{"id":"b207a62d-0d12-4246-830d-ea6225aa8520","arxiv_id":"2506.07280","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Frozen CogVideoX1.5, adapted with LoRA on 3 to 30 input-output videos, performs segmentation, pose estimation, and abstract reasoning (ARC-AGI 16.75%) with modest but real generalization.","lead":"Video diffusion models can be repurposed as few-shot visual task solvers by training lightweight LoRA adapters on short input-output videos, while keeping the base model frozen. The approach handles segmentation, pose estimation, style transfer, and even abstract reasoning puzzles, suggesting video generation may be a general interface for vision.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central interpretation rests on the unverified claim that CogVideoX pretraining excludes ARC-like data; without a contamination audit, the 16.75% ARC accuracy does not establish emergent latent reasoning.","rationale":"The reader's conditional verdict already identifies the data-contamination premise as the weakest assumption, and my independent reading lands on the same point. The paper is genuinely interesting: the framework is simple, the simple-task results (geometric transforms, style transfer, segmentation, pose) are credible, and the authors are transparent about failure modes and shortcuts in Appendix C. Those strengths do not, however, secure the headline interpretation. The phrase \"has not been pre-trained on any data resembling ARC-AGI\" is an unsupported factual assertion about a proprietary, internet-scale training corpus. Since the ARC-AGI result is the paper's flagship reasoning evidence and its stated contribution, the conclusion that VDMs possess latent visual understanding would be substantially weakened if that premise failed. A targeted held-out experiment with novel ARC-style tasks would settle the question without relying on access to CogVideoX's full training data. I therefore keep the reader's CONDITIONAL verdict: the method and lower-level results are credible, but the abstract-reasoning claim should not be accepted until the contamination concern is addressed.","tokens_in":15193,"tokens_out":4328,"duration_ms":62216,"concrete_test":"Construct a held-out suite of ARC-style reasoning tasks whose object shapes, colors, grid sizes, and transformation rules are verifiably absent from both ARC-AGI and CogVideoX's likely training distribution (e.g., newly generated puzzles with a novel visual vocabulary, plus a nearest-neighbor check against any accessible training-data samples). Fine-tune the same LoRA pipeline on this suite and evaluate with the same protocol. If accuracy on the novel tasks is statistically indistinguishable from chance while ARC accuracy stays at 16.75%, pretraining overlap remains a live explanation; if novel-task accuracy is far above chance, the latent-reasoning interpretation is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3.1 states that CogVideoX1.5 \"has not been pre-trained on any data resembling ARC-AGI,\" but no evidence or audit is provided. CogVideoX is trained on large-scale internet video of unknown composition, and ARC-AGI inputs are simple colored grids — exactly the kind of visual content that can appear in screenshots, animations, coding tutorials, or synthetic video common on the web. If ARC-like grid puzzles or visually similar reasoning tasks were present in the pretraining corpus, the reported 16.75% accuracy could be explained by retrieval or distribution overlap rather than by a general latent visual understanding. This matters because contribution 3 (\"first to successfully apply VDMs to ARC-AGI\") and the conclusion that \"VDMs possess latent visual understanding that can be unlocked through minimal supervision\" depend on the reasoning results being genuinely novel relative to pretraining. The claim is load-bearing not because the authors are dishonest, but because it is an unverified empirical premise stated as fact, and it is testable.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a few-shot fine-tuning framework that repurposes a frozen image-to-video diffusion model (CogVideoX1.5, with LTX-Video for comparison) for diverse vision tasks. Each task is encoded as a set of input-output image pairs rendered as short transition videos, and LoRA adapters are trained on these videos while the base model remains frozen. At inference the final frame of the generated video is taken as the prediction. Experiments cover geometric transformations, style transfer, jigsaw puzzles, colorization, inpainting, binary segmentation, pose estimation, TinyImageNet grid classification, and the ARC-AGI and ConceptARC reasoning benchmarks. The authors report strong generalization from 3 to 30 training examples and argue that this success reveals latent visual understanding in video diffusion models, positioning VDMs as adaptable visual learners and potential vision foundation models.","tokens_in":15399,"tokens_out":2792,"duration_ms":35455,"significance":"If the central claim holds, the framework is a valuable probe of the representations learned by video diffusion models and a promising step toward using VDMs as generalist vision backbones. The paper has clear strengths: the task-to-video reformulation is simple and unified, the LoRA-based adaptation is parameter-efficient, the ablation study over interpolation methods, LoRA modules, and rank is useful, and the appendix provides detailed hyperparameters, compute costs, and an unusually candid discussion of failure modes such as underspecification and color shifts. The comparison across two VDMs and the inclusion of public reasoning benchmarks are also positive. However, the central interpretation depends on several load-bearing empirical premises that are not yet adequately supported: the absence of ARC-like data in CogVideoX pretraining, the validity of the proposed Match Rate metric, and the fairness of the checkpoint-selection procedure for ARC and ConceptARC. These issues do not invalidate the framework but they do affect whether the paper can support its broad claims about emergent latent understanding.","major_comments":[{"comment":"The claim that CogVideoX1.5 \"has not been pre-trained on any data resembling ARC-AGI\" is stated without evidence, yet it is load-bearing for contribution 3 and for the conclusion that VDMs possess latent visual understanding. CogVideoX is trained on large-scale, undisclosed internet video, and ARC-AGI's simple colored grids are exactly the kind of visual content that can appear in screenshots, animations, or synthetic video. If ARC-like or visually similar reasoning tasks were present in pretraining, the 16.75% accuracy could reflect distribution overlap rather than a general latent reasoning ability. Please provide a contamination audit, report any known pretraining-data composition, or reframe the claim to avoid asserting absence of overlap; a concrete control would be to test on newly generated ARC-like tasks that cannot have been in the pretraining corpus.","section":"Section 4.3.1"},{"comment":"The ARC-AGI and ConceptARC evaluation selects validation predictions from checkpoints at which the model's prediction on the training set is correct, and otherwise uses the final available predictions. This protocol uses training-set labels to choose among checkpoints, which can inflate the reported accuracy relative to a fixed evaluation schedule and makes the numbers in Table 3 and Figure 9 difficult to interpret as a genuine few-shot generalization result. Please report results under a fixed checkpoint or averaged over multiple checkpoints, and quantify the variance induced by the selection procedure.","section":"Appendix A.4.2"},{"comment":"The Match Rate metric is introduced as \"well correlated with pose estimation quality,\" but no evidence for this correlation is provided, and the metric explicitly does not penalize spurious or excessive predictions. The threshold of 1.5 times the average inter-head distance, with a default of 20 pixels when no heads are available, is ad hoc. Because Match Rate is used in Tables 1 and 2 to support conclusions about pose estimation performance, please validate it against a standard metric such as PCK or OKS on a subset of runs, and either incorporate a false-positive penalty or justify why the visual verification described in the appendix is sufficient.","section":"Appendix A.4.1"},{"comment":"The conclusion that VDMs \"possess latent visual understanding that can be unlocked through minimal supervision\" goes beyond what the current experiments demonstrate. Few-shot success on held-out inputs shows that the LoRA adapters can learn a mapping from a handful of examples, but it does not by itself establish that the knowledge was already latent in the frozen VDM rather than learned by the adapters from the provided examples. To support the latent-knowledge interpretation, please include a baseline in which LoRA is trained on the same few examples but with a model that lacks the VDM's pretraining (for example, a randomly initialized or weakly pretrained encoder), or an analysis of how performance changes as pretraining quality varies.","section":"Section 6"}],"minor_comments":[{"comment":"The table footnote states that some rows are duplicated because they correspond to the same run, but the duplicated entries are not visually marked; please indicate them with a symbol or a note in the table itself to avoid confusion.","section":"Table 1"},{"comment":"The sentence \"we perform several runs and report the average results\" does not state the number of runs or the variance; please include the number of seeds and standard deviations in Table 1 and Table 2, at least for the n=3 and n=5 settings where variability is expected to be highest.","section":"Section 4.2"},{"comment":"The statement that only one ARC-AGI case required manual grid-size handling and that \"visual inspection confirmed\" correctness is not verifiable from the text; please provide a reproducible rule for such cases or exclude them from the reported accuracy.","section":"Appendix A.4.2"},{"comment":"The inference procedure in step 3 is written as \"Apply denoising for t=T,...,1 to vt\", but the notation for the noise scheduling is inconsistent with Eq. (1); please align the notation so that v_T denotes the fully noised sample and v_0 the denoised output.","section":"Section 3.3"},{"comment":"Reference [16] is listed as \"Yoav HaCohen et al.\" while the in-text citation uses \"LTX-Video\"; please ensure the reference list entry matches the author order of the actual paper.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the framework is interesting, but the ARC reasoning claim rests on an unverified no-contamination premise for a proprietary model, and the evaluation protocol for ARC/ConceptARC uses a training-set-based checkpoint selection that may inflate results. The Match Rate metric also needs validation. These are fixable with additional experiments or with careful reframing of the claims, so I recommend major revision rather than rejection. Given the high compute cost, a reasonable path would be to validate the metric and the checkpoint procedure on a subset of tasks, and to replace the strong \"no ARC-like data\" assertion with a documented statement about what is and is not known about the pretraining data."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The framework is simple and the appendix is unusually candid, but the headline claim about emergent latent reasoning is staked on the ARC-AGI number, and that number is the softest part of the paper.\n\nWhat's new: the specific recipe - LoRA adapters on a frozen image-to-video model, tasks encoded as transition videos - is a clean way to turn an I2V generator into a few-shot task learner. The experiments span geometric transforms, style transfer, jigsaw, colorization, inpainting, segmentation, pose, TinyImageNet grid classification, and two reasoning benchmarks. That breadth is real. The ablations on interpolation method, LoRA target modules, and rank are useful, and the comparison with LTX-Video supports a scaling story. The paper is also honest about failure modes: the appendix documents shortcut learning, color shifts, depth decoding errors, and the ambiguity of one-shot tasks.\n\nWhere it goes soft: the central interpretation leans on ARC-AGI, and the ARC results sit on two unverified choices. First, Section 4.3.1 states as fact that CogVideoX1.5 has not been pre-trained on anything resembling ARC-AGI. There is no audit. ARC inputs are colored grids, exactly the kind of content that appears in screenshots, coding tutorials, and synthetic video on the web. If any of that was in training, 16.75% could be distribution overlap rather than latent reasoning. Second, Appendix A.4.2 describes checkpoint selection: the authors only use a validation prediction if the same checkpoint got the training set right. That's a form of training-set oracle, and it can inflate accuracy. The Match Rate metric is a smaller issue - it's a proxy and the paper says so - but it ignores false positives, so the pose numbers should be read with caution.\n\nI don't think these flaws sink the simpler contributions. The segmentation and transformation results are plausible. But the conclusion \"VDMs possess latent visual understanding that can be unlocked through minimal supervision\" is stronger than the evidence. The ARC result needs a contamination audit and a conservative evaluation protocol before it can carry that weight.\n\nOverall: this deserves a serious referee. The method is worth engaging, the writing is clear, and the limitations section is a model of honesty. I'd send it to peer review with the understanding that the reasoning claims need substantial revision. I'd probably cite it for the framework. I'd bring it to reading group, because it's a useful discussion of what few-shot adaptation actually demonstrates.","headline":"A solid few-shot adapter paper whose headline 'emergent reasoning' claim is weaker than the paper's own best evidence; the ARC number is not load-bearing until contamination and checkpoint selection are addressed.","tokens_in":15920,"tokens_out":2653,"would_cite":true,"duration_ms":31003,"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":"A frozen video diffusion model learns new vision tasks from 3–30 examples, including abstract reasoning on ARC-AGI.","keywords":["video diffusion models","few-shot learning","LoRA fine-tuning","visual reasoning","ARC-AGI","generalist vision models","transfer learning","latent understanding"],"falsifier":"Audit CogVideoX1.5's training data for ARC-like grid puzzles, or run the identical LoRA transition-video procedure on a VDM pretrained only on natural videos: if that model still scores near 16.75% on ARC-AGI, the result does not depend on grid-like exposure, while if it collapses to near zero the paper's key premise fails.","tokens_in":15027,"feed_emoji":"🎬","tokens_out":4024,"duration_ms":45061,"temperature":0.7,"pith_summary":"The paper argues that video diffusion models, trained only to generate coherent video, internalize structured visual knowledge that can be re-directed to new tasks with minimal supervision. It proposes turning each input–output image pair into a short transition video and fine-tuning LoRA adapters on a frozen image-to-video model. With only 3 to 30 such examples, the model generalizes to unseen inputs across segmentation, pose estimation, style transfer, and abstract reasoning. The authors read this as evidence that VDMs possess latent visual understanding rather than being mere generative engines.","feed_headline":"Frozen video model learns new visual tasks from a few examples","feed_subtitle":"LoRA tuning on 3–30 transition videos unlocks segmentation, pose, and abstract reasoning in CogVideoX1.5.","key_machinery":"The load-bearing mechanism is the transition-video representation: each task pair $(x_i, y_i)$ is turned into an $F$-frame sequence $v = [x_i, v_2, \\dots, v_{F-1}, y_i]$ by an interpolation function $\\varphi$, and a frozen image-to-video diffusion model is adapted by LoRA on these sequences. At inference the model is conditioned on a new input $x$ and the final generated frame $\\hat{y}$ is taken as the prediction. The interpolation choice matters: a discrete sharp transition is used for ARC-AGI reasoning tasks, while quadratic interpolation works best for segmentation and pose.","core_discovery":"The central claim is that a pre-trained video diffusion model, CogVideoX1.5, can be repurposed for a wide range of visual tasks by fine-tuning lightweight LoRA weights on a handful of input–output pairs rendered as transition videos. On the public ARC-AGI evaluation set, this procedure reaches 16.75% accuracy under the two-attempt protocol, a level the paper compares favorably against several large language models, and on ConceptARC it averages 33.12% with three attempts per test input. The authors conclude that successful adaptation from so few examples reveals latent visual understanding already present in the pretrained VDM, and that video generation provides a powerful inductive bias for unifying perception, generation, and reasoning.","pith_inferences":["If the latent-understanding interpretation is correct, the interpolation trajectory itself acts as a task prior: smooth trajectories for continuous transformations and sharp transitions for discrete reasoning, suggesting that learning the interpolation jointly with the task could improve results.","The ARC-AGI result would be sharpened by an audit of CogVideoX1.5's pretraining data; without that audit, the 16.75% accuracy could reflect distribution overlap with grid-like puzzles rather than general visual reasoning.","Composable LoRA modules, which the paper names as future work, would let a frozen VDM stack skills such as segmentation followed by style transfer without retraining.","A direct comparison against an image diffusion model trained on the same transition videos would isolate whether temporal coherence, rather than the generative prior alone, supplies the few-shot generalization."],"forward_implications":["A single frozen video diffusion model can be repurposed for many visual tasks by swapping lightweight LoRA weights instead of retraining the full model per task.","The few-shot regime of 3–30 examples suggests the pretrained prior does most of the work, which is evidence for transferable visual knowledge inside VDMs.","Scaling helps: CogVideoX1.5 (5B) outperforms LTX-Video (2B) at every training-set size, paralleling the scaling behavior seen in language models.","Video-based reasoning is a viable new route for ARC-style abstract visual reasoning, distinct from direct LLM prompting on grids.","Video diffusion models could serve as the backbone of future visual foundation models that adapt rapidly to novel tasks."],"supporting_citations":[{"why":"CogVideoX1.5 is the pretrained image-to-video diffusion model that all main experiments adapt.","marker":"[17]"},{"why":"LoRA is the parameter-efficient adaptation method used to keep the base model frozen.","marker":"[28]"},{"why":"Denoising diffusion probabilistic models supply the training objective and sampling procedure for the VDM.","marker":"[26]"},{"why":"The ARC-AGI benchmark defines the abstract reasoning task and evaluation protocol.","marker":"[6]"},{"why":"The ARC Prize evaluation protocol fixes the two-attempt limit and exact-match scoring used for ARC results.","marker":"[7]"},{"why":"RealGeneral is the closest prior work fine-tuning CogVideoX1.5 for multi-task visual learning, and the paper contrasts its full fine-tuning with their few-shot LoRA approach.","marker":"[33]"},{"why":"ConceptARC provides the controlled concept-level reasoning benchmark and its evaluation methodology.","marker":"[37]"},{"why":"DreamBooth supplies the dataset for geometric transformation and style transfer experiments.","marker":"[47]"},{"why":"LTX-Video is the second VDM used in the comparison showing that generalization scales with model size.","marker":"[16]"}],"fun_headline_variants":["Video model masters new tasks from just a few clips","LoRA on video transitions unlocks diverse visual tasks","Frozen video diffusion model learns from handful of examples","From video generation to few-shot vision: CogVideoX adapts","Video model's hidden visual knowledge emerges with few examples"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the assertion that CogVideoX1.5 was not pretrained on data resembling ARC-AGI; if its training corpus contained grid-like puzzles, the 16.75% accuracy could come from distribution overlap rather than latent visual understanding.","fun_headline_variants_meta":{"raw":{"variants":["Video model masters new tasks from just a few clips","LoRA on video transitions unlocks diverse visual tasks","Frozen video diffusion model learns from handful of examples","From video generation to few-shot vision: CogVideoX adapts","Video model's hidden visual knowledge emerges with few examples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1472,"prompt_tokens":875,"completion_tokens":597,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":519}},"tokens_in":491,"tokens_out":597,"duration_ms":6403,"temperature":1.0,"reasoning_tokens":519,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:37:50.035135+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Audit CogVideoX1.5's training data for ARC-like grid puzzles, or run the identical LoRA transition-video procedure on a VDM pretrained only on natural videos: if that model still scores near 16.75% on ARC-AGI, the result does not depend on grid-like exposure, while if it collapses to near zero the paper's key premise fails.","supporting_citations":[{"cited_title":"The conceptarc benchmark: Evaluating understanding and generalization in the arc domain.Trans","cited_arxiv_id":null,"evidence_quote":"ConceptARC provides the controlled concept-level reasoning benchmark and its evaluation methodology."}],"review_version":1}