{"id":"bc89dc0e-7f86-49ce-82bb-f0581b46c26a","arxiv_id":"2608.07981","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PhyS adds physics-aware video data, teacher distillation, and windowed reward routing to make streaming world models generate more physically plausible long rollouts.","lead":"This paper trains video generation models to respect basic physics by fine-tuning them on 120,000 real-world interaction videos, then distilling the knowledge into a small fast model and aligning it with a physics reward. A reader interested in world models or embodied AI should look at whether physics-aware training can make long video rollouts more reliable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage 3's V-JEPA2+RAFT reward is validated only on in-distribution training prompts (0.65 pairwise agreement, Fig. 7a); if agreement drops on benchmark prompts, RL gains may reflect reward overfitting rather than physical plausibility.","rationale":"The reader's weakest assumption is precisely the load-bearing point. The paper's claimed mechanism is that freezing a V-JEPA2 predictor as a physical-consistency reward and routing it via TCR yields more physically plausible rollouts. If that reward is an invalid or non-generalizing proxy, the entire RL stage loses its justification. I considered other potential concerns—missing code/data (reproducibility), no error bars (statistical significance), and uneven PhyGenBench categories (acknowledged)—but those affect confidence rather than directly falsifying the causal claim. The reward-validity issue is unique in that it can explain the reported benchmark improvements as an artifact of optimizing an in-distribution proxy. The concrete test of measuring reward-human agreement on benchmark prompts would settle whether the 0.65 agreement extends across distributions. If it does, the concern is mitigated; if not, the central claim weakens. Thus UNCHANGED: the reader's CONDITIONAL verdict remains appropriate, pending this validation.","tokens_in":14670,"tokens_out":8991,"duration_ms":91156,"concrete_test":"Construct a held-out evaluation set of ~500 prompts sampled from the benchmark distributions (PhysicsIQ and VideoPhy at minimum). Generate rollouts with the final PhyS model and the forcing baselines, then have multiple human raters perform pairwise preference judgments on physical plausibility. Compute the agreement (e.g., Spearman rank correlation or pairwise accuracy) between the V-JEPA2+RAFT reward (and V-JEPA2 alone) and the human judgments on this benchmark set. If the agreement is significantly lower than the 0.65 reported on training prompts (e.g., <0.55, or within chance), the reward does not generalize to the evaluation distribution, and the RL component's reported gains cannot be attributed to genuine physical understanding.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PhyS's central empirical claim rests on Stage 3: online RL with Temporal Credit Routing improves physical plausibility of streaming rollouts. The optimization signal is a composite reward R = R_VJ + 0.4*R_Flow (Section 4.3), where R_VJ compares V-JEPA2's predicted future representation from context C_j with the representation encoded from generated window W_j, and R_Flow encourages motion via RAFT. The only evidence that this reward tracks physical plausibility is Fig. 7a: a 0.65 pairwise ranking agreement with human raters on 1,000 'training prompts'—presumably from the RL training distribution. This is moderate and in-distribution. The paper does not report reward-human agreement on any benchmark distribution (PhysicsIQ, VideoPhy, VideoPhy2, PhyGenBench). If V-JEPA2's consistency notion (e.g., representation matching) fails to capture physical violations on out-of-distribution challenging prompts, the RL stage could optimize the proxy while the benchmark gains stem from reward hacking (e.g., inflating motion via the RAFT term). The 0.65 agreement is also below typical human-human agreement and is not compared against a chance baseline. Since the abstract's headline improvements (23.7%/14.8%/31.4% over forcing baselines) include the RL contribution, this reward-validity gap is load-bearing. The paper also lacks any analysis of reward hacking or correlation between the reward and the evaluation metrics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PhyS, a three-stage framework for injecting physical priors into streaming video world models. Stage 1 constructs PhyS-120K, a dataset of 120,804 real-world physical-interaction videos with structured annotations from Qwen3-VL, and performs physics-aware SFT on a Wan2.1-14B bidirectional DiT teacher. Stage 2 distills the teacher into a 1.3B causal DiT using DMD under Self, Rolling, or Causal Forcing. Stage 3 applies online RL with a composite V-JEPA2 + RAFT reward and Temporal Credit Routing (TCR), which computes per-window group-normalized advantages and routes them to temporally aligned denoising actions. The paper reports that PhyS improves PhysicsIQ by 18.2% over the teacher and by 23.7%, 14.8%, and 31.4% over the three forcing baselines, with additional gains on VideoPhy, VideoPhy2, and PhyGenBench, and on long-horizon rollouts.","tokens_in":14978,"tokens_out":5013,"duration_ms":50124,"significance":"If the reported results are robust, the paper makes a valuable contribution: it introduces a large real-world physics-interaction dataset, a concrete three-stage pipeline for transferring physical priors from a bidirectional teacher to a lightweight causal generator, and a temporally localized credit assignment method that is well motivated and supported by ablations showing an advantage over global reward broadcasting. The experimental design is generally sound in that the final evaluation metrics are external benchmarks rather than the training reward, so the headline gains are not circular by construction. However, the empirical validation has several gaps that currently prevent full confidence in the central claims: missing uncertainty estimates, no quantitative comparison to prior physics-aware methods, and insufficient evidence that the Stage 3 reward is a valid proxy for physical plausibility on the evaluation distributions.","major_comments":[{"comment":"No error bars, confidence intervals, or multiple-seed results are reported for any metric in Tables 1 and 2. Several comparisons involve small absolute differences (e.g., Self Forcing PhysicsIQ 17.4 vs 17.9 with SFT teacher; VideoPhy2 SA 23.6 vs 23.9), which cannot be distinguished from training noise without variance estimates. Please report means and standard deviations over at least three independent runs (or equivalent) for the main comparisons, or justify why single runs are reliable.","section":"Tables 1 and 2"},{"comment":"The only validation of the reward model is a 0.65 pairwise ranking agreement with human preferences on 1,000 training prompts, with no chance baseline, confidence interval, or evaluation on benchmark distributions. Because the headline RL gains are optimized against this reward, the paper should report reward-human agreement on held-out benchmark prompts (e.g., from PhysicsIQ, VideoPhy, VideoPhy2, PhyGenBench) and analyze the correlation between the reward and the final evaluation metrics. Without this, the Stage 3 improvements could reflect optimization of a proxy that does not track physical plausibility on the evaluation distribution.","section":"Section 4.3 and Fig. 7(a)"},{"comment":"The main results compare only against forcing baselines and the teacher model; no quantitative comparison is made to existing physics-aware video generation methods such as WISA, VideoREPA, PhysMaster, or ProPhy, which are discussed in the Related Work. To support the claim that PhyS is a state-of-the-art physics-aware streaming world model, at least one shared-benchmark quantitative comparison with a prior physics-aware method is needed; the qualitative examples in Fig. 5 are not sufficient.","section":"Section 5.2"},{"comment":"The category-level PhyGenBench results show substantial regressions: Rolling Forcing + TCR Material drops from 0.305 to 0.197, and Optics regresses for both Self Forcing (0.451 to 0.402) and Rolling Forcing (0.493 to 0.440). The text attributes this to limited category coverage in PhyS-120K, but this explanation is speculative and the Material regression is large. Please analyze whether the reward or routing induces category-specific failure modes and report per-category results with variance.","section":"Table 1, PhyGenBench rows"},{"comment":"Key hyperparameters of the composite reward and routing—lambda_flow=0.4, window length and stride, group size G=8, and KL coefficient beta=0.015—are not ablated. Since the reported gains depend on these choices, a sensitivity analysis or at least a discussion of their impact would strengthen the claim that TCR is robust rather than tuned to the specific configuration.","section":"Section 5.4 and Table 4"}],"minor_comments":[{"comment":"The percentages such as 18.2%, 23.7%, 14.8%, and 31.4% are relative improvements, but the abstract does not state this explicitly; please clarify to avoid ambiguity.","section":"Abstract"},{"comment":"The expectation in Eq. (9) is not explicitly defined; please specify that it is taken over prompts, the group of G rollouts, and the stochastic denoising transitions.","section":"Section 4.3, Eq. (9)"},{"comment":"Please include a chance-level baseline (0.5) and confidence intervals for the pairwise agreement values in Fig. 7(a), and state how many human raters and prompts were used.","section":"Fig. 7(a)"},{"comment":"The Wan2.1-1.3B row at 40 NFE has a dash for PhysicsIQ; please explain why this model is not evaluated on that benchmark.","section":"Table 1"},{"comment":"The text mentions 'random human inspection' of the VLM-generated annotations but gives no numbers; please report how many clips were audited and the resulting agreement rate.","section":"Section 3"},{"comment":"The sentence 'its window length matches the causal block size, with a half-window stride' leaves the actual frame counts unspecified; please give the concrete number of frames per block and window stride.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is promising and the experimental design avoids the most obvious circularity by using external benchmarks for evaluation. The main risk I see is the Stage 3 reward validity: the 0.65 agreement on training prompts is moderate, and without benchmark-distribution reward evaluation or reward-metric correlation analysis, the RL gains are not fully tied to physical plausibility. I would ask the authors to address the missing error bars, add a quantitative comparison with at least one prior physics-aware method, and provide the reward-correlation analysis. These are substantial but feasible additions, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the dataset is a real contribution and the TCR routing idea is a sensible, modest extension of Flow-GRPO that appears to help. The paper's central claim — that window-level physical rewards routed to temporally aligned denoising actions beat global reward broadcasting — is probably right, but the evidence for the reward proxy is the weakest rung in the ladder.\n\nWhat's actually new: PhyS-120K, a curated set of 120K real-world physical-interaction videos with a process-centric taxonomy and structured annotations (object properties, causal state transitions). That's a reusable asset. The ablations back it up: Table 3 shows structured annotations beat generic captions and WISA-80K on PhysicsIQ and VBench when training scale is matched. TCR itself is straightforward — windowed rewards, group normalization per window, overlap-routed advantages — and the equations in Section 4.3 are coherent. Table 4 does the right experiment, showing overlap routing beats global and uniform routing, and that the V-JEPA2 + RAFT combo is better than either alone.\n\nThe soft spots, in order. No error bars anywhere in Tables 1 and 2, and the main tables don't include any prior physics-aware method (WISA and VideoREPA appear only in qualitative comparisons). The PhyGenBench category regressions — Rolling Forcing Material drops from 0.320 to 0.197 — are acknowledged and plausibly explained by limited dataset coverage, but that's still a big drop to paper over with an average. The load-bearing issue is the reward model: 0.65 pairwise agreement with human raters on 1,000 training prompts, no chance baseline, no evaluation on any of the benchmark distributions, and no analysis of the correlation between reward and the reported metrics. If V-JEPA2's representation-matching notion of physical consistency degrades on harder, out-of-distribution prompts, Stage 3 could be optimizing a proxy that doesn't transfer. That concern is real, but it's not damning: the final metrics are external benchmarks, so the circularity burden is low, and the consistent TCR gains over Flow-GRPO across three forcing baselines suggest the effect is genuine.\n\nWho should read it: anyone working on streaming world models, physics-aware video generation, or RL for diffusion models. The dataset alone justifies a look, and the TCR idea is worth a read for people doing credit assignment in autoregressive video RL.\n\nRecommendation: send to peer review, but with a clear ask for error bars, comparisons to at least one prior physics-aware method, and reward-human agreement on benchmark prompts. Also push for the promised code and data — without them the reproducibility is limited.","headline":"A solid extensional result: the PhyS-120K dataset and the TCR credit-routing tweak are real contributions, but the reward-proxy validation is the load-bearing weak spot.","tokens_in":15519,"tokens_out":2800,"would_cite":true,"duration_ms":28854,"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":"The paper claims that temporal credit routing—applying window-level physical rewards to temporally aligned denoising actions—is what makes online RL improve the physical plausibility of few-step causal video generators.","keywords":["world models","video generation","physical plausibility","temporal credit assignment","reinforcement learning","distribution matching distillation","physics-aware dataset","streaming generation"],"falsifier":"Run the full TCR pipeline with the V-JEPA2 reward replaced by a fixed random score or an inverted score and measure PhysicsIQ; if the model still improves, the gains are not caused by the physical-reward signal. Complement this by measuring V-JEPA2's pairwise agreement with human ratings on held-out PhysicsIQ prompts; a drop toward chance would indicate the training-distribution agreement does not transfer.","tokens_in":14482,"feed_emoji":"🎬","tokens_out":5511,"duration_ms":51153,"temperature":0.7,"pith_summary":"The paper claims that the physical plausibility of streaming video world models can be substantially improved by a three-stage training pipeline. First, a new dataset of 120,804 real-world physical-interaction videos, annotated with object properties and causal event sequences, is used to fine-tune a 14B bidirectional diffusion transformer. That teacher is distilled into a 1.3B few-step causal diffusion transformer, and the student is then aligned by online reinforcement learning. The decisive optimization idea is Temporal Credit Routing: instead of giving every denoising step the same video-level reward, the reward is computed on overlapping temporal windows and routed to the denoising actions that produced those frames. If the results are right, this raises PhysicsIQ scores by 18.2% over the teacher and by 23.7%, 14.8%, and 31.4% over the three forcing-based distillation baselines, with additional gains on VideoPhy, VideoPhy2, and PhyGenBench.","feed_headline":"Routing rewards in time lifts physics video scores by 31%","feed_subtitle":"A three-stage recipe—physics videos, distillation, window-level RL—improves streaming world models on four benchmarks.","key_machinery":"The mechanism that carries the argument is Temporal Credit Routing (TCR), a three-step procedure applied to each group of rollouts. Windows $W_j$ slide over the generated frames; each window is scored by a composite reward $R_{i,j} = R^{\\mathrm{VJ}}_{i,j} + \\lambda_{\\mathrm{flow}} R^{\\mathrm{Flow}}_{i,j}$, combining V-JEPA2 representation agreement with a RAFT optical-flow term; rewards are normalized per window across the group; and the resulting advantages are summed with overlap weights $O_{m,j} = |B_m \\cap W_j| / \\sum_k |B_m \\cap W_k|$ to give each causal block a block-level advantage. This block advantage is then used in the Flow-GRPO policy objective with clipping and KL penalty. The paper's argument is that this localization—overlapping windows rather than one global score—is what lets the policy update correct local physical errors without being diluted by easy or hard temporal regions.","core_discovery":"The paper's central claim is that the main bottleneck in physically plausible streaming generation is not the generator architecture but temporal credit assignment in the alignment stage. It argues that a globally broadcast reward cannot tell a causal model which of its blocks produced a physical violation, so it treats the physical-consistency score as a window-level signal: a frozen V-JEPA2 predictor compares predicted and actual future-frame representations within overlapping windows, supplemented by a RAFT optical-flow term, and the resulting group-relative advantages are routed to temporally aligned denoising transitions. The authors claim this produces more physically coherent and causally consistent long-horizon rollouts than video-level reward broadcasting (Flow-GRPO) across PhysicsIQ, VideoPhy, VideoPhy2, and PhyGenBench, and that the effect holds for all three forcing-based causal distillation baselines.","pith_inferences":["A testable extension is to swap in other window-level scorers, such as physics simulators or human-preference models, inside the same TCR loop; the framework is agnostic to the scorer, so the comparison would isolate whether V-JEPA2's representation agreement is the best proxy for physical plausibility.","The reported 0.65 pairwise agreement with human ratings is measured on training prompts, so a natural next experiment is to measure reward-model agreement on the benchmark distributions themselves; a drop toward chance would suggest part of the PhysicsIQ gain is optimizing the reward rather than genuine physical understanding.","The uneven PhyGenBench category results point to a dataset-coverage bottleneck: expanding PhyS-120K with optics and material-interaction clips could directly close those gaps, a prediction the paper's own regression discussion supports but does not test.","Because TCR only assumes blockwise causal generation and window-level rewards, it could transfer to next-frame autoregressive video models and interactive world models beyond diffusion transformers, though the paper does not demonstrate this."],"forward_implications":["If the central claim is correct, physics-aware supervised fine-tuning on real-world interaction videos is enough to improve a bidirectional teacher on all four physics benchmarks, so data curation can inject physical priors without changing the architecture.","TCR improves each forcing-based causal baseline by 8.8% to 17.5% on PhysicsIQ over its Flow-GRPO counterpart, which means the routing, not just the reward, drives the gain.","The alignment gains appear without increasing inference cost: the 1.3B causal student still uses 4 to 5 network function evaluations.","Long-horizon rollouts at 15 and 20 seconds improve on temporal quality, dynamic degree, visual stability, and VideoPhy physical commonsense, suggesting the benefit does not fade as the rollout extends."],"supporting_citations":[{"why":"Defines Self Forcing, the autoregressive self-rollout training paradigm that the causal student builds on and one of the three forcing baselines.","marker":"Huang et al. 2025"},{"why":"Flow-GRPO supplies the online RL policy optimization objective that Temporal Credit Routing modifies.","marker":"Liu et al. 2025a"},{"why":"Distribution matching distillation (DMD) is the method used to compress the 14B teacher into the few-step 1.3B causal student.","marker":"Yin et al. 2024"},{"why":"V-JEPA2 is the frozen predictor whose representation agreement between predicted and actual future frames forms the semantic component of the physical reward.","marker":"Assran et al. 2025"},{"why":"RAFT provides the optical-flow reward term that encourages sufficient motion in generated videos.","marker":"Teed and Deng 2020"},{"why":"Qwen3-VL is the vision-language model used to produce the structured physics annotations in PhyS-120K.","marker":"Bai et al. 2025"},{"why":"Wan2.1 supplies both the 14B bidirectional teacher backbone and the 1.3B causal student architecture.","marker":"Team Wan et al. 2025"},{"why":"PhysicsIQ is the primary benchmark for context-conditioned image-to-video physics evaluation.","marker":"Motamed et al. 2026"},{"why":"VideoPhy provides the physical commonsense and semantic adherence evaluation for text-to-video generation.","marker":"Bansal et al. 2025a"},{"why":"VideoPhy2 provides the action-centric physical commonsense benchmark with joint success-rate reporting.","marker":"Bansal et al. 2025b"}],"fun_headline_variants":["Window-level rewards fix physics in streaming world models","Temporal credit routing boosts physics video scores by 31%","120K physics videos teach streaming world models deep priors","Window-level RL credit assignment improves physics world models","PhyS: temporal credit routing for physically plausible rollouts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen V-JEPA2-plus-RAFT reward actually measures physical plausibility; the paper supports this with a 0.65 pairwise agreement with human ratings on 1,000 training prompts, and if that agreement does not hold on the benchmark distribution, the reported RL gains would not reflect genuine physics understanding.","fun_headline_variants_meta":{"raw":{"variants":["Window-level rewards fix physics in streaming world models","Temporal credit routing boosts physics video scores by 31%","120K physics videos teach streaming world models deep priors","Window-level RL credit assignment improves physics world models","PhyS: temporal credit routing for physically plausible rollouts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000719,"raw_usage":{"total_tokens":3280,"prompt_tokens":1050,"completion_tokens":2230,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":2151}},"tokens_in":666,"tokens_out":2230,"duration_ms":15567,"temperature":1.0,"reasoning_tokens":2151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:35:00.058931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full TCR pipeline with the V-JEPA2 reward replaced by a fixed random score or an inverted score and measure PhysicsIQ; if the model still improves, the gains are not caused by the physical-reward signal. Complement this by measuring V-JEPA2's pairwise agreement with human ratings on held-out PhysicsIQ prompts; a drop toward chance would indicate the training-distribution agreement does not transfer.","supporting_citations":[{"cited_title":"2020 , organization=","cited_arxiv_id":null,"evidence_quote":"RAFT provides the optical-flow reward term that encourages sufficient motion in generated videos."}],"review_version":1}