 9 and Theorem 19) is a solid technical contribution, and the memory/time comparisons in Tables 2 and Figure 5 support the scalability story. However, the convergence theory is significantly weaker than the abstract implies: the main theorem concerns continuous-time gradient flow under assumptions that are partly unverified, and the algorithm actually executed is discrete Adam. The paper is therefore valuable as an empirical scaling contribution with a partial theory, but the theoretical claims need to be reframed or substantially strengthened.","major_comments":[{"comment":"Assumption 6(iii) is the load-bearing hypothesis of the descent argument. The bound on ||Ξ_θ|| in Eq. (76) is used in Lemma 22 to prove the pointwise alignment ⟨v,w⟩ ≥ 0, and Lemmas 10, Theorem 11, and Theorem 12 all inherit it. The appendix remark after Lemma 22 states that this condition is 'generally intractable to verify analytically' and cites Figure 4a as empirical evidence. However, Figure 4a plots JFB/AD gradient alignment during training; it does not measure Ξ_θ as defined in Eq. (75), nor does it compare it to the threshold ρ_θ in Eq. (76). Because the assumption is essentially the statement that the gradient mismatch is below the value needed for the proof, the convergence guarantee is conditional on an unverified condition that is close in content to the desired conclusion. The manuscript should either provide a direct numerical certification of (75)–(76) along actual trainin","section":"Assumption 6(iii), Eqs. (75)–(76), Lemma 22 and Remark"},{"comment":"Theorem 12 proves convergence for the continuous-time gradient flow (Eq. (28)), but Algorithm 1 and all experiments use discrete Adam updates (line 23 of Algorithm 1; 'We use the Adam optimizer' in the Implementation Details). No argument is given to connect the gradient-flow result to the discrete stochastic optimizer actually used. The abstract and the contribution list state that the training update converges, which is misleading. The paper does note the gradient-flow limitation in a remark after Theorem 12, but the main text should be corrected to state clearly that the convergence guarantee does not apply to Algorithm 1, or the theory should be extended to a discrete-time setting.","section":"Theorem 12 vs. Algorithm 1 and Section 'Implementation Details'"},{"comment":"The descent theorem also depends on Assumptions 7 and 8, which are stated as regularity/variance bounds but are not independently verified. Assumption 8.2 in particular bounds deviations in terms of inf_{φ∈∂Cθ E_x[J_x]} ||φ|| and ||E_x[d^JFB_x]||, and the threshold δ_{v,θ} itself depends on ρ_θ − ||Ξ_θ|| from the unverified Assumption 6(iii). Figure 4b shows that time averages of gradients remain bounded, but this is far weaker than the specific inequalities required in Assumptions 7 and 8. The claim that Figure 4 'supports' these assumptions should be moderated; at present the proof is conditional on several assumptions that are not checked quantitatively.","section":"Assumptions 7 and 8; Figure 4b"}],"minor_comments":[{"comment":"The caption says the plot serves as a 'direct verification of Theorem 11.' Since the plot only measures gradient alignment on one training run and does not verify Assumptions 6–8, please rephrase to 'consistent with Theorem 11' or 'suggests descent direction is satisfied in practice.'","section":"Figure 4a caption"},{"comment":"The sentence 'We also tested multiple learning rates ranging from0.01and1→, es10 −5' is garbled. Please correct the typographical error and state the exact learning-rate range tested.","section":"Appendix, 'Additional Numerical Results'"},{"comment":"Table 1 reports point estimates for the baselines without standard errors or confidence intervals. Table 4 provides multi-seed statistics only for the proposed DYS-JFB method and only for three tasks. Adding repeated-seed results for the baselines where they converge would make the 'competitive performance' claim more supportable.","section":"Tables 1 and 4"},{"comment":"The 'stop-gradient' annotation is ambiguous. Please clarify that it applies to the DYS fixed-point iterations used to compute the forward solution, while the final DYS evaluation in line 16 enables gradient flow.","section":"Algorithm 1, line 12"},{"comment":"The claim that DYS 'requires no problem-dependent step-size tuning' should be phrased more carefully: the contraction result holds for any ζ∈(0,1), but the experiments fix ζ=0.5. This is a convenience, not a complete absence of hyperparameters.","section":"Discussion, 'Key advantages'"}],"recommendation":"major_revision","confidential_remarks":"The empirical scaling results are the strongest part of the paper and may well be publishable even if the theory is only conditional. However, the theoretical contribution as presented is not yet commensurate with the abstract's claims. I would encourage the editor to request a revision that either directly verifies Assumption 6(iii) along training trajectories (e.g., by computing the block matrices in Eq. (75) during rollout) or explicitly reframes the convergence result as conditional. The gap between gradient-flow theory and the Adam algorithm also needs to be addressed in the main text. I did not find evidence of problematic citation practices; the reliance on the authors' own prior JFB work is natural and properly cited."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, useful paper with a real scaling result, but the convergence theorem is conditional on a hard-to-verify coupling assumption and is proved for gradient flow rather than the Adam used in practice.\n\nThe genuinely new piece is the combination of Davis–Yin splitting and Jacobian-free backpropagation for trajectory-level policy training with embedded CBF-QP layers. The scaling to 1200 state dimensions and 400 control dimensions is impressive and well beyond the prior ~16-state limit. The contraction analysis for the DYS operator, showing a uniform bound across all Clarke generalized Jacobians, is a real technical contribution—prior work only handled a single selection. The experiments are well structured: they compare against automatic differentiation and implicit differentiation baselines, report memory and time, show barrier values stay positive, and are careful about the fairness of the comparisons.\n\nThe soft spots are real but not fatal. First, the convergence theorem is for continuous-time gradient flow, while Algorithm 1 uses Adam. The paper acknowledges this, but it means the headline 'convergence guarantees' do not cover the actual training procedure. Second, Assumption 6(iii) is load-bearing: the bound on the slack-induced gradient mismatch is exactly what makes Lemma 22 and the descent theorem work. The paper admits it is 'generally intractable to verify analytically' and points to Figure 4a, which shows JFB/AD gradient alignment rather than a direct check of the mismatch matrix. That is indirect evidence at best, and the stress-test note is right that the proof becomes conditional on a condition nearly as strong as the alignment it is meant to establish. Third, the 'hard safety guarantee' is continuous-time; the discrete-time implementation with RK4 does not come with an inter-sample guarantee, though the plotted barrier values are reassuring.\n\nNone of this is an internal contradiction. The theory is honest about its assumptions, and the empirical story stands. For the paper to fully deliver on its promises, a referee should ask for either a direct numerical check of Assumption 6(iii), a relaxation that makes it verifiable, or a scoped statement of the convergence result. Releasing code would also help substantially.\n\nWho should read this: anyone working on safe RL, CBF-based control, or differentiable optimization layers. It deserves a serious referee. I would send it out and likely accept conditionally after the theory-practice gap is addressed.","headline":"The paper's scaling result is the real contribution and the DYS contraction analysis is a genuine advance, but the convergence theorem rests on an unverified coupling assumption and is proved only for gradient flow, not the Adam used in practice.","tokens_in":39527,"tokens_out":3503,"would_cite":true,"duration_ms":30215,"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":"This paper claims that CBF safety filters can be learned end-to-end at 1,200 state dimensions by replacing exact QP differentiation with operator splitting and Jacobian-free gradients.","keywords":["control barrier functions","safety filters","end-to-end learning","Jacobian-free backpropagation","Davis-Yin splitting","nonsmooth optimization","Clarke stationarity","multi-agent control"],"falsifier":"Compute the norm of the gradient-mismatch matrix Xi_theta defined in Assumption 6(iii) for a trained policy at a fixed point where safety constraints are active; if it exceeds the threshold rho_theta, the pointwise alignment lemma and the descent theorem are violated.","tokens_in":38602,"feed_emoji":"🤖","tokens_out":5415,"duration_ms":46405,"temperature":0.7,"pith_summary":"This paper tries to show that hard safety constraints can be folded into neural feedback-policy training at scales previously out of reach. Prior end-to-end methods that differentiate through CBF-based quadratic-program safety filters were limited to about 16 state dimensions, while the authors report state dimensions up to 1,200 (100 quadcopters) and control dimensions up to 400. Their recipe combines an operator-splitting solver (Davis-Yin splitting) with an approximate gradient (Jacobian-Free Backpropagation), which avoids both unrolling many solver iterations and solving the large linear systems of implicit differentiation. The paper also provides a nonsmooth-analysis proof that the resulting biased gradient update is a descent direction and converges to a Clarke stationary point under gradient flow. If correct, this makes safety-critical end-to-end control learning practical in a regime where exact-gradient methods run out of memory or blow up.","feed_headline":"Safety-filter training scales to 1,200 state dimensions","feed_subtitle":"Jacobian-free gradients let neural controllers learn through hard barrier constraints without the old 16-dimension ceiling.","key_machinery":"The central object is the Davis-Yin splitting fixed-point operator T_theta for the lifted CBF-QP projection (control u plus slack s), whose fixed point gives the safety-filtered control. DYS decomposes the projection into simple closed-form projections and a gradient step, making each iteration cheap. The second mechanism is Jacobian-Free Backpropagation: instead of computing (I - dT/dy)^{-1} dT/dtheta, it uses only dT/dtheta, cutting per-step gradient cost from O((m+c)^3 + (m+c)^2 p) to O((m+c)p). Clarke generalized Jacobians handle the nonsmoothness of the projection, letting the proof show every generalized Jacobian of T_theta is a contraction and that the JFB update is a descent directio","core_discovery":"The authors claim that end-to-end training of neural feedback controllers with embedded control-barrier-function safety filters, previously limited to roughly 16 state dimensions, can be pushed to at least 1,200 state and 400 control dimensions without giving up the hard safety guarantee h(z(t)) >= 0. The route is to solve the CBF quadratic program by Davis-Yin splitting, a three-operator fixed-point iteration whose steps are cheap projections, and to differentiate through the fixed point with Jacobian-Free Backpropagation, which replaces the exact fixed-point derivative with a single partial derivative. They further claim convergence of the resulting trajectory-level training scheme: under","pith_inferences":["Inference: The convergence guarantee is for continuous-time gradient flow, not for the discrete Adam updates actually used in the experiments; a discrete-step version is a natural next step and is not supplied by the paper.","Inference: If the slack-variable alignment condition (Assumption 6(iii)) could be checked cheaply online, the method could be extended to adaptive step sizes or to detect when JFB's biased gradient stops being a safe descent direction.","Inference: The reported scalability still relies on fixed-point iterations per time step; for very large agent counts or tight horizons, the number of iterations, rather than memory, may become the limiting factor."],"forward_implications":["The reported ceiling of end-to-end CBF-QP policy learning moves from about 16 state dimensions to 1,200 state and 400 control dimensions, demonstrated on single-integrator, double-integrator, and quadcopter multi-agent tasks.","Safety remains hard by construction: the embedded filter enforces h(z(t)) >= 0 at every rollout, and the reported minimum barrier values stay positive.","Training memory is cut by up to roughly 10x compared with automatic-differentiation unrolling and implicit-differentiation baselines, and per-epoch time is lower whenever the baseline runs.","The convergence proof gives a theoretical foundation for training through nonsmooth safety layers, establishing convergence to a Clarke stationary point under the JFB gradient flow.","Because the forward solver and the differentiation rule are decoupled, any algorithm that computes the CBF-QP solution can be used in the forward pass while retaining the JFB backward pass and its guarantees."],"fun_headline_variants":["CBF safety filters scale to 1,200 dimensions","Jacobian-free training pushes safe control to 1,200 states","Operator splitting enables high-dim safety learning","From 16 to 1,200 dimensions in safe control learning","Scalable end-to-end learning with hard safety guarantees"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the slack variables introduced to convert the barrier inequality into an equality do not distort the approximate gradient too much; the authors state this condition is generally intractable to verify analytically and support it only with numerical experiments, so the descent and convergence theorems stand or fall on that alignment.","fun_headline_variants_meta":{"raw":{"variants":["CBF safety filters scale to 1,200 dimensions","Jacobian-free training pushes safe control to 1,200 states","Operator splitting enables high-dim safety learning","From 16 to 1,200 dimensions in safe control learning","Scalable end-to-end learning with hard safety guarantees"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1729,"prompt_tokens":646,"completion_tokens":1083,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":390,"completion_tokens_details":{"reasoning_tokens":1003}},"tokens_in":390,"tokens_out":1083,"duration_ms":8171,"temperature":1.0,"reasoning_tokens":1003,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:40:22.208647+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the norm of the gradient-mismatch matrix Xi_theta defined in Assumption 6(iii) for a trained policy at a fixed point where safety constraints are active; if it exceeds the threshold rho_theta, the pointwise alignment lemma and the descent theorem are violated.","supporting_citations":[],"review_version":1}