{"id":"cfb76de7-4fbc-4b1e-aaf5-da2c20a54836","arxiv_id":"2602.23050","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A constrained-optimisation objective plus an extended-Kalman VAE architecture improves system identification and prediction accuracy in deep state-space models.","lead":"This paper argues that training deep state-space models by maximizing the usual ELBO objective can fail to learn the true dynamics, and introduces a constrained-optimization training scheme plus a new Kalman-filter-based architecture (EKVAE) to fix it. On pendulum and reacher control tasks, the new model predicts test sequences more accurately than RNN-based baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CO's benefit depends on a feasible reconstruction constraint D0; App. A.7.1 shows the same framework collapses when D0 is only 15% tighter, making the headline improvement conditional on a fragile hyperparameter.","rationale":"The reader's weakest_assumption—feasibility of D0—is also the most load-bearing concern I can identify. The paper's central claim is that CO is a general improvement over plain ELBO/annealing training for DSSMs, and the mechanism by which it helps is the reconstruction constraint. If that constraint is infeasible, the Lagrangian saddle point does not exist and the method collapses; the paper's own Fig. 9 demonstrates this collapse at 1.15D0*. This is not a minor implementational detail: it means the reported successes are conditional on a hyperparameter that must be found empirically, with no theoretical or algorithmic guarantee. The reader's CONDITIONAL verdict already captures this. I considered whether a different concern was more fundamental, such as the EKVAE's 'neural linearisation' not being the true Jacobian of its own transition function, or the lack of released code/error bars. The neural-linearisation issue is real but less load-bearing because EKVAE's recognition model is still a valid variational distribution even if it is not the exact EKF posterior; the ELBO remains a bound. The missing code/error bars are a reproducibility weakness, not a direct attack on the argument. The D0 feasibility issue, by contrast, is explicitly acknowledged in the manuscript and directly threatens the general applicability of the proposed framework. A concrete reproduction of the D0 grid, especially on the reacher image benchmark not covered by Fig. 9, would settle whether the framework is robust. If a range of D0 values works, the central claim stands; if only a knife-edge choice works, the verdict should move toward REJECT. With current evidence, CONDITIONAL remains the appropriate verdict, so no change is recommended.","tokens_in":23333,"tokens_out":10382,"duration_ms":112635,"concrete_test":"Run the full CO training for VHP-EKVAE on the reacher RGB-image benchmark with D0 ∈ {0.8, 0.9, 1.0, 1.1, 1.15, 1.2} × Dmax, using at least 25 seeds per setting. Record (i) whether the constraint D ≤ D0 is actually satisfied at convergence, (ii) the system-identification R2 for angular velocity, and (iii) prediction MSE. If no D0 in this range is both feasible and clearly better than the annealing baseline—or if the feasible interval is a single isolated point that shifts across seeds—then the claim that CO significantly improves system identification is not robust. This directly tests the feasibility assumption that the paper itself identifies as fragile in Fig. 9.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that constrained optimisation (CO) significantly improves system identification and prediction accuracy for DSSMs—requires the constraint D(θ,φ) ≤ D0 to be feasible for the chosen D0. The paper sets D0 by the heuristic D0 = 0.9 × Dmax, where Dmax is the distortion achieved by ordinary ELBO training (App. A.1). This is not a guarantee; it is an empirical choice. Critically, App. A.7.1 and Fig. 9 show that with D0 = 1.15D0* the constraint cannot be satisfied and training collapses. This is a self-admitted failure mode of the same CO mechanism that produces the headline gains. If no feasible D0 exists—or if the feasible region is so narrow that it cannot be identified without expensive search—then the method does not provide the claimed improvement, and the comparison to annealing baselines in Table 1 is not a robust demonstration of a general framework. The paper provides no theoretical condition for feasibility, only a heuristic that happens to work on the tested benchmarks. The main tables also report a single D0 per model/dataset with no error bars, and no code/data are released, so it is possible that the reported R2 and MSE gains are tied to a carefully chosen D0 rather than to the CO formulation itself. This is the most load-bearing assumption because it is both necessary for the method's success and explicitly shown to fail nearby.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the training of deep state-space models (DSSMs), arguing that maximizing the sequential ELBO does not guarantee correct identification of the underlying dynamics. It proposes a constrained-optimization (CO) framework in which the rate is minimized subject to a reconstruction constraint D(θ,φ)≤D0, with a Lagrange multiplier updated by an extension of REWO, plus a variational hierarchical prior (VHP). It then introduces EKVAE, an RNN-free DSSM that combines amortized inference over auxiliary variables with extended Kalman filtering/smoothing and a neural linearization of the transition model. Experiments on pendulum and reacher (angle and RGB image observations) report that CO improves R2-based system identification and prediction MSE for DKS, DVBS, and EKVAE, and that EKVAE achieves the best prediction accuracy, with a demonstration of disentangled position–velocity representations for reward-free policy learning.","tokens_in":23756,"tokens_out":6251,"duration_ms":58755,"significance":"If the empirical claims hold, the paper makes a useful conceptual point: rate–distortion balance should be controlled explicitly in DSSM training, and the proposed CO formulation is a principled way to do this. The derivations in Sec. 3 and App. A are detailed, and the paper extends the method to two established DSSM families (DKS, DVBS) as well as to its own EKVAE. The EKVAE architecture is a clean alternative to RNN-based DSSMs, and the disentangled representation result has practical value for control. The paper also provides some robustness evidence (25-run statistics in Fig. 10) and clearly reports failure modes in App. A.7.1. A further strength is that the main prediction metric is held-out test MSE against ground-truth states, not a term in the training objective, so the evaluation is not circular. However, the central result is conditioned on a D0 heuristic whose sensitivity is not fully characterized, and the headline table lacks error bars; these are the main issues to resolve.","major_comments":[{"comment":"The main claim that CO improves system identification depends on the reconstruction constraint D(θ,φ)≤D0 being feasible. Appendix A.7.1 reports a sharp performance drop at D0=1.15D0*, and states that the constraint cannot be satisfied there. This is load-bearing, but the explanation is unclear: with D0*=0.9Dmax from App. A.1, D0=1.15D0*=1.035Dmax is actually looser than the distortion already achieved by ordinary ELBO training, so one would expect feasibility. Please clarify the normalization of the D0 sweep, provide the full sensitivity curve, and discuss whether a feasible D0 can be identified without expensive search for new datasets. As written, the paper demonstrates the heuristic works on the tested benchmarks but does not support the stronger claim of a general framework.","section":"App. A.7.1 / Fig. 9; Sec. 3"},{"comment":"The headline quantitative results in Table 1 are single-point estimates with no error bars or significance tests. The 25-run statistics in Fig. 10 cover only one model/dataset combination (DKS on pendulum). Since the paper uses the language 'significantly improves' and 'outperforms', the main comparisons (especially EKVAE CO vs. annealing on all three datasets, and the smoothing vs. filtering differences in Table 2) should be reported with means, standard deviations, and ideally significance tests, or the corresponding code should be made available so the claims can be reproduced.","section":"Table 1; App. A.7.2"},{"comment":"The D0 heuristic D0=0.9Dmax uses the reconstruction distortion of the same model under ordinary ELBO training. In settings where ordinary ELBO training fails to identify the dynamics (as shown for annealing in Table 1), Dmax may correspond to a poor local optimum. The paper does not investigate whether the reported improvements are robust to the choice of the reference run used to compute Dmax, nor does it provide guidance for choosing D0 when no reliable ELBO-trained model is available. This should be addressed explicitly, either by an adaptive D0 scheme or by a sensitivity study over Dmax estimates.","section":"Sec. 6.2 / App. A.7.1"}],"minor_comments":[{"comment":"The phrase 'as we show' in the abstract could be softened or the key evidence cited, since the point is only demonstrated empirically in Sec. 6.2.","section":"Abstract / Sec. 1"},{"comment":"The figure caption uses D*_0 without defining it in the caption; please define it as the heuristic value D0=0.9Dmax for readers who do not read App. A.1.","section":"Fig. 9 / App. A.7.1"},{"comment":"H=(δij)∈R^{Da×Dz} is described as 'rectangular identity matrix' but the text should state explicitly that this assumes Da≤Dz; otherwise the first Da dimensions of zt are not well-defined.","section":"Sec. 4.2, Eq. (14)"},{"comment":"No code or data release is mentioned. For a paper whose central contributions are empirical, providing code would greatly strengthen reproducibility; at minimum, a statement of availability should be added.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely a NeurIPS-style submission; the CO extension is conceptually close to [16], and EKVAE builds on KVAE [8], so the novelty is incremental but sufficient if the empirical claims are robust. The main risk is the D0 feasibility sensitivity and lack of error bars in the main table; I would encourage the editor to request a revision that addresses these before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does two useful things. First, it demonstrates a known failure mode—ELBO maximization can produce high likelihood models that haven't actually learned the system dynamics—and shows that a constrained optimization (CO) wrapper, borrowed from the authors' VAE work, consistently fixes this across DKS, DVBS, and their own EKVAE. Second, it introduces the EKVAE, which replaces RNN-based recognition with extended Kalman smoothing plus a learned locally-linear transition: the 'neural linearisation' with base matrices is a neat trick, and the time-invariant auxiliary map H for disentangling position and velocity is clever. The empirical gains on pendulum and reacher are real, as far as I can tell, and the paper is refreshingly honest about where CO helps and where it doesn't. The soft spots are in proportion. The biggest is D0: the paper sets it as 0.9 times the distortion from ordinary ELBO training, and then shows in App. A.7.1 that tightening it by only 15% makes the constraint infeasible and training collapses. That means the headline gains sit on a heuristic that the paper itself demonstrates is fragile. There's no theoretical condition for feasibility, so the 'general framework' claim is weaker than the empirical results. This is a moderate flaw, not a fatal one—the heuristic works on the tested benchmarks, and the paper flags the failure mode openly. Second, the main results table has no error bars. The appendix does include a 25-run comparison for one annealing-vs-CO case, which helps, but the headline numbers in Table 1 are single runs. Without code or data, I can't check the exact values, and the EKVAE rate derivation has a Monte Carlo marginalization step that deserves a careful referee's eye. The tied smoother used as the recognition model also could use tighter analysis. Who's this for? Anyone working on deep state-space models or model-based RL from pixels. It's not a breakthrough, but it's a solid contribution that pushes the subfield forward. I'd take the CO mechanism seriously, and the EKVAE is worth building on. My recommendation: send it to peer review. Ask the authors for error bars, code/data release, and a more careful treatment of D0 sensitivity. The central claim about ELBO decoupling from dynamics holds up well enough to justify referee time.","headline":"A credible, incremental paper that shows constrained optimization improves DSSM dynamics learning, with a genuinely interesting RNN-free EKVAE; the main caveat is the fragile D0 hyperparameter and missing error bars/code.","tokens_in":676,"tokens_out":703,"would_cite":true,"duration_ms":22991,"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":"Maximising the evidence lower bound does not ensure a deep state-space model learns the true dynamics; a constrained-optimisation reformulation, with a learned prior and Kalman-based inference, recovers them and improves prediction.","keywords":["deep state-space models","system identification","evidence lower bound","constrained optimisation","rate-distortion tradeoff","extended Kalman filter","amortised variational inference","disentangled representation"],"falsifier":"On a synthetic linear-Gaussian dynamical system with known transition matrices, train the identical architecture under the CO objective and under plain ELBO maximisation at matched reconstruction quality. If the CO-trained model does not recover the true transition matrices more accurately and does not deliver lower multi-step prediction error, the paper's central claim of improved system identification is falsified. A cheaper partial check, already implicit in the paper's Fig. 9, is that pushing D0 outside the feasible range collapses the method — delimiting the claim's scope.","tokens_in":23225,"feed_emoji":"🎯","tokens_out":31065,"duration_ms":239741,"temperature":0.7,"pith_summary":"This paper tries to establish a negative result and a positive fix: training deep state-space models (DSSMs) by maximising the evidence lower bound (ELBO) can produce models whose latent states do not track the true dynamics of the observed system, because the same ELBO value is compatible with many different balances between reconstruction quality and compression. The proposed fix is a constrained-optimisation framework that rewrites the sequential ELBO as a rate–distortion Lagrangian and enforces a reconstruction threshold, so the model must learn to reconstruct observations well before spending representational capacity on dynamics; a learned hierarchical prior for the initial state avoids the over-regularisation and broken generative models that plague a fixed Gaussian prior. On this base the paper builds the extended Kalman VAE (EKVAE), which replaces RNN-based recognition with analytic extended Kalman filtering/smoothing, a network-predicted linearisation of the transition map, and a linear auxiliary observation model. If the paper is right, DSSM training should be formulated as constrained optimisation rather than plain ELBO maximisation, RNN-free Kalman inference becomes a viable route to system identification — including from raw RGB images — and the learned position–velocity-disentangled representations support reward-free model-based reinforcement learning.","feed_headline":"Recover true dynamics by constraining, not merely maximising, the ELBO","feed_subtitle":"Reformulating deep state-space training as constrained optimisation recovers true states and predicts more accurately.","key_machinery":"The load-bearing device is the rate–distortion decomposition of the sequential ELBO, F = −D − R, turned into the Lagrangian L = R + λ(D − D0), with D the reconstruction distortion and R the KL rate. A sign-aware update rule for λ enforces D ≤ D0 first, then optimises the rate — removing the degeneracy where many (D,R) pairs give the same ELBO. Two mechanisms carry the EKVAE: neural linearisation — the transition p(zt+1|zt,ut) = N(Fψ zt + Bψ ut, Qψ) built from a softmax-weighted mixture of learned base matrices, acting as a state-dependent Jacobian in place of a Taylor expansion — and a time-invariant linear auxiliary model pψ(at|zt) = N(at|Hzt, R) whose rectangular-identity H pins the first","core_discovery":"High ELBO values do not imply a DSSM has learned the true dynamics: annealed baselines match constrained models' ELBO while inferred latents barely correlate with ground truth (R² ≈ 0.04 for pendulum velocity). The remedy is the Lagrangian L = R + λ(D − D0), whose λ-update enforces a reconstruction constraint before dynamics learning; applied to deep Kalman smoothers, deep variational Bayes smoothers, and the new EKVAE, it yields high state correlation and large prediction gains. The EKVAE — extended Kalman smoothing with network-predicted Jacobians and a linear auxiliary observation model — gives the best prediction accuracy in the comparisons, identifies the reacher dynamics from 64×64 RGB","pith_inferences":["Because the rate–distortion degeneracy is generic to ELBO-style objectives, this warning should generalise beyond the tested benchmarks: in any sequential latent-variable setting where the latent state is meant to be physically meaningful, checking inferred latents against known state variables should become a routine diagnostic.","The same Lagrangian machinery could be run in the mirror direction — constraining rate rather than distortion — to enforce a target level of compression or invariance; the paper does not explore this variant.","Neural linearisation — predicting the Jacobian of the transition as a network output rather than differentiating through it — is a transferable design idea for any filtering/smoothing pipeline whose transition function is expensive, implicit, or non-smooth.","The learned hierarchical prior adapts to the dataset's distribution of initial states, so the benefit should scale with how poorly a fixed Gaussian covers those states; datasets with dense initial-state coverage should need the prior less — a testable prediction."],"forward_implications":["ELBO maximisation alone is an unreliable training signal for system identification: models with similar test ELBO can differ wildly in how well their latent states match the true system states (R² from 0.99 down to ~0.04 across configurations in Table 1).","Applying the CO framework to established smoothers (deep Kalman smoothers and deep variational Bayes smoothers) raises their prediction accuracy substantially and lets them identify dynamics that annealing schedules fail to recover.","RNN-based transition models (as in the Kalman VAE and recurrent state-space models) create a non-Markovian latent space: velocity information lives in the recurrent hidden state, so predictions conditioned on a smoothed initial state are markedly worse than those conditioned on a later filtered state.","The EKVAE outperforms the compared DSSMs in prediction accuracy and can identify the reacher system directly from 64×64 RGB images, whereas the RNN-based baselines in the paper cannot.","Disentangled position–velocity representations allow reward functions to be defined purely by encoding goal positions or velocities in latent space, which suffices to learn swing-up, sustained rotation, and reach-avoid policies on the tested control tasks."],"fun_headline_variants":["ELBO alone misleads: constrained training recovers true dynamics","High ELBO ≠ true dynamics: constrained DSSMs learn reality","Constrain the ELBO to fix deep state-space learning","Don't trust ELBO: constrained training recovers true states"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"Everything rests on the reconstruction threshold D0 being feasible: it is set by the heuristic D0 = 0.9 × the best distortion the same model attains under ordinary ELBO training, the appendix shows an abrupt performance collapse at 1.15×D0* attributed to an unsatisfiable constraint, and if no feasible D0 can be found the constrained framework does not help.","fun_headline_variants_meta":{"raw":{"variants":["ELBO alone misleads: constrained training recovers true dynamics","High ELBO ≠ true dynamics: constrained DSSMs learn reality","Constrain the ELBO to fix deep state-space learning","Don't trust ELBO: constrained training recovers true states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000665,"raw_usage":{"total_tokens":2843,"prompt_tokens":685,"completion_tokens":2158,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":429,"completion_tokens_details":{"reasoning_tokens":2087}},"tokens_in":429,"tokens_out":2158,"duration_ms":16964,"temperature":1.0,"reasoning_tokens":2087,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:29:10.663371+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic linear-Gaussian dynamical system with known transition matrices, train the identical architecture under the CO objective and under plain ELBO maximisation at matched reconstruction quality. If the CO-trained model does not recover the true transition matrices more accurately and does not deliver lower multi-step prediction error, the paper's central claim of improved system identification is falsified. A cheaper partial check, already implicit in the paper's Fig. 9, is that pushing D0 outside the feasible range collapses the method — delimiting the claim's scope.","supporting_citations":[],"review_version":1}