{"id":"7e3009ef-0a65-4de3-9082-e051ff6ddf09","arxiv_id":"2506.14168","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A decoder-only autoregressive video model with continuous tokens, frame-wise causal attention, and a next-frame diffusion loss reports a higher VBench-I2V score than Cosmos I2V with a much smaller model and dataset.","lead":"VideoMAR is a video generation model that predicts each frame using an autoregressive transformer over continuous tokens, masking part of the current frame and learning with a diffusion-style loss. The paper reports higher VBench-I2V scores than the Cosmos baseline while training on far less data and compute, but the comparison uses a private dataset and the Cosmos tokenizer itself.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed SOTA margin may be an artifact of the progressive-temperature schedule: Table 5 shows temperature 0.9 cuts Dynamic Degree from 24.80 to 9.87, and Cosmos is not evaluated under the same schedule.","rationale":"I read the central claim as an empirical efficiency/quality statement: a 1.4B continuous-token masked autoregressive model trained on 0.5M pairs beats the Cosmos autoregressive baselines on VBench-I2V with far less compute/data. The most load-bearing condition is that the VBench advantage reflects a genuine modeling advantage, not an artifact of evaluation or hyperparameters. The tokenizer concern raised by the reader is plausible but less decisive: since the baseline is the Cosmos platform that provides Cosmos-Tokenizer, the tokenizer is likely a shared pretrained component, and excluding it from both sides is standard practice when counting image-to-video training data and generation-backbone compute. The temperature issue, by contrast, is fully internal to the paper. Table 5 shows the aggregate gain from the proposed inference schedule coincides with a drop in Dynamic Degree from 24.80 to 9.87, while the selected progressive setting yields 10.98 versus Cosmos-13B's 18.70. The paper's dismissal of this drop as failure cases is not supported by the monotonic sweep. Without a temperature-matched baseline run and error bars, the 0.60-point gap in Table 3 is too fragile to support the 'surpasses SOTA' headline. The issue is addressable by a controlled re-evaluation, so the appropriate verdict remains CONDITIONAL rather than REJECT or ACCEPT.","tokens_in":13915,"tokens_out":12511,"duration_ms":133060,"concrete_test":"Re-run the Table 3 comparison with temperature matched: evaluate Cosmos I2V (5B or 13B) under the same progressive temperature schedule used by VideoMAR (or its discrete-token analogue, lower sampling temperature for later frames), and evaluate VideoMAR-stage2 at fixed temperature 1.0 and fixed 0.9, on identical VBench-I2V prompts/samples with at least 5 seeds. If Cosmos's total score reaches or exceeds 84.82, or if VideoMAR at temperature 1.0 falls below Cosmos's 84.22, the headline superiority is an inference-hyperparameter artifact, not a robust modeling advantage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (surpassing Cosmos I2V on VBench-I2V) rests on Table 3's 0.60-point gap (84.82 vs. 84.22 for VideoMAR-stage2 vs. Cosmos-13B), but the appendix's own temperature ablation shows this gap is driven by an inference-time hyperparameter that suppresses motion. Table 5 reports Dynamic Degree falling from 24.80 at temperature 1.00 to 9.87 at temperature 0.90, while Subject Consistency jumps from 89.70 to 97.17 and Total Score from 82.19 to 84.82. The chosen 'Ours' progressive schedule yields Dynamic Degree 10.98, less than half of Cosmos-13B's 18.70. Thus the aggregate improvement largely reflects making videos more static, not better temporal modeling; Appendix B's claim that the drop is 'mainly attributed to failure cases' is contradicted by the monotonic 15-point decline in that one dimension across the sweep. Because Cosmos was evaluated with its official recommended parameters (Section 5.2) rather than with the same low-temperature/progressive schedule, the headline superiority is not a controlled comparison. This is load-bearing: if Cosmos is re-run under the same temperature schedule, or if VideoMAR is run at temperature 1.0, the order could invert. The absence of error bars makes the 0.60-point margin additionally fragile.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VideoMAR, a decoder-only masked autoregressive image-to-video model operating on continuous tokens from a pretrained video VAE. The method combines frame-wise causal attention with intra-frame bidirectional masked generation, a next-frame diffusion loss, a short-to-long temporal curriculum, two-stage progressive-resolution training, and a progressive temperature schedule at inference. The central empirical claim is that VideoMAR surpasses the previous autoregressive state of the art, Cosmos I2V, on the VBench-I2V benchmark while using 1.4B parameters, 0.5M training pairs, and 64 H20 GPUs, compared to Cosmos's 5B/13B parameters, 100M pairs, and much larger compute. Additional claims include substantial inference speedup through KV cache and spatial parallel generation, and training-free spatial and temporal extrapolation via 3D-RoPE.","tokens_in":14225,"tokens_out":3271,"duration_ms":34046,"significance":"If the headline comparison were properly controlled, the result would be significant: it would show that a compact continuous-token masked autoregressive model can match or beat a much larger discrete-token autoregressive model on a standard benchmark, with far fewer resources, and can extrapolate to unseen resolutions and durations. The framework is clean, the ablations in Table 4 are internally consistent, and the reported inference-time speedups (e.g., 1941s to 134s in Table 2) are concrete and credible. The qualitative extrapolation examples in Figure 4 and Appendix D are useful demonstrations of a desirable capability. However, the paper's central quantitative claim is not yet robustly supported because the comparison with Cosmos is confounded by an inference-time temperature schedule, the resource accounting excludes the pretrained tokenizer, and the reported VBench scores lack error bars. These issues make the headline margin of 0.60 points fragile and potentially an artifact of the evaluation procedure.","major_comments":[{"comment":"The headline comparison with Cosmos is confounded by the progressive temperature schedule. Table 5 shows that lowering the temperature from 1.00 to 0.90 increases the Total Score from 82.19 to 84.82 while reducing Dynamic Degree from 24.80 to 9.87. VideoMAR-stage2's reported 84.82 in Table 3 is obtained under the progressive schedule that yields Dynamic Degree 10.98, less than half of Cosmos-13B's 18.70. The appendix's claim that the dynamic-degree drop is 'mainly attributed to failure cases' is contradicted by the monotonic decline in that dimension across the sweep. Because Cosmos is evaluated with its own recommended parameters rather than under the same temperature schedule, the reported 0.60-point advantage does not isolate the model's temporal modeling ability. The authors should report VideoMAR at temperature 1.0, or re-run Cosmos under the progressive schedule, or otherwise provide a controlled comparison that separates the model from the inference-time hyperparameter.","section":"Section 5.2 / Table 3 / Appendix B (Table 5)"},{"comment":"The resource-efficiency claim is incomplete because the visual tokenizer, Cosmos-Tokenizer, is a pretrained component developed by the Cosmos team. The statement 'All the weights are trained from scratch' in Section 5.1 can only refer to the VideoMAR backbone; the tokenizer's pretraining data and compute are external to the stated 0.5M video-text pairs and 64 H20 GPUs. Consequently, the comparison '1.4B parameters, 0.5M data, 64 GPUs vs. 5B/13B parameters, 100M data, 10000 GPUs' is not apples-to-apples unless the tokenizer is explicitly excluded or its cost is also counted for both sides. The authors should clarify this boundary or restate the efficiency claims accordingly.","section":"Section 5.1 / Abstract and Section 5.2"},{"comment":"There is an internal inconsistency in the reported headline number for VideoMAR-stage2: the abstract and Section 5.2 state 84.51, while Table 3 lists 84.82. Since the entire paper's central claim hinges on this number, the discrepancy must be resolved and the correct value used consistently throughout.","section":"Abstract / Section 5.2 / Table 3"},{"comment":"The evaluation lacks any measure of variance. The paper states that 5 samples per prompt are generated for VBench-I2V, but no error bars, confidence intervals, or per-seed deviations are reported. Given that the claimed advantage over Cosmos-13B is 0.60 points (84.82 vs. 84.22), the result may be within sampling noise. The authors should provide standard deviations, multiple evaluation seeds, or a significance test, or at least discuss the expected variability of VBench scores.","section":"Section 5.1 / Section 5.2"}],"minor_comments":[{"comment":"The title contains a typo: 'Generatio' should be 'Generation'.","section":"Title"},{"comment":"Several column headers ('Moti.', 'I2V Subj.', 'Back. Cons.', etc.) are abbreviated without being expanded in the caption; please define all abbreviations for readability.","section":"Table 3"},{"comment":"The text says the baseline 'Total mask' uses 'w/ Temperature strategy', but the first row of Table 4 shows Temperature as ✗. This is confusing; please correct either the text or the table so the baseline configuration matches the table.","section":"Section 5.4"},{"comment":"The progressive temperature schedule is described in Section 4.3 as varying 'from 1 to 0.9 across frames', while Appendix B specifies the exponential form 0.9 + 10^-(t+1). These formulations are not obviously equivalent; please state the exact schedule used in the main experiments.","section":"Section 4.3 and Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper's central claim is the SOTA comparison against Cosmos. The temperature confound is the most serious issue; it is fixable but requires re-running the comparison under controlled settings. The tokenizer-accounting issue is also significant for the efficiency narrative. If the authors address these points thoroughly, the paper could become a solid contribution, but in its current form the headline result is not reliable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The method is a sensible engineering combination: frame-wise causal attention plus a next-frame diffusion loss on continuous tokens, with 3D-RoPE for extrapolation and a curriculum for long sequences. The internal ablations in Table 4 are consistent and show each component helps. The efficiency results—spatial parallel generation with temporal KV cache—are plausible and worth reporting.\n\nBut the headline claim does not hold up. First, the numbers don't match: the abstract and Section 5.2 say 84.51, Table 3 says 84.82. That is fixable, but it does not inspire confidence. More importantly, the comparison with Cosmos is not controlled. Appendix Table 5 shows that lowering temperature from 1.0 to 0.9 raises Total Score from 82.19 to 84.82 while Dynamic Degree drops from 24.80 to 9.87. The chosen 'Ours' schedule lands at 10.98. Cosmos is evaluated with its own recommended parameters, not the same progressive schedule, so the 0.6-point aggregate margin over Cosmos 13B may simply reflect making videos more static. The appendix's claim that the dynamic-degree drop is mostly due to failure cases is hard to square with the monotonic 15-point decline across the sweep. Add the absence of error bars on 5 samples per prompt, and the SOTA claim is fragile.\n\nThere is also the tokenizer issue: using Cosmos-Tokenizer as the visual encoder means part of the video prior comes from large-scale external pretraining. That does not invalidate the method, but it undercuts the '0.5% data / 0.2% GPU' comparison as stated.\n\nWho should read this? People working on autoregressive video generation and continuous-token modeling will find the architecture and training recipe worth a look. The extrapolation results are interesting. But the evaluation needs real work before the competitive claim can be believed.\n\nRecommendation: send it to peer review—it deserves referee time—but with a clear expectation of major revision. Authors should fix the number discrepancy, report error bars, rerun Cosmos under the same temperature schedule or show the comparison at multiple temperatures, and discuss the dynamic-degree trade-off honestly.","headline":"A plausible continuous-token AR video model with honest internal ablations, but the claimed SOTA over Cosmos is likely an artifact of the progressive temperature schedule and the comparison is not controlled.","tokens_in":14755,"tokens_out":2869,"would_cite":false,"duration_ms":29779,"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 a 1.4B-parameter decoder-only autoregressive video model with continuous tokens outperforms a 13B discrete-token baseline on VBench-I2V while using 0.5% of the training data and 0.2% of the GPU budget.","keywords":["autoregressive video generation","continuous tokens","masked autoregressive models","image-to-video generation","next-frame diffusion loss","3D rotary position embedding","temporal extrapolation","VBench-I2V"],"falsifier":"Retrain VideoMAR from scratch with a video tokenizer trained only on the same 0.5M video-text pairs, with no external pretraining, and measure the VBench-I2V score; if the score falls far below 84.82 or the data and compute requirements jump, then the pretrained tokenizer rather than the autoregressive paradigm is carrying much of the reported result.","tokens_in":13711,"feed_emoji":"🎬","tokens_out":9057,"duration_ms":79790,"temperature":0.7,"pith_summary":"VideoMAR is a decoder-only autoregressive image-to-video model that generates continuous tokens frame by frame while using masked prediction inside each frame. The paper's central claim is that temporal causality (each frame attends only to earlier frames and to visible tokens in the current frame) combined with spatial bi-directionality is the right first principle for video autoregressive models, and that this design can outperform the previous state-of-the-art autoregressive image-to-video baseline on the VBench-I2V benchmark with far fewer parameters, far less training data, and far less GPU compute. The authors attribute the result to a next-frame diffusion loss that trains masked tokens without breaking causality, short-to-long temporal curriculum learning, progressive-resolution training, and a progressive temperature schedule that suppresses error accumulation in late frames. They also report training-free spatial and temporal extrapolation through 3D rotary position embeddings, yielding longer and higher-resolution videos than the training grid.","feed_headline":"Autoregressive video generation beats a 13B rival on 0.5% of the data","feed_subtitle":"VideoMAR, a 1.4B continuous-token masked autoregressive model trained on 0.5M pairs, tops Cosmos I2V on VBench-I2V.","key_machinery":"The central object is the frame-wise causal attention mask paired with the next-frame diffusion loss. The mask lets a token in frame t attend to all tokens in earlier frames and to the visible tokens in frame t, while leaving attention bidirectional within the frame, so temporal generation stays autoregressive and spatial generation stays parallel. The next-frame diffusion loss randomly masks part of a chosen frame, masks all later frames, leaves earlier frames complete, and applies the diffusion loss only to the chosen frame's masked tokens, which removes the training-inference gap that earlier mask-based video autoregressive methods face. Supporting mechanisms include a pretrained continuous video tokenizer, 3D rotary position embeddings for extrapolation, temporal short-to-long curriculum learning, a two-stage progressive-resolution schedule, and a progressive temperature schedule that lowers temperature for later frames.","core_discovery":"VideoMAR establishes, on the paper's own terms, that masked autoregressive generation in continuous space is a viable and resource-efficient paradigm for image-to-video generation when applied temporally frame-by-frame and spatially within each frame. Concretely, a 1.4B-parameter decoder-only transformer trained on 0.5M video-text pairs with 64 GPUs attains a VBench-I2V total score of 84.82 in the second training stage, surpassing the 84.22 of the 13B discrete-token baseline and the 84.16 of the 5B baseline, both trained on 100M pairs. The design keeps complete previous frames as context, applies a frame-wise causal attention mask, and optimizes a token-wise diffusion loss only on the masked tokens of a randomly selected frame, with all later frames fully masked during that training step. The paper further claims that the same checkpoint generates arbitrary combinations of longer duration and higher resolution through 3D-RoPE, and that spatial masked parallel generation combined with temporal KV cache reduces inference time by more than 10 times relative to raster-scan next-token prediction.","pith_inferences":["The headline efficiency comparison implicitly counts the pretrained continuous video tokenizer as free; if that tokenizer's external pretraining data and compute were included, VideoMAR's resource advantage over the baseline would shrink, and the comparison would be less apples-to-apples.","The same architecture is a natural candidate for unifying text-to-image, text-to-video, image-to-video, video-to-video, and video editing in one decoder-only network, which the paper lists as future work.","A direct testable extension is to train VideoMAR with a tokenizer trained only on the same 0.5M video-text pairs; if performance collapses, the tokenizer's pretrained visual prior is the hidden large input rather than the autoregressive design.","Because constant low temperature improves quality but lowers the dynamic-degree score, an adaptive per-frame temperature chosen by a lightweight quality or motion predictor could outperform the fixed exponential schedule the paper adopts."],"forward_implications":["If the central claim is correct, discrete tokenization is not required for high-quality autoregressive video; continuous tokens with a masked diffusion loss can match or beat much larger discrete-token autoregressive models.","If the reported efficiency holds, video autoregressive training becomes feasible with small data and compute budgets, lowering the entry cost for this paradigm.","If the 3D-RoPE extrapolation result holds, a single fixed-resolution checkpoint can generate videos at arbitrary aspect ratios and durations without chunk-wise splitting.","If the progressive temperature strategy works as reported, late-frame collapse from exposure bias can be suppressed by a simple inference-time schedule that lowers temperature for later frames.","If the frame-wise causal mask closes the training-inference gap, mask-based autoregressive video models can retain both frame-by-frame generation and spatial parallel decoding with KV cache acceleration."],"supporting_citations":[{"why":"Supplies the pretrained continuous video tokenizer and the Cosmos autoregressive image-to-video baseline that VideoMAR must beat, along with the data and compute comparison.","marker":"[1]"},{"why":"Supplies the VBench-I2V benchmark and the nine evaluation dimensions on which VideoMAR reports scores.","marker":"[15]"},{"why":"Supplies the continuous-token masked autoregressive image-generation paradigm and the token-wise diffusion loss that VideoMAR adapts to video.","marker":"[18]"},{"why":"Supplies rotary position embeddings, which VideoMAR extends to 3D to obtain spatial and temporal extrapolation.","marker":"[25]"},{"why":"Represents the prior masked autoregressive video method that doubles sequence length with complete teacher forcing, from which VideoMAR's next-frame loss differs.","marker":"[37]"},{"why":"Represents the prior autoregressive video method without vector quantization that separates temporal and spatial modeling, which VideoMAR contrasts with its unified frame-wise causal design.","marker":"[8]"}],"fun_headline_variants":["VideoMAR: 1.4B model beats 13B on 0.5% of the data","1.4B continuous-token AR video model outdoes 13B on VBench-I2V","VideoMAR: 1.4B, 0.5M pairs, 64 GPUs, outperforms Cosmos","VideoMAR: continuous-token AR video generation, 1.4B, beats Cosmos","VideoMAR: 1.4B AR model, 0.5% data, surpasses Cosmos I2V"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The efficiency claim presupposes that the pretrained continuous video tokenizer, which was trained on large-scale external video data, does not count as part of VideoMAR's training cost; if it is counted, the comparison with the baseline is no longer apples-to-apples.","fun_headline_variants_meta":{"raw":{"variants":["VideoMAR: 1.4B model beats 13B on 0.5% of the data","1.4B continuous-token AR video model outdoes 13B on VBench-I2V","VideoMAR: 1.4B, 0.5M pairs, 64 GPUs, outperforms Cosmos","VideoMAR: continuous-token AR video generation, 1.4B, beats Cosmos","VideoMAR: 1.4B AR model, 0.5% data, surpasses Cosmos I2V"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001188,"raw_usage":{"total_tokens":5260,"prompt_tokens":1031,"completion_tokens":4229,"prompt_tokens_details":{"cached_tokens":1024},"prompt_cache_hit_tokens":1024,"prompt_cache_miss_tokens":7,"completion_tokens_details":{"reasoning_tokens":4091}},"tokens_in":7,"tokens_out":4229,"duration_ms":310979,"temperature":1.0,"reasoning_tokens":4091,"cache_read_input_tokens":1024,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:17:52.326894+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain VideoMAR from scratch with a video tokenizer trained only on the same 0.5M video-text pairs, with no external pretraining, and measure the VBench-I2V score; if the score falls far below 84.82 or the data and compute requirements jump, then the pretrained tokenizer rather than the autoregressive paradigm is carrying much of the reported result.","supporting_citations":[],"review_version":1}