{"id":"2a34aa12-0e50-4358-bce7-ecb885b8e4ad","arxiv_id":"2412.07544","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A contractive dynamical-system policy, built from recurrent equilibrium networks and coupling layers, that guarantees out-of-sample recovery and is shown to beat stable baselines on imitation benchmarks.","lead":"This paper introduces SCDS, a policy model built from contractive dynamical systems that guarantees all rollouts converge to a single trajectory even from unseen starting states. The authors report large out-of-sample error reductions on handwriting, manipulation, and navigation benchmarks, and offer a theoretical bound on deployment loss.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 2.1 is not established for unconstrained coupling layers: a RealNVP-style map can destroy the Euclidean contraction property, so Theorem 4.1's premise is not guaranteed.","rationale":"The reader's stated weakest assumption — the invalid MSE triangle inequality in the proof of Theorem 4.1 — is real and by itself invalidates the proof as written. My pass, however, located a more basic gap: the policy's global Euclidean contractivity is not actually guaranteed for the proposed coupling layers. App. A.1 only proves the linear-projection step and then invokes bijective-map preservation of contraction; in the Euclidean norm with a uniform α, that preservation requires global bi-Lipschitzness, which unconstrained RealNVP coupling layers do not possess. The explicit two-point computation with g(x1,x2)=(x1, x2 e^{x1}) and ż=-z shows the contraction ratio can be made arbitrarily large, so no α satisfying Definition 2.1 exists. Consequently the central theoretical contribution — a certifiable, unconstrained contractive policy class with an OOS loss upper bound — is not established as written. To be fair, the empirical study is substantial (LASA, Robomimic, Isaac Lab deployment), the REN foundation is real, and the proof gaps are likely repairable: restrict coupling layers to globally bi-Lipschitz maps with computable constants, replace the MSE triangle inequality with the valid factor-2 inequality, and correct the R vs. R² step in Eq. 19. But in the current manuscript, the advertised guarantees do not follow, so the reader's REJECT verdict should stand.","tokens_in":26581,"tokens_out":25837,"duration_ms":269720,"concrete_test":"Run the explicit counterexample: latent dynamics ż=-z in R² (contractive with γ=1, α=1), output map g(x1,x2)=(x1, x2 e^{x1}) — one RealNVP coupling layer with scale function s(x1)=x1 and zero shift. For initial conditions z_a(0)=(0,M) and z_b(0)=(-1, M e), the initial output distance is 1. At t=0.1, the output distance is M e^{-0.1}(e^{1-e^{-0.1}}−1) ≈ 0.0905 M, while contraction with any fixed α would require it to be ≤ α e^{-0.1}. Taking M large violates the inequality for every finite α, so Definition 2.1 fails and Proposition 2.1 is falsified for the proposed parameterization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's main theoretical claim is that ϕθ is contractive for every θ with a global α in Definition 2.1, and that this contractivity yields the OOS loss bound. The load-bearing step is Proposition 2.1 (App. A.1): the output map (Eq. 5) preserves contraction. The proof handles the linear projection but then cites Manchester & Slotine for 'bijective maps preserve contractivity.' That result preserves contraction in a pulled-back metric; it does not provide a uniform Euclidean α for the composed output map under the L2 norm of Definition 2.1. This is not a minor technicality: an unconstrained RealNVP coupling layer g(x1,x2)=(x1, x2 e^{x1}) is bijective but not globally bi-Lipschitz. With latent dynamics ż=-z (γ=1, α=1), initial states (0,M) and (-1, M e) have output distance 1 at t=0, but at t=0.1 the output distance is ≈0.0905 M, so the ratio to e^{-0.1} times the initial distance grows linearly with M. Hence no single finite α satisfies Eq. 2 for all initial conditions. The loss-bound proof also has independent defects — Eq. 16 applies a triangle inequality to MSE, which is invalid, and Eq. 19 bounds Σ||y0−y0^m||² by R²/M although Assumption 4.1 only gives ≤R — but even if those were repaired, the upper bound would rest on an unproven contractivity premise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SCDS, a state-only imitation learning method whose policy is a continuous-time dynamical system composed of a contractive recurrent equilibrium network (REN) in a latent space, a linear projection, and a stack of RealNVP coupling layers. The authors claim that this composition is contractive for every parameter choice, enabling unconstrained optimization, and they derive upper bounds on the worst-case and expected out-of-sample loss (Theorem 4.1 and Corollary 4.1.1). The method is evaluated on LASA and Robomimic datasets, with additional deployment in the Isaac Lab simulator, reporting strong out-of-sample recovery compared with stable and non-stable baselines.","tokens_in":26924,"tokens_out":7226,"duration_ms":74136,"significance":"If the theoretical results were correct, this would be a valuable contribution: a contractive-by-construction policy class with unconstrained training and explicit worst-case out-of-sample guarantees would address a real gap in imitation learning. The empirical results are promising, the experiments are fairly extensive, and the authors ship code. However, the central theoretical claims rest on two independent mathematical errors—in the preservation of Euclidean contractivity through coupling layers and in the use of a triangle inequality for the MSE loss—so the main advertised guarantee is not established. The empirical findings may still be of interest, but the paper's headline contribution is the rigorous OOS bound, which is currently unsupported.","major_comments":[{"comment":"Theorem 4.1's proof uses the 'triangle inequality for MSE' to claim ℓ(ŷ, y^m) ≤ ℓ(ŷ, ŷ^m) + ℓ(ŷ^m, y^m). Mean squared error does not satisfy the triangle inequality: for scalar values 0, 1, 2, MSE(0,2)=4 while MSE(0,1)+MSE(1,2)=2. The missing cross term is not bounded in the proof, so the displayed upper bound in Theorem 4.1 and the bound in Corollary 4.1.1 do not follow from the preceding inequalities. This is a load-bearing step for the paper's main 'rigorous' deployment guarantee.","section":"App. A.2, Eq. (16)"},{"comment":"The preservation of Euclidean contractivity through the output map is not established. First, Lemma A.1 gives ||Pv||²/||v||² ≥ σ_min(P)², hence ||v|| ≤ ||Pv||/σ_min(P), so the denominator in Eq. (13) should be σ_min(P), not σ_min(P)². Second, the paper explicitly advocates latent dimensions Nz > Ny for expressiveness, in which case P ∈ R^{Ny×Nz} necessarily has a nontrivial null space, σ_min(P)=0, and the displayed bound is vacuous. Third, the citation to Manchester & Slotine for 'bijective maps preserve contractivity' concerns contraction in a pulled-back metric, not a single global Euclidean α in the norm of Definition 2.1. A RealNVP coupling layer such as (x1,x2) ↦ (x1, x2 e^{x1}) is bijective but not bi-Lipschitz; combined with a contracting linear latent flow, the ratio of output-trajectory distance to e^{-γt} times the initial output distance can be made arbitrarily large, so no finite α independent of the initial condition satisfies Eq. (2). Consequently the premise of Theorem 4.1—that the composed policy ϕθ is contractive in the output space—is not guaranteed for the proposed architecture.","section":"App. A.1, Proposition 2.1 and Eq. (13)"},{"comment":"The constant α in the reported upper bound is not a verified certificate: the paper states it is 'approximated empirically using a Monte Carlo approach' after training. Since Proposition 2.1 does not provide a computable α from the network parameters, the numerical values labeled L^{MSE}_{ub} are post-hoc estimates rather than upper bounds certified by the theory. The claim that the bound 'rigorously establish[es] the reliability of our method in deployment' is therefore stronger than what is delivered.","section":"Sec. 5, 'Upper bound' paragraph and Corollary 4.1.1"}],"minor_comments":[{"comment":"Definition 2.1 allows an arbitrary Lp norm, but the paper then fixes the L2 norm; this should be stated in the definition itself to avoid ambiguity in Eq. (2) and in the proofs.","section":"Definition 2.1"},{"comment":"When Nz > Ny, P cannot have full column rank, so z(0) is only a least-squares approximation and ŷ(0) ≠ y0 in general; the paper acknowledges this, but the subsequent contractivity statements for output trajectories starting from exact y0 require an additional argument that is not provided.","section":"Sec. 2.3, Eq. (6)"},{"comment":"The set in Eq. (12) is called a 'multi-focal ellipse', but for M>2 it is a convex multi-center region rather than an ellipse in the usual sense; also App. A.2 refers to it as a 'conic initial set', which appears to be a typo.","section":"Assumption 4.1 and App. A.2"},{"comment":"The geometric-series factor (e^{-2γ}-1)/(e^{-2γ/H}-1) is positive because both numerator and denominator are negative; the authors should state this explicitly to avoid confusion about the sign of the bound.","section":"Theorem 4.1"},{"comment":"The L^{MSE}_{ub} values would be more informative if the paper reported the number of Monte Carlo samples, the variance of the α estimate, and a comparison of the bound with the observed OOS loss under the same distribution; the claim that the bound is 'relatively tight' is not quantified.","section":"Sec. 5, Table 1"}],"recommendation":"reject","confidential_remarks":"The two load-bearing proof errors—the MSE triangle inequality and the preservation of a uniform Euclidean contraction constant through unbounded coupling layers—are not local typos; repairing them requires either altering the architecture (e.g., restricting coupling layers to be globally bi-Lipschitz with certified constants, which would weaken the 'unconstrained optimization' claim) or changing the loss to a metric and re-proving the bound. Given that the paper's central advertised contribution is the theoretical OOS guarantee, I recommend rejection. The empirical study could form the basis of a revised submission if the theoretical claims are substantially reworked."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth reading, but the central theory does not hold up. The architecture—continuous RENs plus coupling layers, state-only training, latent-space learning—is a genuine and useful combination, and the OOS improvements (up to 2.5x over baselines on Robomimic) are impressive. The empirical work is careful: multiple datasets, ablations, simulator deployment, and a public codebase. Credit where due: this is a solid systems/empirical contribution.\n\nThe problem is the claimed guarantee. Proposition 2.1 says the output map preserves contractivity for every θ with the same Euclidean α. The proof leans on Manchester & Slotine's result that bijective maps preserve contractivity, but that is in a pulled-back metric, not the L2 norm used in Definition 2.1. A RealNVP coupling layer like g(x1,x2)=(x1, x2 e^{x1}) is bijective but not globally bi-Lipschitz; with latent dynamics ż=-z and initial states (0,M) and (-1,Me), the output distance ratio grows unboundedly with M. So no finite global α exists in general. This is not a technicality—it removes the premise of Theorem 4.1.\n\nTheorem 4.1's proof also applies the triangle inequality to MSE. Squared Euclidean distance does not satisfy it; the cross term is silently dropped. So the upper bound is not derived. The stress-test also flags Eq. 19, but on close reading that step seems fine: the M in the denominator is accounted for, and the inequality chain works. The real defects are the two above.\n\nA caveat: the bound in Corollary 4.1.1 uses α estimated post-training, which is fine for reporting an empirical bound, but it does not make the bound a priori. That is a lesser issue.\n\nWhere does this leave the paper? The empirical claims stand, but the advertised “rigorous OOS guarantee” is unproven. The paper would be much stronger if the authors either added explicit Lipschitz constraints on the coupling layers (e.g., spectral normalization) or reframed the results as empirical findings plus a heuristic bound. As written, the load-bearing theory fails.\n\nWho should read it: anyone building contractive imitation policies will find the architecture and experiments useful. The proof is not reliable. I would not cite the theory, but I might cite the empirical architecture with a caveat. This deserves a serious referee—the flaws are subtle enough that expert judgment is needed—but as a reviewer I would reject in current form.","headline":"Useful architecture and strong OOS results, but the central contractivity guarantee and the OOS loss bound are not proven; the paper should not be accepted as-is.","tokens_in":27442,"tokens_out":2536,"would_cite":false,"duration_ms":28148,"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 imitation policies shaped as contractive dynamical systems recover out-of-sample states with a provable upper bound on loss, because every rollout contracts exponentially toward the expert trajectory.","keywords":["imitation learning","contractive dynamical systems","out-of-sample recovery","recurrent equilibrium networks","coupling layers","neural ordinary differential equations","soft dynamic time warping","robotic manipulation"],"falsifier":"Take $a=0$, $b=1$, $c=2$: the squared-error loss from $a$ to $c$ is $4$, while the sum of the losses via $b$ is $1+1=2$, so the triangle inequality used in the proof of Theorem 4.1 (Appendix A.2, Eq. (16)) fails for MSE. A complete falsifier would be an explicit contractive policy and an out-of-sample state whose measured MSE loss exceeds the claimed upper bound.","tokens_in":26405,"feed_emoji":"🤖","tokens_out":8701,"duration_ms":78710,"temperature":0.7,"pith_summary":"The paper targets a specific failure mode of imitation learning: a policy trained on expert demonstrations can become unreliable when deployment starts from a state outside the training data. Its proposal, SCDS, is an imitation policy built as a contractive dynamical system, so that any rollout exponentially converges toward a single trajectory regardless of initial condition or perturbation. Contractivity is built into the architecture rather than enforced during training: a recurrent equilibrium network in a latent space, followed by a linear projection and invertible coupling layers, remains contractive for every parameter choice, which allows unconstrained gradient-based optimization. The paper derives an upper bound on the worst-case and expected out-of-sample mean-squared-error loss under an initial-state locality assumption, turning contractivity into a quantitative recovery certificate. A sympathetic reading is that this gives imitation learning a transient-phase guarantee that asymptotic stability alone does not provide.","feed_headline":"Contractive policies guarantee out-of-sample recovery","feed_subtitle":"Built as contractive dynamical systems, imitation policies get provable out-of-sample recovery","key_machinery":"The load-bearing construction is the SCDS policy, a continuous-time dynamical system $\\dot{z} = f_\\theta(z)$ in a latent space, paired with an output map made of a linear projection and $K$ invertible coupling layers. The latent dynamics are a recurrent equilibrium network (REN), a nonlinear model whose parameterization guarantees contractivity at a chosen rate $\\gamma$ for all parameters $\\theta$; the coupling layers are bijective maps that preserve that contractivity, and the initial latent state is set with a pseudoinverse so the planned state approximates the observed initial state. This architecture lets the paper reduce the out-of-sample recovery question to a comparison of rollouts from nearby initial conditions, which contractivity then turns into an exponential-decay estimate used in Theorem 4.1.","core_discovery":"The central claim is that the policy in Eq. (7) is contractive for every parameter choice, with the contraction rate preserved through the output transformation, and that this contractivity directly yields a deployable recovery guarantee. Theorem 4.1 states that, when the out-of-sample initial state lies in a multi-focal ellipse around the training initial states, its MSE loss is bounded by a weighted sum of in-sample rollout errors plus a term that shrinks with the contraction rate and with the size of the uncertainty region; Corollary 4.1.1 extends the same bound to any distribution over that region. The paper presents empirical evidence that these guarantees translate into lower out-of-sample error than stability-only baselines on LASA handwriting motions and Robomimic manipulation tasks, including simulated deployment on a Franka arm and a Jackal robot.","pith_inferences":["The bound's proof uses the triangle inequality for the squared-error loss, which mean squared error does not satisfy; a direct corollary is that the certificate would become rigorous if the loss were replaced by a true metric, such as Euclidean distance, or if the proof were amended to account for the missing cross term.","Because the bound is computable from the current state, one could extend SCDS to an online safety monitor that pre-empts rollouts whose certified loss exceeds a threshold, an application the paper does not discuss.","The inverse-distance weighting in Eq. (9) is a nearest-neighbour-style soft partition of the initial-state space; it would be natural to test whether alternative kernels or learned weights tighten the bound while preserving the zero-loss-at-demonstration property.","For multi-modal expert behaviors, such as the Robomimic tasks where demonstrations do not share a target, a single contractive system can only contract to an average behavior; an extension would need mixture or mode-conditional contractive policies."],"forward_implications":["Any rollout started from an out-of-sample state contracts exponentially toward rollouts from in-sample states, so the transient phase is controlled and the target is reached from unseen initial conditions.","Because contractivity holds for every parameter value, training can be done with unconstrained gradient descent, avoiding the accuracy-versus-stability trade-offs of constrained optimization.","The bound in Theorem 4.1 is computable from precomputed in-sample errors and inverse-distance weights, so it can serve as a deployment-time certificate for a learned policy.","State-only training removes the need for expert velocity measurements, eliminating a common source of accumulated error in imitation learning.","Higher contraction rates or tighter initial-state uncertainty regions directly tighten the certified worst-case and expected loss."],"supporting_citations":[{"why":"Defines contractivity and the exponential-convergence inequality (Definition 2.1) that the whole policy and bound rest on.","marker":"Lohmiller & Slotine (1998)"},{"why":"Supplies the unconstrained continuous-time REN parameterization that makes the latent dynamics contractive by design.","marker":"Martinelli et al. (2023)"},{"why":"Establishes the REN model class and its representational and stability properties, which the paper extends to state outputs.","marker":"Revay et al. (2023)"},{"why":"Gives the result that bijective maps preserve contractivity, used in Proposition 2.1 for the coupling layers.","marker":"Manchester & Slotine (2015)"},{"why":"Provides the normalizing-flow and coupling-layer construction that forms the bijection block of the output map.","marker":"Papamakarios et al. (2021)"},{"why":"The Neural ODE framework used to solve the initial-value problem and backpropagate through rollouts.","marker":"Chen et al. (2018)"},{"why":"Supplies the differentiable soft-DTW loss used to compare rollouts with demonstrations of different lengths.","marker":"Cuturi & Blondel (2017)"},{"why":"Provides the LASA handwriting dataset and the stable-dynamical-system baseline context for empirical evaluation.","marker":"Khansari-Zadeh & Billard (2011)"},{"why":"Provides the Robomimic manipulation demonstrations used to evaluate high-dimensional performance.","marker":"Mandlekar et al. (2021)"}],"fun_headline_variants":["Provable out-of-sample recovery via contractive policies","Guaranteed recovery from unseen states via contractive imitation","Contractive imitation policies ensure efficient unseen-state recovery","Out-of-sample recovery made provable by contractive policies","Contractive imitation policies: efficient, provable OOS recovery"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The quantitative upper-bound theorem assumes that the squared-error loss satisfies the triangle inequality, but mean squared error does not: the cross term is dropped in the proof in Appendix A.2. If that assumption fails, the stated worst-case and expected-loss bounds do not follow from contractivity alone.","fun_headline_variants_meta":{"raw":{"variants":["Provable out-of-sample recovery via contractive policies","Guaranteed recovery from unseen states via contractive imitation","Contractive imitation policies ensure efficient unseen-state recovery","Out-of-sample recovery made provable by contractive policies","Contractive imitation policies: efficient, provable OOS recovery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00093,"raw_usage":{"total_tokens":3926,"prompt_tokens":831,"completion_tokens":3095,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":447,"completion_tokens_details":{"reasoning_tokens":3016}},"tokens_in":447,"tokens_out":3095,"duration_ms":20480,"temperature":1.0,"reasoning_tokens":3016,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:44:37.161332+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $a=0$, $b=1$, $c=2$: the squared-error loss from $a$ to $c$ is $4$, while the sum of the losses via $b$ is $1+1=2$, so the triangle inequality used in the proof of Theorem 4.1 (Appendix A.2, Eq. (16)) fails for MSE. A complete falsifier would be an explicit contractive policy and an out-of-sample state whose measured MSE loss exceeds the claimed upper bound.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines contractivity and the exponential-convergence inequality (Definition 2.1) that the whole policy and bound rest on."},{"cited_title":"Manchester, Luca Furieri, and Giancarlo Ferrari-Trecate","cited_arxiv_id":null,"evidence_quote":"Supplies the unconstrained continuous-time REN parameterization that makes the latent dynamics contractive by design."},{"cited_title":"Recurrent equilibrium networks: Flexible dynamic models with guaranteed stability and robustness","cited_arxiv_id":null,"evidence_quote":"Establishes the REN model class and its representational and stability properties, which the paper extends to state outputs."},{"cited_title":"Neural ordinary differential equations","cited_arxiv_id":null,"evidence_quote":"The Neural ODE framework used to solve the initial-value problem and backpropagate through rollouts."},{"cited_title":"Soft- DTW : a differentiable loss function for time-series","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable soft-DTW loss used to compare rollouts with demonstrations of different lengths."},{"cited_title":"Learning stable nonlinear dynamical systems with G aussian mixture models","cited_arxiv_id":null,"evidence_quote":"Provides the LASA handwriting dataset and the stable-dynamical-system baseline context for empirical evaluation."},{"cited_title":"What matters in learning from offline human demonstrations for robot manipulation","cited_arxiv_id":null,"evidence_quote":"Provides the Robomimic manipulation demonstrations used to evaluate high-dimensional performance."}],"review_version":1}