{"id":"5db263bd-b2d0-4b6e-b861-b849fb5f2fdc","arxiv_id":"2508.06742","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Sampling causal structure hypotheses from a feature-attribution-derived distribution, instead of committing to a single causal graph, makes learned robot dynamics models more robust to noise and change at a fraction of the compute.","lead":"This paper introduces CADY, a robot dynamics model that learns a distribution over possible causal relationships between a robot's state, its commands, and the next state, and uses random samples of those relationships to shape its predictions. Tests on simulated and real robots show the model stays accurate under noisy or missing sensors and sudden dynamics changes, while needing far less compute than baseline model-based planners.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"IG-derived causal structure may be a placebo: no control shows that the Bernoulli mask probabilities p carry causal information beyond generic stochastic masking.","rationale":"The reader's weakest assumption correctly identifies the IG-to-causality gap as the central risk. I push it one step further: the stronger concern is that the method could be completely insensitive to the values of p, because no control with uninformative p is reported. The paper's empirical results are genuine and the computational savings are concrete, but they do not yet establish that causal information, rather than stochastic mask diversity, drives the robustness improvements. This is an empirical question that the same experimental setup can answer with three additional mask distributions. Since the current evidence supports a conditional verdict and the concern is addressable, I keep the reader's CONDITIONAL disposition unchanged rather than escalating to REJECT.","tokens_in":14911,"tokens_out":3870,"duration_ms":47843,"concrete_test":"Re-run the Cartpole/Pusher frozen- and noisy-sensor protocols (Section V-D) and the Jackal intervention protocol (Section V-E) with three additional mask distributions: (i) uniform Bernoulli p=0.5, (ii) a random permutation of the learned p matrix, and (iii) the known ground-truth p for Cartpole. Compare degradation percentages and post-intervention MSE. If uniform or shuffled p performs within noise of the learned-p CADY, the causal content of p is not load-bearing and the central claim fails; if uniform p degrades toward PETS-level performance and ground-truth p matches/exceeds learned p, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that sampling masks from an estimated causal structure distribution improves robustness. Those mask probabilities are estimated with Integrated Gradients on f_C trained on the same dataset D (Eqs. 3-4), and the paper itself states 'IG was not designed as a causality test' (Section IV-A). The only ground-truth check, Fig. 6, visibly misses the true theta-dot -> x-dot edge and is run in the same Cartpole environment used for training. If IG values do not track causal effects, the sampled masks are just stochastic feature masks, and the robustness gains could come from regularization rather than from causal structure. The fixed-vs-distribution ablation (Section V-D) shows only that resampling from the learned p helps over one fixed graph; it does not compare against a uniform p=0.5 or a shuffled p, so it cannot distinguish causal information from generic mask diversity. Thus the abstract's causal robustness claim rests on an unvalidated proxy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CADY, a probabilistic encoder-multidecoder dynamics model that masks its latent features with binary matrices sampled from a learned distribution over causal structures. The structure distribution is estimated by computing Integrated Gradients on a separately trained 'fully wired' network and normalizing the attributions to Bernoulli probabilities (Eqs. 3–4). The authors argue that marginalizing over structural uncertainty yields dynamics models that are more robust to missing/noisy observations and unexpected interventions, while using far fewer parameters and FLOPs than PETS. Experiments cover Cartpole, Pusher, and a Jackal ground vehicle in simulation and the real world, with robustness, intervention, and computational-cost evaluations.","tokens_in":15159,"tokens_out":3817,"duration_ms":45341,"significance":"If the causal-structure claim is substantiated, the paper offers a computationally cheap alternative to causal discovery for improving model-based planning robustness. The empirical evaluation is genuinely broad and includes several well-designed controls: held-out noise and missing-feature tests, a parameter-count control PETS(S) in the intervention study, and a real-world zero-shot deployment. The reported gains are large (e.g., Jackal Mission 1 success 85.7% vs. 55.7% at noise variance 0.01; post-intervention one-step MSE 0.124 vs. 2.039) and the computational savings are substantial. However, the central mechanism—that the Integrated-Gradients-derived mask probabilities carry causal information—is not isolated from generic stochastic masking or feature selection, so the paper's headline causal explanation remains conditional on additional controls.","major_comments":[{"comment":"The causal structure distribution p_ij is defined as normalized Integrated Gradients of a network f_C trained on the same observational dataset D used to train the dynamics model. As the paper itself states, 'IG was not designed as a causality test.' The only ground-truth validation is the Cartpole comparison in Fig. 6, which visibly underestimates the θ̇→ẋ edge and is in the same environment used for training. The central claim that masks sampled from P(p) are 'causally informed' is thus not yet established. A concrete test: compare CADY's performance against the same architecture with (i) uniform p=0.5, (ii) p with shuffled entries, and (iii) IG computed from a model trained on label-randomized data. If these controls perform comparably, the gains are due to stochastic masking/regularization, not causal structure. This is load-bearing for the abstract's causal claim and must be address","section":"Section IV-A, Eqs. (3)–(4)"},{"comment":"The ablation showing that resampling from the learned p improves performance over a fixed graph is presented as evidence for the value of structural uncertainty. However, the fixed graph is presumably derived from p itself, and the comparison conflates the specific learned probabilities with the effect of stochastic mask sampling. Without a uniform-p or shuffled-p control, the result cannot distinguish causal information from generic mask diversity. The paper should report the fixed-graph result alongside the proposed uniform/shuffled controls, and clarify how the 'fixed graph' is selected (threshold, most-likely edge, etc.).","section":"Section V-D, fixed-graph ablation"},{"comment":"The structural validation is limited to a single environment (Cartpole) and even there the estimated distribution misses the true θ̇→ẋ edge with low probability. For Pusher and Jackal no ground-truth structure is provided at all, yet the robustness narrative depends on the masks being causally meaningful in those settings. If the causal claim is retained, the authors should either provide structural validation in at least one more environment or show that the missing edge is not load-bearing for the robustness results. Alternatively, the text should be revised to characterize the masks as 'predictive-feature masks' rather than causal structures, with the causal terminology softened accordingly.","section":"Section V-B, Fig. 6"}],"minor_comments":[{"comment":"The smoothing function s(·) and the clipping bound ρ_min are introduced in Eq. (4) but only defined in Section V-A ('cubic root' and 0.02). Define them at first use for clarity.","section":"Section IV-A, Eq. (4)"},{"comment":"PETS(S) is first mentioned in Fig. 10 but not defined in the baselines list; please define it explicitly (presumably 'PETS with reduced parameters') at first occurrence.","section":"Section V-A, Baselines"},{"comment":"The figure caption says the ground truth is for the state delta, but the text 'the figure does not show the ∆ symbol for ease of visualization' appears only in the body; please state this in the caption to prevent misinterpretation of the x_t → x_{t+1} relation.","section":"Fig. 6"},{"comment":"Reference [25] appears in the bibliography between [24] and [26] but is not in numeric order in the text citation sequence; please reorder or renumber.","section":"References"},{"comment":"The parameter counts for CADY are very small (e.g., 230 for Cartpole). It would help to report the architecture breakdown (encoder/decoder sizes) or at least confirm that these counts include all trainable parameters.","section":"Table I"},{"comment":"Minor typos: 'FLOPS' vs. 'FLOPs' is inconsistent; 'P(pD)' in Section IV-B is used before defining p_D. Please proofread notation consistency.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The empirical results are strong and would likely be of interest to the RA-L readership, but the paper's central causal claim is not yet adequately supported. The requested controls (uniform/shuffled p, randomized-label IG) are feasible within the paper's experimental setup and would directly test whether the causal structure distribution is doing the advertised work. If the authors are unable to distinguish from generic stochastic masking, the paper could be resubmitted with the causal claims appropriately weakened to feature-mask-based robustness, which may still be valuable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: the empirical core is genuinely useful, but the causal framing outruns the evidence. The actual contribution is a cheap encoder-multidecoder dynamics model that masks its latent space using Bernoulli probabilities computed from Integrated Gradients on a fully-wired network trained on the same observational dataset. That pipeline is new in the cited literature, and the robustness evaluations are properly forward-looking: frozen/noisy sensors, sudden gain reversals, and a parameter-matched PETS control in the intervention experiment. The compute savings (20-250x fewer FLOPs/params) and the consistency of the gains across Cartpole, Pusher, and Jackal simulations make the method worth knowing about. I discount the field demo (video only, no numbers).\n\nThe soft spot is exactly what the stress-test note flags: the IG-derived edge probabilities are never shown to carry causal information beyond generic stochastic masking. There's no control with uniform p=0.5 or a shuffled p, so the fixed-vs-distribution ablation only shows that resampling helps, not that the resampling distribution is causal. The paper even admits IG 'was not designed as a causality test,' and the only ground-truth check (Cartpole, Fig. 6) visibly misses the theta-dot -> x-dot edge. So the robustness gains are real in the narrow sense, but I wouldn't trust the causal interpretation until a control distinguishes structure from noise. The authors' \"causally-informed/inspired\" hedge in the introduction is honest; the abstract and title don't carry that hedge.\n\nMinor: time/distance stats condition on success, which favors the reliable method by construction; compute comparison is against PETS only, not against other lightweight models. These are addressable.\n\nBottom line: solid, well-executed empirical paper with an unvalidated mechanism. It deserves a serious referee and could be useful for people building robust model-based planners, but the causal claim needs an ablation that isolates structure before I'd cite it as causal. If I were editing, I'd send it for full review with a request for that control.","headline":"The method is a cheap, genuinely robust latent-masking trick, but the causal mechanism is unproven — the paper deserves review for the empirics, not the causal framing.","tokens_in":15697,"tokens_out":3151,"would_cite":true,"duration_ms":33201,"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 claims that a robot's learned dynamics model is more robust to sensor failure and dynamics changes when it samples causal masks from a distribution over structures rather than relying on a single causal graph.","keywords":["causal structure distribution","integrated gradients","dynamics learning","model-based planning","robustness","encoder-multidecoder","structural causal model","latent masking"],"falsifier":"On a system with a known causal graph, compare CADY's behavior under three mask sources: Integrated-Gradient estimates, the ground-truth graph, and uniform random masks with matched sparsity. If random masks match IG masks in robustness, the causal estimator is not carrying the result; if ground-truth masks do not beat IG masks, the structure distribution itself is not the cause. A second check: rescale one input feature by a constant and see whether p_ij shifts; IG is sensitive to scaling, so a causally meaningful estimator should not arbitrarily flip the masks.","tokens_in":14715,"feed_emoji":"🤖","tokens_out":8106,"duration_ms":81778,"temperature":0.7,"pith_summary":"This paper claims that a robot's learned dynamics model becomes more robust to real-world disruptions if, instead of committing to a single causal graph, it samples masks from a distribution over causal structures. The authors build CADY, an encoder-multidecoder network whose latent features are masked by binary matrices drawn from a Bernoulli distribution; the edge probabilities are estimated with Integrated Gradients on a fully-wired network. Across Cartpole, Pusher, and a simulated and real Jackal ground vehicle, CADY keeps planning accurate when sensors freeze or go noisy, adapts quickly to unmodeled control-gain interventions, and uses 20x-250x fewer FLOPs and parameters than the PETS ensemble baseline. The broader claim is that accounting for structural uncertainty is a cheap way to make model-based planning robust.","feed_headline":"Causal mask sampling makes robot models 20x cheaper and more resilient","feed_subtitle":"Sampling masks from an estimated structure distribution keeps learned models accurate when sensors freeze or go noisy.","key_machinery":"The load-bearing object is the causal structure distribution $P(p)$, a product of independent Bernoulli variables $E_{ij}$ indicating an edge from input $i$ to next-state variable $j$. Its parameters are set to normalized Integrated Gradients, $p_{ij} := \\mathrm{IG}_{ij}(x, x'=0; f_C)$, clipped to $[\\rho_{\\min}, 1-\\rho_{\\min}]$, so feature attribution stands in for causal discovery. Every forward call samples a mask $M \\sim P(p)$ and applies it in the latent space, $\\tilde{z}_j = z \\odot m_{:,j}$, before a per-output decoder produces a Gaussian over the next state. This turns structural uncertainty into a cheap training-time regularizer and an inference-time marginalization.","core_discovery":"The discovery is that uncertainty over causal structure, not just over function parameters, belongs inside a learned dynamics model. CADY represents the dynamics as $f(x) = f_{\\mathrm{dec}}(f_{\\mathrm{enc}}(x) \\odot M)$ with $M \\sim P(p)$, where $P$ is a product of independent Bernoulli edge distributions over the bipartite graphs allowed by the Markovian state-transition assumption. Each $p_{ij}$ is estimated as a normalized, clipped Integrated Gradient of a fully-wired contribution model, and every forward pass draws a fresh causal mask. In experiments this marginalization over plausible structures yields up to 13% lower performance degradation under frozen or noisy sensors, one-step post-","pith_inferences":["If the mechanism is general, the same encoder-multidecoder plus Bernoulli-mask recipe could apply to any controlled dynamical system with sparse interactions, not just robots.","A natural testable extension is to replace the Integrated-Gradients estimator with a proper Bayesian structure-learning method; if the robustness gains survive that swap, the architecture is the contribution and the attribution estimator is incidental.","The authors' own Cartpole sanity check shows one mismatched edge ($\\dot{\\theta}_t \\to \\dot{x}_{t+1}$); intervening on that specific edge would be a sharper test of whether the estimated masks track true causal structure."],"forward_implications":["A dynamics model that samples causal masks at inference time degrades 1.6-13% less than PETS when a sensor freezes or becomes noisy, with the gap growing in higher-dimensional systems.","Under an unmodeled control-gain intervention, CADY's one-step prediction error is 0.124 versus 2.039 for PETS, and fine-tuning on recent data recovers faster.","On the Jackal ground vehicle under GPS-like state noise (variance 0.01), mission success rises to 85.7% from PETS's 55.7%; on a steep mountain mission, CADY keeps 27% success where PETS falls to 2.8%.","Marginalizing over a distribution of structures rather than fixing one graph improves performance by 3.2% on Cartpole and 5.8% on Pusher.","CADY uses 20x fewer FLOPs and parameters than PETS on Pusher and more than 250x fewer on Jackal, including the cost of sampling the mask."],"supporting_citations":[{"why":"Supplies Integrated Gradients, the attribution method used to estimate the Bernoulli edge probabilities p_ij.","marker":"[9]"},{"why":"Defines structural causal models and the observational-equivalence argument for why a single graph is hard to identify.","marker":"[1]"},{"why":"Supports the claim that explicitly modeling causal structure improves generalization and robustness.","marker":"[2]"},{"why":"Provides the combinatorial count of DAGs used to justify bounding the search to bipartite dynamics graphs.","marker":"[5]"},{"why":"PETS is the main model-based baseline whose performance and computational cost CADY is compared against.","marker":"[41]"},{"why":"CMI, a causally-inspired dynamics model using conditional mutual information, is a baseline for causal structure estimation.","marker":"[8]"},{"why":"REG, an L1-regularized causal-state-abstraction baseline, is compared for causal masking and planning.","marker":"[43]"}],"fun_headline_variants":["Sampling causal structures boosts robot model robustness 13%","Causal uncertainty in dynamics models reduces compute and improves resilience","CADY: Sampling causal masks makes robot planning robust to noisy sensors","Learning causal structure distributions for cheaper, sturdier robot models","Causal mask sampling cuts robot model cost 20x and boosts resilience 13%"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that Integrated Gradients on a fully-wired network trained on the same observational data yield valid edge probabilities for a causal structure distribution; if attribution values do not track true causal effects, the masks can drop predictive features or keep spurious ones and the claimed mechanism collapses.","fun_headline_variants_meta":{"raw":{"variants":["Sampling causal structures boosts robot model robustness 13%","Causal uncertainty in dynamics models reduces compute and improves resilience","CADY: Sampling causal masks makes robot planning robust to noisy sensors","Learning causal structure distributions for cheaper, sturdier robot models","Causal mask sampling cuts robot model cost 20x and boosts resilience 13%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3019,"prompt_tokens":763,"completion_tokens":2256,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":2180}},"tokens_in":507,"tokens_out":2256,"duration_ms":18651,"temperature":1.0,"reasoning_tokens":2180,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:36:09.406770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a system with a known causal graph, compare CADY's behavior under three mask sources: Integrated-Gradient estimates, the ground-truth graph, and uniform random masks with matched sparsity. If random masks match IG masks in robustness, the causal estimator is not carrying the result; if ground-truth masks do not beat IG masks, the structure distribution itself is not the cause. A second check: rescale one input feature by a constant and see whether p_ij shifts; IG is sensitive to scaling, so a causally meaningful estimator should not arbitrarily flip the masks.","supporting_citations":[{"cited_title":"Toward causal representation learning","cited_arxiv_id":null,"evidence_quote":"Supports the claim that explicitly modeling causal structure improves generalization and robustness."},{"cited_title":"Counting unlabeled acyclic digraphs","cited_arxiv_id":null,"evidence_quote":"Provides the combinatorial count of DAGs used to justify bounding the search to bipartite dynamics graphs."},{"cited_title":"Deep reinforcement learning in a handful of trials using probabilistic dynamics models","cited_arxiv_id":null,"evidence_quote":"PETS is the main model-based baseline whose performance and computational cost CADY is compared against."},{"cited_title":"Task independent causal state abstraction","cited_arxiv_id":null,"evidence_quote":"REG, an L1-regularized causal-state-abstraction baseline, is compared for causal masking and planning."}],"review_version":1}