{"id":"d30ef9ab-985e-499d-9100-06b04595cc84","arxiv_id":"2501.01339","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper sketches NFPF, a normalizing-flow particle filter with jointly learned linear latent dynamics, but provides only qualitative and self-admittedly insufficient CartPole experiments.","lead":"The report proposes NFPF, a method that combines a normalizing flow observation model, a learned linear latent dynamics model, and a particle filter to estimate hidden states from images. A generalist should read it as an early architecture sketch, because the author explicitly says the experiments are too limited to confirm the method works.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training objective Eq. (9)–(11) is not computable from images alone: latent states X are never inferred during training, and Eq. (11) also decouples θ from φ,ψ even though ŷ_t = gθ^{-1}(y_t) keeps them coupled.","rationale":"The reader's weakest assumption—that latent states X are treated as known deterministic nodes although they are never observed or inferred—is precisely the load-bearing gap in the central claim. My additional point about Eq. (11) is not a substitute for that gap but a further internal inconsistency: even if X were available, the stated separation of the minimization into independent θ and φ,ψ problems is false because ŷ_t depends on θ. Both issues are concrete and reproducible from the equations as printed, and both undermine the strongest claim that the method 'allows joint training' from image observations. The paper is transparent about its experimental weakness and does include the author's own limitation statements, which counts in its favor as honesty, but it does not change the fact that no valid image-only training algorithm is actually specified. I agree with the REJECT verdict: the architecture sketch may be a plausible direction, but as written the central claim is unsupported, and no amount of qualitative plotting can repair the missing inference and the algebraic error in Eq. (11).","tokens_in":5840,"tokens_out":4063,"duration_ms":44129,"concrete_test":"Take a single training image y_t from the CartPole dataset and write out all quantities needed to evaluate Eq. (10): x_t or x_{t-1} is required, but neither is defined from y_t. Concretely, attempt to implement the training loop exactly as described (no extra encoder, no approximate posterior) and compute a numerical value of Eq. (11) for a minibatch of 10 trajectories. If the only way to get a finite value is to supply latent states from the simulator or to add an inference network absent from the paper, then the claimed image-only joint training procedure does not exist. As a separate check, re-derive Eq. (11) while keeping θ in the first term and verify whether ∂/∂θ log p(ŷ_t | μϕ(Aψ x_{t-1} + Bψ u_t), Σ) is nonzero; if it is nonzero, the decoupled minimization stated in Eq. (11) is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that NFPF can be trained on image observations only. Eq. (9) defines the objective as log p(Y|X), where X is a sequence of latent states, but no inference procedure is specified: there is no encoder, no variational posterior, no importance-sampling distribution, and no initial-state distribution p(x_k) to integrate over. Eq. (10) then replaces x_t with the deterministic recursion Aψ x_{t-1} + Bψ u_t, which does not depend on the observed y_t; with x_{t-1} itself unobserved, the right-hand side is undefined for a training batch of images. Thus Eq. (11) is not actually computable. A second independent defect compounds this: the minimization is split as min_{φ,ψ} over the first term plus min_θ over the second term, but the first term contains ŷ_t = gθ^{-1}(y_t), so it depends on θ; the claimed decoupling is algebraically false. The reported experiments (2 particles, 10 training trajectories, no baselines, no quantitative comparison) cannot compensate for the missing training procedure, because the plots in Figs. 4 and 5 are produced by a particle filter whose proposal is a standard normal latent state, not by a model whose latent states were inferred from the training images. The Discussion's own admission that the plots 'might not convince the reader' and the listed limitations do not repair this: they confirm that the central claim is not supported by the supplied evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method called Normalizing Flows based Particle Filter (NFPF) for simultaneous latent state estimation and latent linear dynamics discovery from image observations. It reviews Kalman filtering, particle filtering, autoencoders, SINDy, and normalizing flows, then derives a training objective in Section 3.1 that is intended to jointly train a normalizing flow observation likelihood, a mean parametrization, and linear dynamics matrices. The experiments apply the method to CartPole image observations and show latent trajectories from a bootstrap particle filter with two particles, without quantitative evaluation or comparisons.","tokens_in":6158,"tokens_out":5428,"duration_ms":55246,"significance":"If the proposed training procedure were well defined, the idea of using a normalizing flow to supply an observation likelihood to a particle filter while jointly learning latent linear dynamics would address a real and useful problem. The paper also offers a readable survey of the relevant components and is unusually candid about its limitations. However, the central algorithmic claim is not supported: the training objective is written over unobserved latent states with no specified inference procedure, and the experiments do not provide quantitative evidence. The significance is therefore not established in the current form.","major_comments":[{"comment":"The training objective in Eq. (9) is written as maximizing log p(Y|X) over a sequence of latent states X, but X is never observed and the paper specifies no encoder, variational posterior, importance-sampling distribution, or initial-state distribution p(x_k) to make this objective computable. Replacing x_t by A_psi x_{t-1} + B_psi u_{t-1} in Eq. (10) does not resolve the problem, because x_{t-1} is itself unobserved and no inference step is defined. Consequently, Eq. (11) is not actually computable from image observations alone, and the claimed joint training of g_theta, mu_phi, and the linear matrices A_t and B_t is not supported. This is the central gap in the paper.","section":"Section 3.1, Eqs. (9)-(11)"},{"comment":"Even setting aside the missing inference procedure, the split of the minimization into separate terms over (phi, psi) and theta is algebraically incorrect. The first term contains y_hat_t = g_theta^{-1}(y_t), so it depends on theta through the inverse of the normalizing flow. The objective therefore cannot be minimized for phi and psi independently of theta as written; the claimed decoupling in Eq. (11) is false.","section":"Section 3.1, Eq. (11)"},{"comment":"The experimental evidence is insufficient to support the central claims. The evaluation uses only 10 training trajectories and a particle filter with 2 particles, reports no quantitative error metrics, no baselines, and no train/test split, and the plots are presented for visual inspection only. The Discussion itself lists these missing elements, including controllability restrictions, LQR evaluation, and better proposal distributions. Without a well-defined training procedure and quantitative validation, the empirical claims for latent state estimation and dynamics discovery are not established.","section":"Section 4, Figures 4 and 5"}],"minor_comments":[{"comment":"The integral in Eq. (4) writes dx_{k-1} in the text; this should be dx_{t-1}.","section":"Section 2.1.1, Eq. (4)"},{"comment":"The statement that the Bayesian filter integrals 'can be computed analytically only when g(x_t) is a linear function' is imprecise, since the EKF is then described as a way to compute them; the EKF linearizes the observation function and yields approximate results, not analytical ones.","section":"Section 2.1.1"},{"comment":"Figure 5 is captioned as showing 100 dimensions while Section 4 states that the latent dimensionality was set to 4; the relation between these two figures should be clarified.","section":"Section 4, Figures 4 and 5"},{"comment":"The initial latent state distribution p(x_k) and the covariance Sigma appear in the architecture figure but are not defined in the training objective; the paper should specify how these quantities are chosen or learned.","section":"Section 3.1 and Figure 2"},{"comment":"The Methods section motivates comparison with SINDy, but no comparative experiments are reported; the paper should either include such a comparison or state that it is left to future work.","section":"Section 2.3 and Section 4"},{"comment":"The text has a report-style tone, e.g., 'This report tries to conclude previous works'; the abstract and introduction should be revised to a journal-paper format.","section":"Abstract and Introduction"}],"recommendation":"reject","confidential_remarks":"The manuscript reads as an early technical report rather than a finished research paper. The central missing piece is a well-defined inference procedure for the latent states during training; without that, Eqs. (9)-(11) do not define a trainable algorithm. Reconstructing such a procedure and adding proper quantitative evaluation would constitute substantial new work rather than a local revision, so I recommend rejection. The literature review and the honest discussion of limitations are useful, but they do not compensate for the absent algorithmic core and validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is an honest but unfinished architecture sketch. The contribution is a legitimate combination of existing pieces - conditional normalizing flow observation likelihood, jointly learned linear latent dynamics, and a bootstrap particle filter - and the paper is transparent that it is a report rather than a validated method. The literature review is fine, and the limitations list in Section 5 is unusually candid. I give it credit for that.\n\nThe soft spots are load-bearing. Equation (9) defines the objective as log p(Y|X) over latent states X, but no inference procedure is specified: no encoder, no variational posterior, no proposal distribution, no initial-state distribution. Equation (10) then replaces x_t with the deterministic recursion A_psi x_{t-1} + B_psi u, but x_{t-1} is unobserved, so the right-hand side is not computable from image observations alone. The split in Equation (11) is also algebraically false: the first term contains y_hat_t = g_theta^{-1}(y_t), so it depends on theta; you cannot minimize over phi, psi and theta separately. The stress-test note is correct on both points.\n\nThe experiments do not fill the gap. Ten training trajectories, two particles, no baselines, no quantitative metrics, no error bars, and no train/test split. The plots in Figures 4 and 5 come from a particle filter with a standard normal proposal, so they do not demonstrate that the learned observation likelihood is actually good. The author's admission in the Discussion that the plots might not convince the reader is accurate, not merely modest.\n\nThere is a kernel of an idea here - NF gives exact density evaluation that SINDy lacks, and tying the flow to latent linear dynamics is plausible - but as written the central claim is unsupported. This is not a paper I would send to a serious referee in its current form. The right move is to desk reject and encourage the author to resubmit after specifying an actual inference procedure for X, fixing the decoupling error, and providing quantitative comparisons with baselines. Until then, the method section does not hold together on its own terms.","headline":"Honest but unfinished architecture sketch: the central training objective is not computable as written, and the experiments are far too weak to support the claims.","tokens_in":593,"tokens_out":962,"would_cite":false,"duration_ms":30897,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A conditional normalizing flow supplies the observation likelihood so that a particle filter can jointly estimate latent states and latent linear dynamics from images alone.","keywords":["latent state estimation","normalizing flows","particle filter","linear dynamics discovery","image observations","posterior filtering distribution","joint training","hidden Markov model"],"falsifier":"On a synthetic image-sequence benchmark with known ground-truth latent states (for example, a pendulum rendered from its true angle and angular velocity), run the proposed joint training from several random initializations. If the learned latent trajectories do not track the true states and the learned matrix A does not predict the true next latent state better than a random linear model, then the claim that images alone yield latent state and linear dynamics is not supported.","tokens_in":5579,"feed_emoji":"🌀","tokens_out":7335,"duration_ms":68143,"temperature":0.7,"pith_summary":"This paper proposes a filtering method, NFPF, that treats image observations as outputs of a latent linear dynamical system. The method uses a conditional normalizing flow to compute the observation likelihood p(yt|xt), and trains the flow, the mean parametrization of its base distribution, and the linear dynamics matrices At and Bt jointly by maximizing the log-likelihood of the observed images. If this works, an agent could recover both a posterior distribution over latent states and the linear dynamics that generated them from raw images alone, with no manually specified observation model. The report includes a CartPole demonstration in which a bootstrap particle filter uses the learned likelihood to produce latent trajectories.","feed_headline":"A flow-based particle filter learns hidden states and dynamics from images","feed_subtitle":"The flow supplies the observation likelihood; the particle filter recovers latent states and linear dynamics.","key_machinery":"The central object is the conditional normalizing flow used as an observation model: an invertible network gθ that maps the latent state x and the image y to a simple Gaussian variable ŷ, so that the observation density is evaluated exactly by the change-of-variables formula p(yt|xt) = p(ŷt|µϕ(xt), Σ)|det J_{$gθ^{{-1}}$}(yt)|. This identity is what makes the whole pipeline differentiable: gradients from the image likelihood flow back through the flow, through the mean parametrization, and through the linear dynamics matrices, enabling joint training of all parameters. In deployment, the resulting likelihood is plugged into the bootstrap particle filter weight update to obtain an approximate posterior filtering distribution.","core_discovery":"The central claim is that the observation likelihood p(yt|xt) of a hidden Markov model with latent linear dynamics can be learned from images by a conditional normalizing flow gθ whose base distribution is Gaussian with mean µϕ(xt) produced by a convolutional network. The dynamics are parametrized as xt = Aψ xt−1 + Bψ ut−1, where fψ outputs the matrices from the current latent state (and optionally the control). Maximizing log p(Y|X) over a batch, expanded via the change-of-variables formula, updates θ, ϕ, and ψ in one differentiable objective; the same likelihood then feeds the particle-filter weight update. The paper thus claims simultaneous latent state estimation and latent linear dynamics discovery from image observations alone, with stability constraints such as spectral radius less than one available for the learned matrices.","pith_inferences":["The training objective in Eq. (9)-(11) treats the latent states xt as known deterministic nodes, yet the paper specifies no encoder or inference step for obtaining them from images during training; a variational or particle-based inference layer would make the objective well-posed and would be a natural completion of the method.","The paper's CartPole plots show two-particle filter trajectories against true latent states, but the report itself notes the evidence is thin; a stronger test would be a synthetic-image benchmark with known ground-truth latent states and quantitative trajectory error.","If the conditional flow were replaced by a dimension-changing flow, the method could drop the mean-parametrization CNN entirely and learn the observation model as a fully invertible encoder-decoder.","Comparing NFPF directly with SINDy-based autoencoders on the same images would clarify whether exact observation likelihoods change the discovered dynamics or only the posterior uncertainty estimates."],"forward_implications":["From image observations alone, the method yields both an approximate posterior filtering distribution p(xt|y1:t) and time-dependent linear dynamics matrices At and Bt.","The learned observation likelihood can be inserted into any particle-filter weight update, not only the bootstrap variant used in the experiments.","Because the dynamics matrices are produced by a network fψ from the latent state, the formulation extends to state-dependent and control-dependent linear dynamics.","The paper identifies latent-space LQR control and the inclusion of controllability constraints as natural next steps if the learned matrices are to be used for control."],"supporting_citations":[{"why":"Supplies the conditional normalizing flow used to model the observation likelihood p(y|x) with an exact density.","marker":"[16]"},{"why":"Provides the particle filter weight update and Bayesian filtering background that the proposed algorithm relies on.","marker":"[15]"},{"why":"Gives the change-of-variables density formula that makes the flow-based likelihood computable.","marker":"[13]"},{"why":"Demonstrates particle filters using conditional normalizing flows, the line of work the paper extends to observation likelihoods.","marker":"[6]"},{"why":"Provides the SINDy autoencoder baseline that also learns latent dynamics from observations and that the method contrasts with.","marker":"[5]"}],"fun_headline_variants":["Flow-based particle filter uncovers latent linear dynamics and states from images","Image-driven flow particle filter reveals hidden states and linear dynamics","Learn latent state and dynamics models from images via a flow particle filter","Particle filter with normalizing flow estimates latent states and discovers linear dynamic","Simultaneous latent state and linear dynamics recovery from image observations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training objective maximizes log p(Y|X) after substituting the latent state xt by the deterministic dynamics recurrence, but the paper never explains how latent states are obtained from image observations during training, so the objective may not be computable for actual image data.","fun_headline_variants_meta":{"raw":{"variants":["Flow-based particle filter uncovers latent linear dynamics and states from images","Image-driven flow particle filter reveals hidden states and linear dynamics","Learn latent state and dynamics models from images via a flow particle filter","Particle filter with normalizing flow estimates latent states and discovers linear dynamics","Simultaneous latent state and linear dynamics recovery from image observations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000271,"raw_usage":{"total_tokens":1526,"prompt_tokens":737,"completion_tokens":789,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":353,"completion_tokens_details":{"reasoning_tokens":700}},"tokens_in":353,"tokens_out":789,"duration_ms":8127,"temperature":1.0,"reasoning_tokens":700,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:29:21.749528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic image-sequence benchmark with known ground-truth latent states (for example, a pendulum rendered from its true angle and angular velocity), run the proposed joint training from several random initializations. If the learned latent trajectories do not track the true states and the learned matrix A does not predict the true next latent state better than a random linear model, then the claim that images alone yield latent state and linear dynamics is not supported.","supporting_citations":[{"cited_title":"Cambridge University Press, 2013","cited_arxiv_id":null,"evidence_quote":"Provides the particle filter weight update and Bayesian filtering background that the proposed algorithm relies on."},{"cited_title":"Differentiable Particle Filters through Conditional Normalizing Flow","cited_arxiv_id":"2107.00488","evidence_quote":"Demonstrates particle filters using conditional normalizing flows, the line of work the paper extends to observation likelihoods."}],"review_version":1}