{"id":"e1285ef7-c68c-4ce5-87bd-2e6272c43644","arxiv_id":"2506.10778","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SlotPi combines a learned Hamiltonian energy module with spatiotemporal attention to improve object-centric video prediction and visual question answering on several datasets.","lead":"SlotPi adds a Hamiltonian-based physical module to slot-based video prediction models, and tests it on object, fluid, and real-world water scenes. The model reports small gains on benchmarks and a new real-world dataset, but no code or error bars are provided yet.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Hamiltonian constraint in SlotPi is not shown to be load-bearing: Eq. (6) sums the physics update with an unconstrained residual at λ=1, and no symplectic/conservation check or same-capacity ablation isolates the physical module.","rationale":"I read the central claim as an architecture claim: a slot-based Hamiltonian module combined with a spatiotemporal residual yields state-of-the-art prediction across rigid-body, fluid, and coupled scenes. The reported numbers support that the full system predicts well on several datasets, but they do not establish that the improvement is attributable to Hamiltonian physics. The reader's weakest assumption identifies the unvalidated phase-space interpretation and the residual in Eq. (6); my reading agrees and adds that Phat_{t+1} is never used, so the architecture is not even a closed Hamiltonian rollout. This does not overturn the paper—the full model may still be a useful predictive architecture—but it changes what can be claimed: without the proposed ablation and conservation check, 'physics-informed' is a label rather than a demonstrated mechanism. Therefore the conditional verdict stands unchanged; the paper should be accepted only if the artifacts are released and the physical-module attribution is established.","tokens_in":18436,"tokens_out":8292,"duration_ms":68785,"concrete_test":"Decisive check on CLEVRER (the dataset with the largest reported gains): train four matched variants—full SlotPi; λ=0 (cST alone); λ=1 with the Hamiltonian gradient update replaced by a same-parameter-count MLP residual; and full SlotPi with Phat_{t+1} fed back as the next-step momentum. Report PSNR/SSIM/LPIPS, FG-ARI/FG-mIoU, and the drift of latent energy H_t over the 42-step rollout. If the non-Hamiltonian or λ=0 variant is within about 0.1 PSNR or 0.5 FG-ARI of full SlotPi, or if |H_t−H_0|/H_0 exceeds about 20%, the Hamiltonian module is not the demonstrated source of the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that SlotPi's strong prediction and adaptability results come from embedding Hamiltonian physics into object-centric reasoning. That claim requires (i) the attention-computed Q_t and P_t in Eq. (1) to be valid canonical coordinates, (ii) H_t in Eq. (2) to be a meaningful Hamiltonian, and (iii) Eq. (4)'s update to actually constrain the final prediction in Eq. (6). None of these is verified. Q_t and P_t are arbitrary attention outputs of the same S_t (P_t additionally attends over history), with no symplectic structure or coordinate independence enforced. H_t is recomputed at every step from a history-dependent attention, and only Qhat_{t+1} from the Euler step enters Eq. (6); Phat_{t+1} from Eq. (4b) is discarded and never fed back. Eq. (6) is a plain sum with λ=1, so the spatiotemporal residual cST can, in principle, account for the full prediction without any help from the physical module. The §4.4 ablation varies λ but omits the claimed λ=0 baseline, and it does not compare against a same-capacity non-Hamiltonian update. Thus the reported state-of-the-art numbers are consistent with the null hypothesis that the gains come from added capacity, not from physical inductive bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SlotPi, a slot-based object-centric dynamics model that combines a Hamiltonian-style physics module (Eqs. 1-4) with a spatiotemporal attention module (Eq. 5) through a weighted sum (Eq. 6). The model is evaluated on video prediction and VQA on OBJ3D, CLEVRER, Physion, on a Navier-Stokes fluid dataset, and on a new real-world fluid-object dataset collected by the authors. The central claims are that SlotPi achieves state-of-the-art results on all datasets and that the Hamiltonian module provides physical constraints that improve predictive accuracy and cross-domain adaptability.","tokens_in":18764,"tokens_out":7225,"duration_ms":71273,"significance":"The paper's potential contribution is a single object-centric framework that unifies rigid-body, fluid, and coupled fluid-object prediction, together with a new real-world dataset and an ablation of Hamiltonian embedding choices. The manuscript is transparent about its two-stage training pipeline, and the limitations section is candid. However, the current evidence does not establish the load-bearing role of the Hamiltonian constraint, and the 'state-of-the-art on all datasets' claim is contradicted by the OBJ3D PSNR result in Table 1. With additional isolation experiments and direct physical-consistency metrics, the contribution could be solid.","major_comments":[{"comment":"The ablation for the physical module is incomplete and does not support the text. The paragraph states that 'When the physical module is removed (i.e., λ=0), the model degrades to a structure similar to STATM or Slotformer, resulting in inferior performance,' but Table 7 reports only λ=1, λ=0.1, and a learnable λ; no λ=0 row is shown. Furthermore, there is no same-capacity non-Hamiltonian update baseline. Because Eq. (6) is a plain sum with λ=1, the unconstrained residual cST could in principle account for the full prediction, so the reported gains over STATM may be due to added capacity rather than to the physics-informed inductive bias. This is the central load-bearing point of the paper and needs a direct ablation.","section":"Section 4.4 / Table 7 / Eq. (6)"},{"comment":"The claim that 'SlotPi achieves state-of-the-art results on all datasets' is not supported by Table 1: SAVi-Dyn attains PSNR 32.94 while SlotPi attains 32.67 on OBJ3D, with SSIM tied. Additionally, no standard deviations or numbers of seeds are reported for any experiment. On CLEVRER the PSNR gap over STATM is only 0.19 dB (30.41 vs 30.22), which is within typical run-to-run noise for this class of models; without repeated runs the 'state-of-the-art' performance claim is not statistically grounded.","section":"Abstract, Contributions, Table 1"},{"comment":"The paper does not verify that the attention-computed Q_t and P_t form a valid Hamiltonian phase space. Q_t and P_t are both functions of the same slot representation S_t (Eq. 1), with no symplectic structure or canonical-coordinate test. Moreover, the predicted momentum Phat_{t+1} from Eq. (4b) is discarded and never used in Eq. (6), and the final prediction adds an unconstrained residual cST with λ=1. No energy conservation, symplecticity, or other direct physical-consistency metric is reported. Therefore the statement in Section 4.1 that the model's predictions are 'more physically consistent' is an inference from image-quality metrics, not a demonstrated property.","section":"Section 3.1-3.2, Eqs. (1)-(6), Section 4.1"},{"comment":"The fluid experiments use the PDE-Refiner dataset (ref [43]) but compare only against FNO and U-Net. PDE-Refiner itself, a strong baseline specifically designed for long-rollout accuracy on this benchmark, is omitted. Without this comparison and without error bars, the claim of 'significantly outperforms other models' on the Navier-Stokes task is not fully established.","section":"Section 4.2, Table 5"}],"minor_comments":[{"comment":"The definition S_t = {s(0,t),...,s(N,t)} ∈ R^{N×D} is inconsistent because it lists N+1 slots while claiming N slots; Eq. (7) sums i=0 to N, which reinforces the inconsistency. Please clarify whether there are N or N+1 slots.","section":"Section 3, notation"},{"comment":"'High Correction Time (HCT)' appears to be a typo; if it is 'High Correlation Time', please correct and define it (e.g., time until correlation drops below a threshold).","section":"Section 4.2, metric name"},{"comment":"'second rand' should be 'second rank' (or 'second place').","section":"Section 4.1, Table 3 discussion"},{"comment":"Please report the numerical values of the initial and final learned λ and the training epoch count; the description 'gradually increases, but it does not exceed 1' is vague.","section":"Section 4.4, Figure 6"},{"comment":"The related-work section covers many references but does not discuss the relationship to generalized Hamiltonian learning methods (e.g., neural symplectic form) in detail; a brief positioning sentence would help.","section":"Section 2, Related Work"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper's topic fits the conference, and the released code/data are valuable. The main risk is that the physical-module contribution is not isolated; I recommend requiring the λ=0 and same-capacity baselines, error bars over at least 3 seeds, and a PDE-Refiner comparison before acceptance. Also, the repeated claim of SOTA should be aligned with Table 1. The heavy reliance on the authors' own STATM baseline is acceptable but should be accompanied by independently reproduced baselines for all tables."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read SlotPi as a direct extension of STATM (same group) with a Hamiltonian-style gradient module added on top, evaluated on object, fluid, and fluid-object benchmarks. The genuinely new piece is the real-world fluid-object dataset—603 training videos of objects floating in dyed water in indoor and outdoor settings. That dataset and the fluid rollout results are the most useful part of the submission, and the qualitative figures suggest the model does capture slow fluid-driven motion.\n\nWhat the paper does not show is that the Hamiltonian module is doing the work the title claims. Eq. (6) is a plain sum of the physical update and an unconstrained spatiotemporal residual, with λ=1 everywhere except the ablation. Nothing prevents the residual from carrying the full prediction. The ablation in §4.4 compares λ=1, λ=0.1, and a learnable λ, but omits the λ=0 baseline that would remove the physics module entirely, and it never compares against a same-capacity non-Hamiltonian update. So the reported gains are consistent with added capacity rather than physical inductive bias. On top of that, Eq. (4b) computes P̂_{t+1} but only Q̂_{t+1} from Eq. (4a) enters Eq. (6); the momentum prediction is discarded. That makes the 'Hamiltonian phase space' story much harder to defend. The stress-test note is fair on all of this.\n\nThe empirical reporting has soft spots in proportion. Table 1 shows SlotPi slightly below SAVi-Dyn on OBJ3D PSNR (32.67 vs 32.94) and equal on SSIM, yet the abstract and contributions say 'state-of-the-art results on all datasets.' That is an overclaim. No error bars or seeds anywhere. The fluid comparison omits PDE-Refiner, which provided the dataset. The real-world comparisons are only against SlotFormer and STATM re-implementations. These are individually minor, but collectively they undercut the strongest claims.\n\nSelf-citation to STATM is not a problem: the method really is an extension of that line, and the paper says so.\n\nWho gets value: people building object-centric world models who want a simple physics-inspired term, and anyone benchmarking on real-world fluid-object scenes. This paper deserves a serious referee: the dataset is useful, the direction is sensible, and the architecture is simple enough to reproduce. But I would want major revisions before acceptance: report the λ=0 and same-capacity ablations, show variance, fix the SOTA claim, and either provide evidence of conservation or symplectic structure in the learned latent dynamics or soften the physics language.","headline":"SlotPi is a credible incremental extension of STATM with a useful new fluid-object dataset, but the paper does not demonstrate that the Hamiltonian module is what drives the reported gains.","tokens_in":19270,"tokens_out":3626,"would_cite":false,"duration_ms":35863,"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":"SlotPi embeds Hamiltonian physics into slot-based object-centric models and claims state-of-the-art prediction across rigid bodies, fluids, and their coupled real-world dynamics.","keywords":["object-centric learning","Hamiltonian neural networks","physics-informed machine learning","video prediction","visual question answering","fluid dynamics","slot attention","world models"],"falsifier":"Run a long rollout on a conservative scene with no external forcing, such as OBJ3D, and track the predicted energy $H_t$ from Eq. (2); if energy drifts as much as it does when the physics module is ablated ($\\lambda=0$), the Hamiltonian term is adding capacity rather than enforcing physics. A sharper check is to test whether the learned update is approximately symplectic by comparing the Jacobian of $(\\hat Q,\\hat P)$ with respect to $(Q,P)$ with a symplectic matrix on held-out trajectories.","tokens_in":18245,"feed_emoji":"⚛️","tokens_out":9340,"duration_ms":102324,"temperature":0.7,"pith_summary":"SlotPi is an object-centric video prediction model that writes Hamilton's equations into the latent dynamics of slot representations. It claims that attention over slots can supply learned generalized coordinates and momenta, and that a learned Hamiltonian computed by self-attention can advance the scene through gradient updates. A parallel spatiotemporal attention module is added to capture non-conservative behavior that Hamiltonian dynamics alone cannot represent. If the claim holds, a single slot-based framework improves prediction quality and predictive visual question answering on rigid-object benchmarks, competes with neural operator methods on Navier-Stokes fluid data, and extends to real-world fluid-object interactions; the authors report state-of-the-art results on every dataset tested.","feed_headline":"Physics-informed slots beat prior world models on every test set","feed_subtitle":"A single architecture embeds Hamiltonian physics in slot attention to improve video prediction and visual question answering.","key_machinery":"The load-bearing object is a latent Hamiltonian layer built entirely from slot attention. Cross-attention maps each slot's history to generalized momentum $P_t$, self-attention maps the current slots to generalized coordinates $Q_t$, and a self-attention energy head with Softplus yields $H_t$. The layer then applies Hamilton's equations and an Euler integrator to produce $\\hat Q_{t+1}$ and $\\hat P_{t+1}$, and the final prediction is $\\hat S_{t+1}=\\lambda \\hat Q_{t+1}+\\hat{ST}_{t+1}$, where $\\hat{ST}_{t+1}$ is the output of a spatiotemporal attention module. The physics layer is what carries the inductive bias; the residual module is what lets the hybrid handle dissipation and non-conservative forces.","core_discovery":"SlotPi's central claim is that constraining slot evolution with a learned Hamiltonian layer makes object-centric dynamics prediction both more accurate and more transferable. Starting from slots $S_t$ extracted by an upstream object-centric model, cross-attention over slot histories yields generalized momenta $P_t$ and self-attention yields generalized coordinates $Q_t$; a self-attention head with Softplus produces per-slot energies whose sum is the Hamiltonian $H_t$. Hamilton's equations $\\dot Q_t=\\partial H_t/\\partial P_t$ and $\\dot P_t=-\\partial H_t/\\partial Q_t$ are integrated with an Euler step to predict $\\hat Q_{t+1}$ and $\\hat P_{t+1}$, and the final slot prediction is the weighted sum of this physical state and an unconstrained spatiotemporal residual. The authors report that SlotPi outperforms SlotFormer, STATM, FNO, and U-Net on OBJ3D, CLEVRER, Physion-Collide VQA, the PDE-Refiner Navier-Stokes benchmark, and their own real-world fluid-object dataset, with the largest margins on the most interaction-heavy scenes.","pith_inferences":["The paper leaves implicit that the biggest gains are reported on CLEVRER, the most interaction-heavy benchmark, so the Hamiltonian prior may matter most when collisions and rearrangements dominate rather than on simple or slow scenes.","Because Eq. (6) adds an unconstrained residual at $\\lambda=1$, SlotPi is best understood as a Hamiltonian-regularized hybrid rather than a strictly conservative simulator, and on non-conservative scenes the residual may carry much of the signal.","A natural extension would be to restrict the residual module to learned dissipation (for example, a Rayleigh-like damping term), which would make the Hamiltonian constraint testable and recover exact symplectic structure in the conservative limit.","The fluid experiments treat image patches as slots rather than discovered objects, suggesting that the object-centric machinery is acting partly as a spatial tokenizer; that reading could connect SlotPi to other PDE surrogate and operator-learning problems."],"forward_implications":["Slot-based world models can carry an explicit energy-based inductive bias without knowing the governing equations in advance.","The same architecture transfers from rigid-body scenes to pure Navier-Stokes fluid prediction, so the physical constraint need not be task-specific.","More physically consistent rollouts translate into higher accuracy on CLEVRER predictive questions and on Physion-Collide, closing part of the gap toward human performance.","The real-world fluid-object dataset demonstrates that the model can capture coupled behaviors such as floating objects being pushed, rotated, and held back by fluid motion."],"supporting_citations":[{"why":"Supplies the SAVi slot-extraction backbone used for the OBJ3D experiments and as the base object-centric model.","marker":"[33]"},{"why":"Provides the spatiotemporal slot-attention predictor that SlotPi extends, and the slot extractor used on CLEVRER and the real-world dataset.","marker":"[39]"},{"why":"Is the transformer-based slot dynamics baseline whose predictions and pretrained weights the paper builds on and must beat.","marker":"[77]"},{"why":"Supplies the slot extractor for the Physion VQA experiments.","marker":"[63]"},{"why":"Provides the benchmark videos and predictive-question VQA setup where the main object-dynamics results are measured.","marker":"[79]"},{"why":"Provides the physical-prediction VQA benchmark, including the human upper bound the model is compared with.","marker":"[5]"},{"why":"Supplies the Navier-Stokes trajectories used as the fluid dataset.","marker":"[43]"},{"why":"Is the neural operator baseline for the fluid prediction comparison.","marker":"[40]"},{"why":"Is the other fluid prediction baseline that SlotPi outperforms on RMSE, MAE, and HCT.","marker":"[59]"},{"why":"Provides the standard Hamiltonian-network formulation whose structure the physics module adapts to slot attention.","marker":"[21]"}],"fun_headline_variants":["SlotPi: Hamiltonian slots beat prior world models everywhere","Physics-infused slots improve prediction and VQA across datasets","SlotPi's learned Hamiltonian boosts object-centric reasoning","Object-centric physics model outperforms on fluids and interactions","SlotPi transfers across object, fluid, and real-world scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the numbers the attention mechanism computes really are positions and momenta of a physical system, so the model's energy estimate is a true energy whose gradients drive correct motion; the paper does not test conservation or symplectic structure.","fun_headline_variants_meta":{"raw":{"variants":["SlotPi: Hamiltonian slots beat prior world models everywhere","Physics-infused slots improve prediction and VQA across datasets","SlotPi's learned Hamiltonian boosts object-centric reasoning","Object-centric physics model outperforms on fluids and interactions","SlotPi transfers across object, fluid, and real-world scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1765,"prompt_tokens":980,"completion_tokens":785,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":707}},"tokens_in":596,"tokens_out":785,"duration_ms":9492,"temperature":1.0,"reasoning_tokens":707,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:19:03.396949+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a long rollout on a conservative scene with no external forcing, such as OBJ3D, and track the predicted energy $H_t$ from Eq. (2); if energy drifts as much as it does when the physics module is ablated ($\\lambda=0$), the Hamiltonian term is adding capacity rather than enforcing physics. A sharper check is to test whether the learned update is approximately symplectic by comparing the Jacobian of $(\\hat Q,\\hat P)$ with respect to $(Q,P)$ with a symplectic matrix on held-out trajectories.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the SAVi slot-extraction backbone used for the OBJ3D experiments and as the base object-centric model."},{"cited_title":"Reasoning-Enhanced Object-Centric Learning for Videos","cited_arxiv_id":"2403.15245","evidence_quote":"Provides the spatiotemporal slot-attention predictor that SlotPi extends, and the slot extractor used on CLEVRER and the real-world dataset."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the slot extractor for the Physion VQA experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the standard Hamiltonian-network formulation whose structure the physics module adapts to slot attention."}],"review_version":1}