{"id":"76bdc422-e67e-4e65-8151-465e0f7ed22f","arxiv_id":"2505.08453","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Systematic analysis of Causal Curiosity in a simulated robotic manipulator shows high accuracy in single-factor and high-granularity settings, but frequent failures when multiple causal factors vary simultaneously.","lead":"This paper stress-tests Causal Curiosity, a reinforcement learning method that estimates hidden physical parameters of objects by choosing robot actions that make object trajectories cluster by parameter value. The authors find the method works well in simple single-factor settings, but often fails when several parameters vary at once, and they propose a modified reward and a PPO planner that sometimes helps.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'impossibility' of multi-factor estimation in §4.4 is not established: the evidence comes only from two open-loop planners with deliberately untuned hyperparameters, so the failures may be planner limitations rather than identifiability limits.","rationale":"The paper's most valuable contribution is its empirical mapping of where Causal Curiosity works, such as single-factor settings and fine granularity, and where it degrades, such as small gaps and simultaneous factor variation. Those measurements are plausible and useful even if the headline is softened. The load-bearing issue is not primarily the lack of error bars or code, both real but secondary, but the inference from 'CEM and an observation-blind PPO variant failed on these combinations with untuned hyperparameters' to 'impossibility of solving some complex experimental situations'. Section 3.6 is decisive here because the authors state that hyperparameters were deliberately not tuned. The PPO planner described in §3.5 is also not a genuinely closed-loop RL agent: wrappers force all environments to execute the same actions and the reward is computed only at the last timestep, so the policy cannot adapt mid-episode. Thus the negative results in Tables 5–8 are properties of a specific search class and budget. The proposed concrete test, a closed-loop policy or heavily tuned CEM on the worst combination, would settle whether the limit is real. Because the reader already conditioned the verdict on exactly this overclaim, I do not move the verdict; I would keep CONDITIONAL, with the condition that 'impossibility' be replaced by 'not achieved by the tested planners' unless the test shows otherwise.","tokens_in":23760,"tokens_out":6988,"duration_ms":76016,"concrete_test":"Re-run the most challenging Section 4.4 combination, for example Spinning Friction as main factor with Size as secondary factor where CEM/PPO report F1 ≈ 0.55, under the same 20 environments and Eq. (6) reward, but with a closed-loop policy that receives per-environment observations during the episode, or with a CEM search using tuned hyperparameters and an order-of-magnitude larger iteration budget. If any tested configuration achieves F1 ≥ 0.9 on held-out trajectories, the reported 'impossibility' is a planner/search limitation rather than an identifiability limit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim, 'we show the impossibility of solving some complex experimental situations... with multiple parameters varying simultaneously', overreaches what Section 4.4 can support. All negative results are produced by exactly two planners: the CEM open-loop planner from [21] and a PPO policy that, as described in §3.5, is modified so that all environments execute the same actions and the reward is injected only at the last timestep. Neither planner can adapt its actions to observations during the episode. Moreover, §3.6 explicitly states that hyperparameters were 'defined without the need for hypertuning processes' because the aim was comparative results, not best possible results. A failure of an untuned, non-adaptive search is not evidence that the causal factor is unidentifiable. The F1/Silhouette reward of Eq. (6) is a clustering proxy: it measures whether one fixed action sequence separates trajectories into two clusters, not whether any admissible experiment, such as an adaptive or multi-stage policy with tuned hyperparameters, could estimate the parameter. The central 'impossibility' conclusion is therefore at best a statement about the two tested planners under one proxy, not about the identifiability of parameters in simultaneous-variation scenarios.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an empirical evaluation of Causal Curiosity, a reinforcement-learning method for estimating causal factors (mass, size, lateral friction, spinning friction, gravity) in the CausalWorld TriFinger robotic manipulation simulator. The authors study five research questions: accuracy across different factors, granularity via repeated bipartition of factor ranges, the effect of reducing the gap between parameter ranges, simultaneous variation of multiple factors, and causally related or confounding factors. They propose augmenting the original Silhouette-based reward with an F1 classification score and compare the original CEM open-loop planner with a PPO-based planner that they adapt to share one action sequence across all environments. The reported results show perfect F1 classification in single-factor experiments, successful estimation at fine granularity, degraded clustering scores when gaps shrink, frequent failures when two factors vary simultaneously, and mostly successful estimation in the two confounding scenarios. The paper concludes that single-factor estimation is robust and highly refinable, that simultaneous variation of two causal factors represents a hard limit, and that the PPO planner offers modest robustness gains over CEM.","tokens_in":24008,"tokens_out":4354,"duration_ms":44334,"significance":"If the central conclusions are accepted, the paper would be a useful benchmark for Causal Curiosity, since it is, to my knowledge, the first systematic accuracy and failure-mode analysis of the method across multiple causal factors, granularities, and confounding structures. The authors are commendably explicit about their comparative rather than best-case experimental design, and the introduction of an F1 classification term is a sensible extension for detecting situations where clustering separates trajectories but not the true factor groups. The paper also provides concrete comparative evidence that factor identifiability is factor-dependent. However, the strongest claimed contribution, the 'impossibility' of multi-factor estimation, is not supported by the experimental design as presented, because the negative results come from two non-adaptive, deliberately untuned planners evaluated under a clustering proxy. The proper contribution of the paper is a careful empirical stress test of two specific planners under that proxy, and the manuscript should be revised to state that scope precisely.","major_comments":[{"comment":"The Introduction's claim that the paper shows 'the impossibility of solving some complex experimental situations ... with multiple parameters varying simultaneously' overstates what Section 4.4 can support. The negative results are obtained with exactly two planners, both of which are non-adaptive open-loop controllers: the CEM planner of §3.5, and the PPO planner that is deliberately modified so that all environments execute identical actions and receive the reward only at the last timestep (§3.5). Section 3.6 additionally states that hyperparameters were 'defined without the need for hypertuning processes' because the aim was comparative, not best-possible, results. A failure of these two untuned, non-adaptive planners under the clustering proxy of Eq. (6) is evidence about the limits of those planners, not about the identifiability of the causal factors. I recommend replacing 'impossibility' with a claim scoped to the tested planners, such as 'open-loop planners with fixed hyperparameters fail on several two-factor combinations under this clustering objective,' and, if feasible, adding a closed-loop or adaptive baseline to test whether the difficulty persists.","section":"§4.4, Introduction, §3.5, §3.6"},{"comment":"The evaluation protocol for the F1 component of Eq. (6) is ambiguous and potentially circular. Section 3.4 describes a protocol in which the clustering function qM is learned from one set of trajectories and then used to classify new objects, but Section 3.6 states that 'the trajectories are used to define a clustering function, which allows to obtain the scores' within the same experiment, without mentioning a held-out set. If F1 is computed on the same 20 trajectories used to define qM, then the reported F1 scores measure training-set separation rather than out-of-sample classification accuracy; this would inflate the F1=1.00 results in Tables 2-4 and weaken the interpretation of the low-F1 results in Tables 5-8. The manuscript should unambiguously state whether F1 is computed on held-out trajectories, and if it is not, the analysis should be rerun with a train/test split.","section":"§3.4, §3.6, Eq. (6)"},{"comment":"The quantitative results are reported only as point estimates, without standard deviations or confidence intervals, even though Section 3.6 states that each experiment comprises 100 replications. This is load-bearing for the central limit claims: for example, Table 5 reports Spinning Friction F1 values of 0.65, 0.55, 0.75, and 0.65 depending on the secondary factor, and without error bars or a chance-level baseline one cannot distinguish reliable above-chance performance from an unreliable one. Reporting mean ± standard deviation over the 100 replications, together with a random-classification baseline for the 20-environment setup, would substantially strengthen the conclusions drawn from these tables.","section":"§3.6, Tables 5-8"}],"minor_comments":[{"comment":"The word 'planer' is used instead of 'planner' in many places (e.g., §3.5, §3.6, and Tables 2-9); a global typo correction is needed.","section":"Throughout"},{"comment":"The text says the PPO planner is implemented from 'the PPO2 implementation of Stable Baselines,' while earlier in the same section Stable Baselines 3 is named; the exact library version and wrappers should be clarified for reproducibility.","section":"§3.6"},{"comment":"Equation (7) renders the anticausal inequality as 'G ≠ fG(L) + N2,' which is unclear; the intended statement is presumably that G cannot be expressed as an additive function of L, and this should be written out explicitly.","section":"§4.5, Eq. (7)"},{"comment":"The exclusion of Mass-Gravity combinations is justified only by saying their effects 'partially counteract each other'; a more detailed explanation is needed, since this omission changes the completeness of the factor-pair analysis.","section":"§4.4, Tables 5-8"},{"comment":"The cluster labels C1-C6 referenced in Table 9 are not marked on Figures 9 and 10; labeling the panels directly would make the table much easier to interpret.","section":"§4.5, Figures 9-10, Table 9"}],"recommendation":"major_revision","confidential_remarks":"The main risk to the paper is the gap between its stated 'impossibility' claim and what the experimental protocol actually shows. The empirical material is valuable, but the authors should be required to either reframe the claim to the tested planners and proxy or add an adaptive baseline and tuned hyperparameters that demonstrate the difficulty persists. The ambiguity about whether F1 is evaluated on the training trajectories should also be resolved before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper you asked about is a useful empirical boundary study of Causal Curiosity, and the authors deserve credit for putting the method under pressure. What's genuinely new: they add an F1 classification term to the reward, use PPO as the search planner rather than only CEM, and run scenarios the original proposal didn't consider — two factors varying at once, and factors with causal relationships including a confounder. The high-precision estimation result in Section 4.2 is the strongest positive finding: they push parameter resolution three orders of magnitude finer, and PPO often beats CEM on clustering score. The Section 4.5 causal/confounding experiments are a reasonable first step, and the authors are properly cautious about extrapolating from them.\n\nThe soft spots are real and not equally serious. The main one is the 'impossibility' claim in the Introduction and Discussion. Section 4.4 shows two planners failing, not that simultaneous multi-factor estimation is impossible. The authors explicitly say hyperparameters were deliberately not tuned, and both planners are open-loop or effectively non-adaptive: the CEM planner samples a fixed action sequence, and the PPO wrapper forces all environments to execute the same actions with reward only at the last timestep. So the negative result is about the tested planners under one clustering proxy, not about identifiability of the parameters. The Discussion's 'hard limit' wording should be toned down.\n\nTwo evaluation-protocol concerns are also legitimate. Section 3.4 describes using new objects for F1 evaluation, while Section 3.6 says the same 20 trajectories are used to define the clustering function and score it. The paper should state plainly whether reported F1 is training or held-out. Most tables lack error bars or per-replication variance, which matters because several key scores (e.g., 0.85, 0.65) are near chance for 20 environments. And while the authors point to the original code, they don't ship their own modified code or data; for an empirical paper, that weakens reproducibility.\n\nNone of this is fatal. The central empirical pattern — single-factor estimation works and can be very precise, while simultaneous factor variation degrades performance — is credible. The problem is the gap between that pattern and the word 'impossibility.' Researchers applying or extending Causal Curiosity in robotics or causal RL will get a practical map of where the method breaks. It deserves serious peer review with major revision: reframe the strong claim, clarify the evaluation protocol, and ideally release the code. I would cite it as an empirical stress-test.","headline":"A genuinely useful empirical stress-test of Causal Curiosity whose central 'impossibility' claim overreaches: the evidence shows two untuned, non-adaptive planners failing, not that simultaneous multi-factor estimation is impossible.","tokens_in":24512,"tokens_out":2116,"would_cite":true,"duration_ms":22410,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Causal Curiosity accurately estimates a single causal factor in a simulated robotic-manipulation task and can refine that estimate three orders of magnitude, but the paper claims that when two factors vary simultaneously the method can…","keywords":["reinforcement learning","causal analysis","parameter estimation","sensitivity analysis","dynamical systems","causal curiosity","robotic manipulation","confounding factors"],"falsifier":"Run the hardest multi-factor cell (spinning friction as target with size as the secondary factor) under an exhaustive grid over the six-step control sequence, or with a closed-loop policy that re-aims the robot based on live object motion: if any search finds F1 = 1.0 where the paper reports 0.55, the claimed impossibility is an artifact of the planners' search rather than a property of the physical system.","tokens_in":1721,"feed_emoji":"🤖","tokens_out":3990,"duration_ms":102875,"temperature":0.7,"pith_summary":"This paper measures how far the Causal Curiosity reinforcement-learning approach can go as a practical tool for estimating unknown causal parameters of a dynamical system, using a simulated robot manipulator pushing an object. The paper claims that in ideal single-factor experiments the method works well: it classifies all five tested factors (mass, size, lateral friction, spinning friction, gravity) perfectly and can refine the estimate to about one-thousandth of the original range by repeated bipartition. It also finds a hard limit: when a second factor varies at the same time, identification of the target factor often fails outright, with spinning friction the hardest and size the easiest to recover. A proposed PPO-based planner plus an F1 classification term in the reward gives modest robustness gains over the original cross-entropy planner.","feed_headline":"Causal curiosity fails when two factors change at once","feed_subtitle":"One-factor estimation is accurate and refinable down to 0.1% of the range; the hard limit appears with simultaneous variation.","key_machinery":"The load-bearing device is the cluster-separation reward of Eq. (6), $R = C(q_M(o^i_{0:T}), z^i) + k \\cdot S(o^i_{0:T})$, where $C$ is an F1 classification score comparing the learned clustering of object trajectories with the true parameter-range labels, $S$ is a Silhouette score measuring how well the two trajectory clusters separate (using soft dynamic time warping as the distance), and $k = 0.1$ weights separation below correctness. This reward is what the planners optimize and what the paper uses to declare success or failure: a parameter is judged estimable exactly when some action sequence makes the trajectories of the two parameter ranges form distinguishable groups. The paper compares two search mechanisms over that reward, the original open-loop cross-entropy model-predictive planner and a closed-loop policy-gradient (PPO) planner, and adds recursive bipartition of the parameter range to push resolution to very fine granularity.","core_discovery":"On the paper's own terms, the central discovery is a map of where Causal Curiosity succeeds and where it breaks. The method's reward—correct classification of each trajectory plus cluster separation—is sufficient to drive accurate parameter estimation when only one factor is varied and the two ranges are well separated; under these conditions the paper reports F1 scores of 1.00 for all five factors and demonstrates refinement to ranges about 0.14% of the initial width. The paper's headline limitation is the claim that it is impossible to solve some of the newly studied multi-factor scenarios: when two uncorrelated factors vary simultaneously, the F1 score for the target factor can drop to as low as 0.55 (spinning friction against size as the secondary factor), and success depends strongly on which factor is the target and which is the confounder. Causally related and confounding factors, modelled with additive noise models, were mostly identified correctly. The paper reads these results as showing that Causal Curiosity is a generalisable method for single-factor estimation, but that robustness to simultaneous variation is system- and factor-dependent.","pith_inferences":["Beyond the paper's own claims, the 'impossibility' result is a statement about the tested planners and observation space, not a formal identifiability proof—a closed-loop policy that conditions actions on live observations, or an observation space enriched with contact forces, might succeed where these planners fail.","The reward's geometric structure suggests a testable prediction: difficulty with two varying factors should track the overlap of the induced trajectory distributions; computing the Bayes-optimal classification error on trajectories before running the planners would settle whether the planners or the physics are the bottleneck.","The clustering score could plausibly be reused as an uncertainty estimate for adaptive partitioning, turning the agnostic bipartition into a Bayesian-style exploration strategy that splits ranges where uncertainty is highest.","The reported factor hierarchy (size easy, spinning friction hard) may serve as an evaluation target for causal representation learning: a model that captures true causal structure should be able to predict this ordering from the system's dynamics alone."],"forward_implications":["In controlled single-factor experiments, all five causal factors are classified with F1 = 1.00 and comparable clustering scores, so the method is factor-agnostic in easy conditions.","Repeated bipartition estimates a parameter to roughly three orders of magnitude finer than the starting range while clustering scores generally improve, implying the method can serve as a high-resolution estimator rather than a coarse classifier.","Shrinking the gap between the two parameter ranges to about 4% of their size lowers clustering scores by about a quarter on average and can cause outright misclassification, so range separation, not just range width, drives reliability.","Simultaneous uncorrelated variation of a second factor breaks identification for some factor pairs: the target factor's F1 can fall to 0.55, and factor identity matters more than the choice of planner.","Causally related and confounding factor scenarios (additive noise models) are mostly identified correctly, suggesting the method can handle structured dependence when the causal graph is known."],"supporting_citations":[{"why":"Proposes the Causal Curiosity method whose accuracy limits this paper measures.","marker":"[21]"},{"why":"Supplies the PPO algorithm used as the alternative planner.","marker":"[31]"},{"why":"Defines the cross-entropy optimization used by the baseline CEM planner.","marker":"[32]"},{"why":"Frames the model predictive control planner that the CEM method optimizes.","marker":"[33]"},{"why":"Defines the Silhouette score used to measure trajectory cluster separation.","marker":"[83]"},{"why":"Defines the soft dynamic time warping distance used to compare trajectories.","marker":"[84]"},{"why":"Supplies the additive noise models used to construct the causally related and confounding factor experiments.","marker":"[91]"}],"fun_headline_variants":["Causal curiosity succeeds alone, fails when factors pair up","Causal curiosity's multi-factor limit: simultaneous changes break estimation","Reinforcement learning causal estimation: single-factor only, multi-factor fails","Causal curiosity: high accuracy for one factor, but two changes at once break it","Simultaneous factor changes expose causal curiosity's biggest limit"],"cache_read_input_tokens":26624,"weakest_assumption_plain":"The paper equates successful parameter estimation with finding a single action sequence whose object trajectories split into two well-separated clusters matching the two parameter ranges, as judged by the F1 and Silhouette scores; if a different way of choosing actions or observing trajectories could separate groups that this scheme cannot, the reported limits are limits of the scheme, not of what is knowable about the parameters.","fun_headline_variants_meta":{"raw":{"variants":["Causal curiosity succeeds alone, fails when factors pair up","Causal curiosity's multi-factor limit: simultaneous changes break estimation","Reinforcement learning causal estimation: single-factor only, multi-factor fails","Causal curiosity: high accuracy for one factor, but two changes at once break it","Simultaneous factor changes expose causal curiosity's biggest limit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001098,"raw_usage":{"total_tokens":4581,"prompt_tokens":944,"completion_tokens":3637,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":3546}},"tokens_in":560,"tokens_out":3637,"duration_ms":22196,"temperature":1.0,"reasoning_tokens":3546,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:54:28.014823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the hardest multi-factor cell (spinning friction as target with size as the secondary factor) under an exhaustive grid over the six-step control sequence, or with a closed-loop policy that re-aims the robot based on live object motion: if any search finds F1 = 1.0 where the paper reports 0.55, the claimed impossibility is an artifact of the planners' search rather than a property of the physical system.","supporting_citations":[{"cited_title":"Annals of operations research 134, 19–67 (2005)","cited_arxiv_id":null,"evidence_quote":"Defines the cross-entropy optimization used by the baseline CEM planner."},{"cited_title":"In: International Conference on Machine Learn- ing, vol","cited_arxiv_id":null,"evidence_quote":"Defines the soft dynamic time warping distance used to compare trajectories."}],"review_version":1}