{"id":"57734da4-92b5-4b07-a0db-97475c3dd4bd","arxiv_id":"2508.07312","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"A mobile-friendly video-text model, MobileViCLIP, claims zero-shot retrieval on par with a much larger model at 55.4x lower latency on phones.","lead":"This paper introduces MobileViCLIP, a compact video-text model that claims to match a large server model's zero-shot retrieval while running 55x faster on a phone. The payoff, if the numbers hold, is video search and understanding that run on-device without a network connection.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 55.4x/parity depends on matched protocol; abstract omits frame count, resolution, token count, precision, and runtime, so the speed ratio may partly reflect protocol differences rather than architecture efficiency.","rationale":"The strongest claim is the efficiency-parity trade-off. The weakest point in the argument is not the architecture itself but the comparison protocol: the 55.4x/6.7x speed ratios and the 6.9%/parity retrieval comparisons are only interpretable if both models see identical inputs, use the same precision, and are measured on the same mobile runtime. The reader's weakest_assumption identifies exactly this issue. The full text is garbled, so tables, ablations, and protocol details cannot be audited; the abstract alone does not state whether the InternVideo2 numbers come from its published protocol or from a matched mobile protocol. Because InternVideo2 shares a senior author, the burden of demonstrating protocol matching is higher, though this is not an allegation of misconduct. The code link is a positive sign, but the code was not verified in this pass. No fatal flaw is visible, so REJECT would be wrong; but the central claim cannot be accepted from the available artifact, so I keep the reader's UNVERDICTED verdict (UNCHANGED). The concrete test above would settle whether the concern lands.","tokens_in":17229,"tokens_out":3972,"duration_ms":41926,"concrete_test":"On the same mobile device, run MobileViCLIP-Small and the released InternVideo2-S14/L14 checkpoints under exactly matched inputs — same frame count, spatial resolution, token length, text prompts, precision, and inference library — and record end-to-end latency and MSR-VTT text-to-video retrieval. Then repeat under InternVideo2's published heavy protocol. If the matched-protocol speed ratio falls materially below 55.4x/6.7x, or the retrieval parity/gap changes by more than about 2 points, the abstract's headline needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — MobileViCLIP-Small is 55.4x faster than InternVideo2-L14 and 6.7x faster than S14, with retrieval parity to L14 and +6.9% over S14 on MSR-VTT — is meaningful only if both models are evaluated under identical conditions. InternVideo2 is typically evaluated with more frames and higher resolution than a mobile-optimized model would use, while MobileViCLIP inherits an image-text backbone and is designed for low mobile latency. The abstract does not disclose frames, spatial resolution, token count, precision, batch size, device, or runtime for either model. Consequently, the 55.4x ratio likely combines genuine architectural efficiency with a lighter input protocol; similarly, the 6.9% MSR-VTT gain over InternVideo2-S14 could shift if the baseline used its published heavier protocol instead of MobileViCLIP's protocol. Because InternVideo2 shares a senior author with this paper, protocol consistency is not independently established. This is an external-validity/auditability concern, not an internal inconsistency: no contradiction is visible in the readable text, but the headline numbers are not yet benchmark-controlled.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MobileViCLIP, an efficient video-text model for mobile devices, built by injecting temporal structural reparameterization into an efficient image-text backbone and training on a large-scale video-text dataset. The central claims are that MobileViCLIP-Small is 55.4x faster than InternVideo2-L14 and 6.7x faster than InternVideo2-S14 on mobile-device inference, while achieving zero-shot retrieval performance similar to InternVideo2-L14 and 6.9% better than InternVideo2-S14 on MSR-VTT. The paper also releases code. The submitted text is heavily corrupted in the version I received, so I could verify the abstract and the main numeric tables but not the full method or implementation details.","tokens_in":17442,"tokens_out":5272,"duration_ms":56898,"significance":"If the headline results are validated under controlled conditions, this is a practically significant contribution: it would demonstrate that a small, mobile-deployable video-text model can approach the zero-shot retrieval accuracy of much larger models at a fraction of the latency, with code released to support reproducibility. The claims are precise and falsifiable, which is a strength. However, the current manuscript does not provide enough benchmarking protocol detail or statistical uncertainty to substantiate the speedup and accuracy ratios. The speed and retrieval comparisons need to be audited under matched input protocols before the central contribution is established.","major_comments":[{"comment":"The headline speedups (55.4x vs InternVideo2-L14, 6.7x vs InternVideo2-S14) are not meaningful without specifying the inference protocol for both models. The abstract does not report input frame count, spatial resolution, number of tokens, precision, batch size, mobile device, inference engine, or latency measurement methodology. InternVideo2 is commonly evaluated with more frames and higher resolution than a mobile-optimized model would use, so the ratios may partly reflect protocol differences rather than architectural efficiency. Please provide a matched-protocol table where both models are evaluated under identical conditions, and state the protocol unambiguously.","section":"Abstract, speed comparison"},{"comment":"The claim of +6.9% over InternVideo2-S14 and parity with InternVideo2-L14 on MSR-VTT requires that both models are evaluated with the same retrieval protocol: dataset split, prompt templates, number of views, frame sampling, and text-side preprocessing. It is not sufficient to compare against published numbers if those numbers were obtained under a different protocol. Because InternVideo2 shares a senior author with this paper, independent auditability is especially important; please provide the exact evaluation scripts and a matched-protocol comparison.","section":"Experiments, zero-shot retrieval"},{"comment":"Single-point retrieval metrics are used throughout. Zero-shot video retrieval is sensitive to random seeds, prompt templates, and inference-time view sampling. Without error bars over multiple runs or a clear statement that the evaluation is deterministic and seed-independent, the reported 6.9% improvement and parity claims may be within run-to-run variation. Please report mean and standard deviation (or equivalent) for the main tables, or justify why the chosen protocol removes all randomness.","section":"Experiments, variance and significance"}],"minor_comments":[{"comment":"'55.4x times faster' is redundant; use '55.4x faster'. Also clarify whether 'faster' means wall-clock latency, throughput, or end-to-end mobile inference time.","section":"Abstract"},{"comment":"The submitted text is severely corrupted in many sections, especially the method and implementation-detail paragraphs. Please ensure the actual submission renders all equations and tables legibly, and that the mobile benchmark (device, SDK, precision, batch size) is described in a caption or a dedicated paragraph.","section":"Experiments, implementation details"},{"comment":"The GitHub link is helpful, but please state the license, the exact evaluation scripts needed to reproduce the MSR-VTT numbers, and the version of InternVideo2 used for the baseline measurements.","section":"Code release"},{"comment":"In the provided text, several tables lack complete captions and column headers are garbled. Please ensure all tables are self-contained with units for latency, throughput, and memory, and with footnotes defining any abbreviations such as R@1, R@5, or mdR.","section":"Tables"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a real deployment gap and the claims are potentially important. My main concern is benchmarking auditability: the speed and retrieval comparisons involve a baseline from the same group, and the current manuscript does not demonstrate matched evaluation protocols. This is fixable with additional experiments and disclosure. I would also ask the editor to verify the integrity of the submitted PDF, since the text provided to me was heavily corrupted and only the abstract, tables, and a few fragments were readable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: MobileViCLIP bolts temporal structural reparameterization onto a mobile-efficient image-text backbone, trains it on video-text data, and reports a small model that is far faster on mobile devices than InternVideo2-S14/L14 while staying close on MSR-VTT retrieval. The headline numbers are concrete and falsifiable: 55.4x faster than L14, 6.7x faster than S14, retrieval parity with L14, +6.9% over S14. Those are new measurements for this specific model, not a rehash.\n\nWhat's genuinely good: the combination is a legitimate application of known pieces (structural reparameterization, MobileCLIP-style backbone) to a new domain, and the paper ships code, which makes independent verification possible. The claims are specific enough that a referee can check them.\n\nSoft spots: the abstract does not state the evaluation protocol—no frame count, spatial resolution, token count, precision, device, or latency methodology. If MobileViCLIP runs with fewer frames at lower resolution than the InternVideo2 numbers come from, the speedup ratio partly measures protocol differences, not architecture efficiency. That's an auditability concern, not a demonstrated flaw, but it means the 55.4x figure is not yet benchmark-controlled. Also, InternVideo2 shares a senior author with this paper, so independent benchmarking isn't established. The full text in the artifact I got is garbled, so I couldn't verify the method or ablations directly; based on the abstract and readable fragments, nothing looks internally inconsistent.\n\nThe paper is an engineering contribution with practical value for edge deployment, not a conceptual breakthrough. It will be useful to people building mobile video-language systems.\n\nI'd send it to peer review. The claims are precise, the code is public, and a referee can demand the missing protocol details. If the latency benchmark holds up under matched conditions, this is a solid paper. If it doesn't, the reviewer will catch it. Either way, it deserves referee time.","headline":"Plausible, useful efficiency result for mobile video-text retrieval, but the headline speedup ratios need matched-protocol details and independent benchmarking to fully trust.","tokens_in":17968,"tokens_out":1535,"would_cite":true,"duration_ms":15899,"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":"A phone-sized video-text model matches a much larger model's zero-shot retrieval while running 55.4x faster.","keywords":["video-text retrieval","zero-shot retrieval","mobile inference","structural reparameterization","efficient video model","CLIP","temporal modeling","MSR-VTT benchmark"],"falsifier":"Run both models on the same mobile device with identical input frames, resolution, token count, and precision, then measure per-video latency and MSR-VTT retrieval; if MobileViCLIP-Small is not roughly 55.4x faster than InternVideo2-L14 under these matched conditions, the central speed claim fails.","tokens_in":17067,"feed_emoji":"📱","tokens_out":7461,"duration_ms":70325,"temperature":0.7,"pith_summary":"This paper tries to establish that a video-text model small enough for mobile devices can match the zero-shot retrieval of much larger video pretraining models. The recipe is to take an efficient image-text backbone, add temporal structural reparameterization so the model learns from video during training, and fold the temporal branches back into the base architecture for deployment. Trained on a large-scale high-quality video-text dataset, MobileViCLIP-Small is reported to run 55.4x faster than InternVideo2-L14 and 6.7x faster than InternVideo2-S14 on mobile, while matching L14's zero-shot retrieval on MSR-VTT and beating S14 by 6.9%. The consequence is that on-device video-text search and classification become practical without the usual accuracy penalty.","feed_headline":"Phone-sized video model matches large-model retrieval","feed_subtitle":"MobileViCLIP-Small runs 55.4x faster than InternVideo2-L14 on mobile while matching its zero-shot retrieval.","key_machinery":"Temporal structural reparameterization: a training-time technique that augments an efficient image-text backbone with temporal modeling branches and then merges those branches back into the original architecture before deployment. Its role is to give the model motion- and order-aware video representations during pretraining while keeping inference cost roughly equal to the image-only backbone, which is what makes the reported mobile speedups possible.","core_discovery":"The central claim is that temporal structural reparameterization lets a lightweight image-text model acquire video understanding without a deployment-time latency penalty. During training, MobileViCLIP uses extra temporal branches that process information across frames; after training, those branches are folded back into the base efficient architecture, so at inference the model runs like the original fast image-text model with video-specific weights. Trained on a large-scale high-quality video-text dataset, MobileViCLIP-Small achieves zero-shot retrieval performance similar to InternVideo2-L14 on MSR-VTT while being 55.4x faster on a mobile device, and outperforms InternVideo2-S14 by 6.9% w","pith_inferences":["The same train-with-temporal-branches, fold-before-deploy recipe could be applied to other efficient image-text backbones to produce mobile video models for new domains, a direction the paper does not run itself.","The reported 55.4x speed ratio likely combines architecture efficiency with input sampling choices such as fewer frames, lower resolution, and fewer tokens; a protocol-matched comparison would isolate the pure architectural advantage.","Given that a small model can match larger ones after high-quality video-text pretraining, scaling data quality rather than model size may be a cheaper route to strong video-text understanding.","The temporal reparameterization idea may transfer to other time-ordered inputs, such as audio or sensor streams, where a deployable model needs to be fast at inference but can afford richer training-time structure."],"forward_implications":["Mobile devices can run zero-shot video-text retrieval with accuracy close to much larger server-side models, making on-device video search and filtering practical.","Because the temporal branches are folded back into the base architecture, video understanding at inference time costs roughly the same as image understanding.","The efficiency gain is not tied to a smaller model alone: the reparameterized design keeps a lightweight backbone fast while still learning temporal cues.","A small model can reach near-parity with a much larger video-text model on MSR-VTT, so mobile deployment need not mean abandoning zero-shot generalization."],"supporting_citations":[],"fun_headline_variants":["Video smarts on phone: 55x faster, same recall","MobileViCLIP: giant-model video skills, phone-size speed","Phone video model matches giant's recall at 55x speed","Folding temporal branches: video understanding without latency"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The headline speed and accuracy ratios presuppose that MobileViCLIP and InternVideo2 are benchmarked with matched input frames, spatial resolution, token counts, mobile runtime, and precision; if the comparison uses the heavy published protocol for InternVideo2 and a lighter protocol for MobileViCLIP, the 55.4x and 6.9% numbers partly measure protocol differences rather than architecture efficiency.","fun_headline_variants_meta":{"raw":{"variants":["Video smarts on phone: 55x faster, same recall","MobileViCLIP: giant-model video skills, phone-size speed","Phone video model matches giant's recall at 55x speed","Folding temporal branches: video understanding without latency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000929,"raw_usage":{"total_tokens":3806,"prompt_tokens":722,"completion_tokens":3084,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":3014}},"tokens_in":466,"tokens_out":3084,"duration_ms":18727,"temperature":1.0,"reasoning_tokens":3014,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T22:12:50.683740+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both models on the same mobile device with identical input frames, resolution, token count, and precision, then measure per-video latency and MSR-VTT retrieval; if MobileViCLIP-Small is not roughly 55.4x faster than InternVideo2-L14 under these matched conditions, the central speed claim fails.","supporting_citations":[],"review_version":1}