{"id":"c9f76cc2-27fc-46a5-b941-e23d721b7c34","arxiv_id":"2508.10858","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A hierarchical direct preference optimization with four alignment levels plus automated data selection improves physical plausibility of text-to-video models.","lead":"This paper trains video generators to prefer physically plausible videos by adding four fine-grained preference signals, instance, state, motion, and semantic, and by selecting training videos from an existing large pool using AI judges. It reports consistent gains on physics and general video benchmarks for CogVideoX and HunyuanVideo.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported physical-plausibility gains may stem from a judge-alignment loop: data selection, hyperparameter tuning, and evaluation all rely on VLM/LLM judgments, leaving no robust held-out evidence that videos are truly more physically correct.","rationale":"The reader's weakest assumption—that caption-level LLM scores of physical fidelity may track wording rather than true physics—is related and important, but it is not the single most load-bearing point. Even if those scores are unreliable, PhysHPO could in principle still improve physical plausibility by using real-world videos as preferred samples in DPO; the selection component might be suboptimal while the headline claim survives. The more decisive vulnerability is that the physical-plausibility measurement itself is generated by automated VLM judges, and those same benchmarks are used to tune the method and select the data pipeline. This creates a potential feedback loop: training signal, model selection, and evaluation all pass through the same family of VLM/LLM statistical heuristics. Therefore the reported gains on Table 2 may reflect alignment with the judge rather than genuine physics. The concrete test I propose—blinded human expert rating on held-out prompts—would directly settle whether the central claim holds. I therefore keep the reader's CONDITIONAL verdict unchanged: the paper is promising but requires independent human or simulation-based validation before the central claim can be accepted at face value.","tokens_in":20669,"tokens_out":10097,"duration_ms":130338,"concrete_test":"Pre-register a blinded human-expert evaluation: sample 100 prompts from VideoPhy/PhyGenBench that were not used for any hyperparameter or data-selection decision; have annotators blind to model label rate concrete physical violations (e.g., object permanence, contact, gravity, conservation) using a fixed rubric and report inter-annotator agreement. If human-rated physical correctness does not improve significantly while VideoPhy/PhyGenBench scores do, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that PhysHPO 'significantly improves physical plausibility'—is supported almost entirely by VideoPhy/PhyGenBench scores that are produced by automated VLM judges. Yet the same kind of LLM/VLM judgment is used inside the training pipeline and for hyperparameter selection. Section 4.2 selects 'good data' by LLM evolving-prompt scoring of captions (s = sd × st × so); Figure 3 chooses the data-selection strategy and dataset size using PhyGenBench; Figure 13 tunes λ, ρ, μ and βerr/βgap on VideoPhy. Table 2 then reports improvements on those same two benchmarks. Appendix C.1 shows that VideoPhy/PhyGenBench scores are thresholded outputs of VideoCon-Physics, an automated VLM judge. If the LLM/VLM rewards captions or videos that merely contain physics-sounding words ('stirring', 'melting', 'collision') rather than verifiable physical behavior, the DPO training signal and the evaluation signal share the same bias: the model can improve benchmark scores by learning to satisfy the judge's textual heuristics without reducing actual physical violations. The user study (15 volunteers, Fig. 5) is the only human check, but no per-dimension numbers, confidence intervals, or significance tests are reported, so it cannot currently break the loop. This is more load-bearing than the caption-scoring assumption alone: even if caption scores are noisy, real-video DPO might still help; but if the evaluation itself is biased and the same benchmarks were used to select the method, the headline improvement is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PhysHPO, a hierarchical cross-modal direct preference optimization framework for improving physical plausibility in text-to-video generation. The method defines four alignment granularities: instance-level overall preference (Eq. 5-6), state-level boundary preference by swapping boundary frames (Eq. 7), motion-level preference using structural information such as optical flow (Sec. 5.3), and semantic-level consistency using VLM-generated captions (Eq. 8). To avoid constructing new datasets, the authors introduce a data-selection pipeline that filters OpenVidHD-0.4M for real-world content, scores captions with LLM-based evolving prompts across 17 physical phenomena, and applies diversity filtering. Experiments on CogVideoX-2B/5B and HunyuanVideo report consistent improvements on VideoPhy, PhyGenBench, and VBench (Table 2 and Table 4), with ablations isolating each loss level (Table 3) and a small user study (Fig. 5). The central claim is that PhysHPO significantly improves physical plausibility and overall video quality of existing advanced models.","tokens_in":21118,"tokens_out":5090,"duration_ms":57963,"significance":"If the reported gains are robust, the paper makes a useful contribution by showing that fine-grained, multi-level DPO, combined with automatic data selection from real-world videos, can improve physics-related generation quality without expensive dataset construction. The framework is clearly structured, the ablation study indicates each loss component contributes, and the additional evaluation on HunyuanVideo provides some cross-model evidence. However, the strength of the claim currently rests on automated benchmarks whose scores are thresholded and were also used for hyperparameter and data-selection choices. The lack of error bars, statistical tests, and an independent human evaluation with quantitative results means the central claim is not yet established at the level claimed. The work is timely and could be publishable after substantial validation improvements.","major_comments":[{"comment":"The reported VideoPhy and PhyGenBench numbers are proportions obtained after binarizing PC and SA at a 0.5 threshold. This creates a nonlinear mapping in which small continuous-score improvements can be amplified into large percentage-point changes. No error bars, seed variance, or significance tests are reported for any row in Table 2 or Table 4. Without these, a gain such as CogVideoX-2B VideoPhy Overall 18.6→25.9 cannot be distinguished from threshold noise. Please report continuous scores, multiple seeds, and confidence intervals or significance tests.","section":"Appendix C.1, Table 2"},{"comment":"Hyperparameters (λ, ρ, μ, βerr/βgap, boundary-frame count, number of generated negatives) are selected using VideoPhy accuracy in Figure 13 and Figure 14, and the data-selection strategy and dataset size are selected using PhyGenBench in Figure 3. Table 2 then reports improvements on those same benchmarks. This is a selection-on-the-test-set loop: the reported gains may partly reflect fitting the benchmark rather than genuine physical-plausibility improvement. A held-out validation set or nested tuning procedure is needed to support the headline claim.","section":"Figure 13, Figure 3, Table 2"},{"comment":"Training-data selection uses LLM/VLM caption scoring (sd×st×so) and VLM-generated semantic captions, while the main evaluations on VideoPhy and PhyGenBench use VideoCon-Physics, an automated VLM judge. If the LLM/VLM rewards physics-sounding wording or superficial cues rather than actual physical behavior, the training signal and the evaluation signal share the same bias, allowing benchmark gains without true physical improvement. The only human check is a 15-volunteer user study with no reported per-dimension scores, confidence intervals, or significance tests. A larger, statistically analyzed human evaluation, or an independently validated judge not used anywhere in training, is necessary to break this loop.","section":"Sec. 4.2, Sec. 5.4, Appendix C.1, Fig. 5"},{"comment":"The state-level non-preferred sample y_l^state is defined as the preferred video with 'the first and last N frames' replaced, but the manuscript never states what they are replaced with. This is the central operation behind LState and Eq. (7), and without specifying the replacement source (e.g., frames from the base model, noise, or another generated video) the method is not reproducible. Please provide the exact construction and justify why it yields a valid non-preferred sample for state consistency.","section":"Sec. 5.2, Eq. (7)"},{"comment":"The motion-level objective is stated as uMotion(x, yw→ymotion_w, yl→ymotion_l, t) 'following Eq. (7)', but the manuscript does not define how the DPO log-ratio is computed on optical flow or other structural representations. Is the structural information used as a conditioning signal, as an additional modal input, or are the log-probabilities evaluated on motion features? This is a central component of the hierarchy and needs an explicit mathematical formulation and implementation description.","section":"Sec. 5.3"}],"minor_comments":[{"comment":"The diversity criterion is described as 'cosine distance d' and a sample is added if d < τ with τ=0.9. For a distance, smaller values mean greater similarity, so d < 0.9 would select near-duplicates rather than diverse samples. If τ is applied to cosine similarity instead, please say so explicitly; otherwise the current text is internally inconsistent.","section":"Sec. 4.3"},{"comment":"The y-axis label reads 'PhysGenBench'; the benchmark name is PhyGenBench.","section":"Fig. 3"},{"comment":"One row label says '+ PhysDPO' but the method is PhysHPO; please correct the typo.","section":"Fig. 15"},{"comment":"The total score s = sd × st × so could be zero if one category is missing, and the scaling or normalization of the three sub-scores is not explained. Please clarify how missing or low sub-scores are handled and how the product is calibrated.","section":"Sec. 4.2"},{"comment":"The text says captions are 'encoded into embeddings' using LLaMA-1 13B, but LLaMA is not an embedding model. Please specify the exact feature extraction procedure (e.g., mean-pooled hidden states) and distance metric.","section":"Sec. 4.3"},{"comment":"The caption says 'Webold the best results' — likely 'We bold the best results'; no bold formatting appears in the rendered table.","section":"Table 2 caption"},{"comment":"The user-study section lists instructions to volunteers but does not provide the actual per-dimension MOS values, sample sizes per condition, or statistical comparisons. Even a small study can be reported with mean, standard deviation, and a paired test.","section":"Appendix C.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a solid engineering core and the hierarchical DPO idea is interesting for the video-generation community. The main risk is the evaluation loop: automatic LLM/VLM judges are used for data selection, hyperparameter tuning, and final evaluation, and the binarized metrics make the reported gains look larger than they may be. This is fixable with held-out validation, continuous scores, and a stronger human study. I do not see a fundamental methodological error that would require rejection, but the current evidence does not support the 'significantly improves' claim without these additions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious referee. The core idea — four-level hierarchical DPO (instance, state, motion, semantic) for video post-training, paired with LLM-driven data selection from existing real-world pools — is a sensible, reusable recipe. The gains on VideoPhy and PhyGenBench are consistent across CogVideoX-2B/5B and HunyuanVideo, and the data efficiency result (21K selected videos beating 433K raw and even WISA-32K) is genuinely interesting. The appendix cost comparison for caption-based vs video-based scoring is practical and useful.\n\nThe soft spots are real but not disqualifying. The headline numbers all come from benchmarks that were also used to tune the method: Figure 13 tunes λ, ρ, μ, βerr/βgap on VideoPhy, and Figure 3 picks data strategy and dataset size on PhyGenBench. That doesn't make the gains fake, but it does mean the reported effect sizes are optimistic and need a held-out evaluation. The judge-alignment concern — LLM scoring for data selection, VideoCon-Physics for evaluation — should be taken seriously; if both track wording rather than real physics, the training signal could partly reward the judge's heuristics. That said, the preferred videos are real-world clips, so the DPO signal is anchored to physical footage, which weakens the worry. Still, the authors should address this explicitly and show that their data-selection scores correlate with human physical-commonsense judgments on a held-out set.\n\nThe loss definitions are under-specified: uState and uMotion are sketched, not derived. There are no error bars or seed variance anywhere, and the user study is reported only as a small MOS summary with no per-dimension numbers or significance tests. The firstness claims are overstated given VistaDPO (same group) and CHip already do hierarchical/cross-modal DPO for video/MM-LLMs; the novelty here is the physics-specific combination and the data-selection angle, which is enough. No code or data is released, making reproduction hard.\n\nBottom line: a solid method paper with a circularity soft spot. Send it to review, but require the authors to address the tuning/evaluation overlap, add ablations on held-out data, and release at least the selected dataset and training config.","headline":"A useful post-training recipe with consistent gains, but the same benchmark used for tuning and evaluation means the headline physical-plausibility claim still needs a held-out check.","tokens_in":21556,"tokens_out":2174,"would_cite":true,"duration_ms":25274,"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":"PhysHPO, a hierarchical cross-modal direct preference optimization, improves the physical plausibility of text-to-video models, lifting CogVideoX-5B's VideoPhy overall score from 39.6 to 45.9.","keywords":["video generation","physical plausibility","direct preference optimization","hierarchical alignment","data selection","text-to-video","post-training","motion modeling"],"falsifier":"Train PhysHPO on a dataset selected by the caption-scoring pipeline but whose captions describe physically impossible events while still scoring high, such as 'water flows uphill into a glass'; if the trained model still improves VideoPhy and PhyGenBench scores, the selection score is not what carries the gain. Conversely, selecting the lowest-scoring captions and observing no drop in physics benchmarks would indicate the LLM scores carry no physical signal.","tokens_in":20602,"feed_emoji":"🎬","tokens_out":11589,"duration_ms":94167,"temperature":0.7,"pith_summary":"This paper sets out to establish that physical plausibility in text-to-video generation can be improved by post-training with a hierarchical cross-modal direct preference optimization (PhysHPO), without building new datasets. Instead of a single video-level preference signal, the method contrasts preferred and non-preferred videos at four granularities: whole-video content, boundary frames, optical-flow motion trajectories, and textual semantics. The preferred videos are chosen by an automated data-selection pipeline that filters existing text-video pools for real-world content, LLM-scored physical fidelity across 17 phenomena, and diversity. On physics benchmarks the approach reports consistent gains, e.g., CogVideoX-5B rises from 39.6 to 45.9 on VideoPhy overall and from 0.45 to 0.61 on PhyGenBench overall, with slight improvements on general-quality VBench.","feed_headline":"CogVideoX-5B jumps 6.3 points on VideoPhy overall","feed_subtitle":"Real-world clips plus four-level preference alignment improve physical plausibility without new datasets.","key_machinery":"The load-bearing object is the hierarchical construction of preference pairs: for each granularity, the paper creates a negative that violates physics specifically at that granularity, then applies a shared DPO-style log-sigmoid contrast. Instance negatives split into 'error' samples (base-model generations semantically aligned but physically flawed) and 'gap' samples (prompt-masked generations missing content); state negatives are made by replacing the first and last $N$ frames of the good video; motion negatives use optical flow rather than pixels; semantic negatives use a VLM-rewritten caption. The second mechanism is the data-selection pipeline, which filters real-world videos with VLMs,","core_discovery":"PhysHPO is a post-training framework that extends Diffusion-DPO's reward reparameterization to four preference objectives. At the instance level, the preferred video is a selected real-world clip and the non-preferred video is either an error-prone generation (the base model's output most visually similar to the preferred) or a gap sample generated from a masked prompt. At the state level, the first and last frames of the preferred video are replaced to build a boundary-violating negative. At the motion level, optical flow is extracted from both videos and the DPO contrast is applied to these structural features. At the semantic level, a vision-language model rewrites the prompt to match the","pith_inferences":["A natural next test is whether caption-level LLM scoring can be replaced by video-based physical verifiers or differentiable physics simulators; if a video-supervised selector works as well or better, the method's dependence on LLM caption judgments would shrink.","The four-granularity decomposition may transfer to other alignment goals in video—safety, style, or identity preservation—where instance-only DPO has been the default.","Because the semantic level optimizes the same video under two captions, PhysHPO effectively performs cross-modal contrastive learning, suggesting a link to vision-language representation learning that the paper does not develop.","The diversity stage uses caption embeddings; using motion or visual embeddings there might yield an even more physically diverse selection, a variant the paper leaves untested."],"forward_implications":["If the claim holds, text-to-video models can be made more physically plausible by post-training on selected existing data, removing the need for expensive manual or synthetic physics datasets.","Data selection becomes a first-class ingredient in video post-training: the reported 21K-video selection outperforms a 433K raw pool and a 32K manually built dataset under identical DPO.","Fine-grained alignment matters: ablations show each of the four levels contributes, and the full hierarchy outperforms instance-only DPO by 4.6 VideoPhy-overall points on CogVideoX-5B.","Physics-aligned models also become more robust on 'impossible prompt' tests (IPV-TXT), suggesting the gains are not a fixed library of physical rules.","General video quality (VBench) is not sacrificed: small improvements are reported alongside the physics gains."],"supporting_citations":[{"why":"Supplies the raw data pool (OpenVidHD-0.4M) from which the physics-focused subset is selected.","marker":"[57]"},{"why":"Defines the 17 physical phenomena and three physics categories used as the physical-fidelity scoring targets, and provides the WISA-32K dataset used as a comparison baseline.","marker":"[75]"},{"why":"Provides the Diffusion-DPO formulation and the vanilla DPO baseline that PhysHPO extends and compares against.","marker":"[72]"},{"why":"Provides the VideoPhy benchmark and VideoCon-Physics judge used to measure physical plausibility.","marker":"[6]"},{"why":"Provides the PhyGenBench benchmark for mechanics, optics, thermal, and materials evaluation.","marker":"[54]"},{"why":"Supplies the Deita-style iterative diversity selection procedure adopted in the data pipeline.","marker":"[53]"},{"why":"Supplies the evolving-prompting strategy used to augment captions for fine-grained physical fidelity scoring.","marker":"[84]"},{"why":"Supplies the Qwen2.5-VL model used for real-world video filtering and for caption rewriting in the semantic-level objective.","marker":"[4]"}],"fun_headline_variants":["PhysHPO: four-level preference tuning lifts VideoPhy by 6.3","Real-world clips plus four-level DPO make videos obey physics","PhysHPO: four-level preference alignment without new datasets","Fine-grained DPO on real clips boosts physical plausibility","PhysHPO: four-level DPO uses existing data for physics"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The paper assumes that LLM scores assigned to captions—after evolving prompting over 17 physical phenomena—measure the physical fidelity of the corresponding videos; if these scores reward wording rather than actual physics, then training the model to prefer high-scoring captions will not transfer to benchmark physical correctness.","fun_headline_variants_meta":{"raw":{"variants":["PhysHPO: four-level preference tuning lifts VideoPhy by 6.3","Real-world clips plus four-level DPO make videos obey physics","PhysHPO: four-level preference alignment without new datasets","Fine-grained DPO on real clips boosts physical plausibility","PhysHPO: four-level DPO uses existing data for physics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001061,"raw_usage":{"total_tokens":4294,"prompt_tokens":760,"completion_tokens":3534,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":3444}},"tokens_in":504,"tokens_out":3534,"duration_ms":26090,"temperature":1.0,"reasoning_tokens":3444,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:13:06.587451+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PhysHPO on a dataset selected by the caption-scoring pipeline but whose captions describe physically impossible events while still scoring high, such as 'water flows uphill into a glass'; if the trained model still improves VideoPhy and PhyGenBench scores, the selection score is not what carries the gain. Conversely, selecting the lowest-scoring captions and observing no drop in physics benchmarks would indicate the LLM scores carry no physical signal.","supporting_citations":[],"review_version":1}