{"id":"0429168e-9bfb-4c11-a919-e10a9a08a26c","arxiv_id":"2506.08887","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A parameter-efficient video-text retrieval method that trains only 0.56M parameters on top of frozen CLIP and achieves 50.5% R@1 on MSRVTT.","lead":"DiscoVLA adapts CLIP to video-text retrieval by fusing image and video features and distilling image-level alignment into video-level alignment. It reports state-of-the-art recall on four video-text retrieval benchmarks with only 0.56 million trainable parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 1.5% R@1 gain over TempMe may be seed noise; no error bars and hyperparameters tuned on MSRVTT test set undermine the SOTA claim.","rationale":"The reader's verdict is CONDITIONAL, and the reader listed missing error bars as a supporting concern but chose PImgAlign caption quality as the weakest assumption. I agree that caption quality is a real risk, but it is not the most load-bearing assumption for the central claim. The B1 ablation shows that even without PImgAlign or AlignDistill, the IVFusion-only variant already surpasses TempMe on MSRVTT ViT-B/32 (201.7 vs 198.6 R@sum), so the claimed advantage does not depend critically on pseudo captions. In contrast, the statistical robustness of the headline result is directly load-bearing: a 1.5 R@1 gap over the strongest baseline, with no variance reporting and hyperparameters selected on the test benchmark, could plausibly vanish under seed noise or overfitting. The proposed concrete test will settle this by providing confidence intervals and an honest validation split. If the difference remains significant, the central claim is supported; if not, the paper should be revised to weaken the SOTA claim or provide multi-seed evidence. Since the reader's verdict already conditionally accepts pending such robustness checks, my read does not change the verdict, hence UNCHANGED.","tokens_in":17272,"tokens_out":12390,"duration_ms":121859,"concrete_test":"Run DiscoVLA and TempMe on MSRVTT ViT-B/16 with 5 random seeds using identical data splits and hyperparameters. For hyperparameter selection, use a validation split (e.g., 9K train / 1K-B) and only use the 1K-A test set for final evaluation. Report per-seed R@1 and R@sum, mean ± std, and a two-sided bootstrap test of the R@1 difference. If the 95% CI of (DiscoVLA - TempMe) includes 0, the central SOTA claim is not statistically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DiscoVLA outperforms prior methods is supported by single-run comparisons. On MSRVTT (ViT-B/16), DiscoVLA achieves 50.5% R@1 versus 49.0% for TempMe (Table 1), a 1.5-point gap. TempMe is a concurrent arXiv preprint by the same first author, and no standard deviation, confidence interval, or multi-seed average is reported anywhere in the paper. Appendix B (Figures 6-9) shows alpha, beta, H_v, and H_l tuned on MSRVTT, and the final values are then reported on all datasets. If the test set (MSRVTT 1K-A) was used for model selection without a separate validation split, the reported number is optimistically biased. In video-text retrieval, seed-level variation is typically on the order of 1 R@1, so the 1.5-point margin could be within noise. The PImgAlign caption-quality assumption identified by the reader is secondary: the B1 ablation (IVFusion only, no pseudo captions) already reaches 201.7 R@sum on MSRVTT ViT-B/32, exceeding TempMe's 198.6, so the headline SOTA does not rest solely on pseudo captions. The load-bearing vulnerability is statistical: without variance estimates, the claimed superiority over the strongest baseline is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DiscoVLA, a parameter-efficient adaptation of CLIP for video-text retrieval. It identifies three discrepancies between image-level and video-level understanding (vision, language, and alignment) and introduces three components: IVFusion, which fuses image-level and video-level features via a lightweight adapter in both the vision and text encoders; PImgAlign, which uses LLaVA-NeXT-generated pseudo image captions to learn fine-grained image-level alignment; and AlignDistill, which distills image-level alignment knowledge into video-level alignment via KL divergence. The method uses only 0.56M trainable parameters and reports state-of-the-art results on MSRVTT, LSMDC, ActivityNet, and DiDeMo.","tokens_in":17605,"tokens_out":6694,"duration_ms":68122,"significance":"If the empirical results are robust, this work makes a solid contribution to parameter-efficient video-text retrieval. The main strength is the IVFusion module, which provides a substantial gain in the ablation (8.7 R@sum over the LoRA baseline) while adding only about 0.05M parameters, and the overall framework is well-structured with extensive comparisons and ablations. The release of code is a strength. However, the central claim of state-of-the-art performance is weakened by the absence of any statistical significance testing and by the use of the MSRVTT test set for hyperparameter selection, which may bias the reported headline numbers.","major_comments":[{"comment":"The paper reports single-run results with no error bars, confidence intervals, or multiple seeds. The improvement over TempMe on MSRVTT is 0.9% R@1 (ViT-B/32) and 1.5% R@1 (ViT-B/16), and on other datasets the margins vary from 0.4 to 2.0 R@1. In video-text retrieval, seed-level variation is typically on the order of 1 R@1, so these margins may be within noise. The claim that DiscoVLA \"significantly outperforms\" prior methods is not supported by statistical evidence. Please report mean and standard deviation over at least 3 random seeds for the main comparisons and for the key ablations, or at minimum provide the number of seeds and a discussion of expected variance.","section":"Section 4.2, Tables 1-4"},{"comment":"The hyperparameters α, β, H_V, and H_L are tuned on MSRVTT using text-to-video results (Figures 6-9). According to Appendix A, the model is trained on the 9,000 train+val pairs and evaluated on the 1K-A test set, so the test set is used for model selection. This makes the MSRVTT results, including the headline 50.5% R@1, optimistically biased. Please clarify whether a separate validation split was used; if not, either use a held-out validation split for hyperparameter selection and then retrain on the combined training+validation set, or show that the chosen hyperparameters are stable by reporting results on a validation split for each dataset.","section":"Appendix B, Figures 6-9; Section 4.1"},{"comment":"The PImgAlign contribution relies on the accuracy of LLaVA-NeXT pseudo image captions, but the paper does not evaluate caption quality directly. The only quantitative evidence is the indirect ablation in Table 8, which shows a 1.6 R@sum drop when video caption guidance is removed. This does not invalidate the overall method, because the IVFusion-only variant (B1 in Table 5) already outperforms prior methods, but it leaves the proposed mechanism of PImgAlign under-validated. Please include a caption-quality evaluation (e.g., comparison with ground-truth frame descriptions or human assessment) or a more direct test of how caption correctness affects retrieval performance.","section":"Section 3.4, Table 8"}],"minor_comments":[{"comment":"The GitHub URL \"DsicoVLA\" appears to be a typo; it should be \"DiscoVLA\".","section":"Abstract"},{"comment":"The contrastive loss expressions appear to be missing a negative sign and proper parentheses; the standard form is -log( exp(sim_ii/τ) / Σ_j exp(sim_ij/τ) ). Please correct the notation.","section":"Equations (2) and (3)"},{"comment":"The sentence \"we evaluate the effects of hyperparameters α, β, H V and H V\" should read \"H V and H L\".","section":"Section 4.3, last paragraph"},{"comment":"In the table caption and text, \"MSTVTT\" should be \"MSRVTT\".","section":"Table 9 caption"},{"comment":"The text \"Utilized ONL Y during the training phase\" contains a typo; it should be \"ONLY\".","section":"Figure 2"},{"comment":"The paper does not report the training-time computational cost of generating pseudo image captions with LLaVA-NeXT. Please add a note on the additional offline training overhead.","section":"Section 3.4 / Experiments"}],"recommendation":"major_revision","confidential_remarks":"The TempMe baseline (arXiv:2409.01156) is a concurrent preprint by the same first author. I do not see a correctness problem in citing it as a baseline, but the authors should ensure that the numbers reported for TempMe match the original paper and that the comparisons are apples-to-apples. Given the small performance margins, the absence of multi-seed results is particularly concerning. The hyperparameter tuning on the MSRVTT test set is a common but questionable practice; if the authors can demonstrate insensitivity to hyperparameters (as they claim in the appendix) or retune on validation splits, that would materially strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a solid parameter-efficient video-text retrieval paper with a fresh combination of ideas and unusually thorough ablations, but the headline SOTA claim is not statistically established. No error bars, no multiple seeds, and hyperparameters tuned on the MSRVTT test set. It still deserves peer review.\n\nWhat's new: the paper identifies three discrepancies between image-level CLIP and video-level retrieval—vision, language, and alignment—and addresses all three. IVFusion is the most interesting piece: it restricts spatio-temporal attention to CLS tokens and fuses image- and video-level features with a lightweight adapter, avoiding the quadratic cost of full token-level video attention. PImgAlign generates pseudo image captions with LLaVA-NeXT using the video caption as guidance, and computes a max-based fine-grained similarity to handle caption-to-image ambiguity. AlignDistill then transfers image-level alignment to the video level. The ablations (Tables 5–8) support the contribution of each module. At 0.56M trainable parameters, the method beats published baselines on four benchmarks. That is a real contribution.\n\nSoft spots: the main issue is statistical. Single-run results and the 1.5% R@1 gap over TempMe on MSRVTT ViT-B/16 could easily be seed noise. The R@sum gaps are larger (e.g., 209.9 vs 206.7), which is more reassuring, but there is no variance estimate anywhere. The hyperparameter sweeps in Appendix B are done on MSRVTT with no mention of a held-out split, so the primary number is optimistically biased. Transferring those hyperparameters to other datasets mitigates the bias but does not remove it. The pseudo-caption assumption is a secondary concern: only one ablation (Table 8) supports it, and caption quality is not measured. But the B1 ablation (IVFusion only) already exceeds TempMe on ViT-B/32 R@sum, so the method does not rest on pseudo captions alone. Also, the code link in the abstract is mistyped as DsicoVLA, which suggests the repository may not be in a usable state.\n\nWho this is for: anyone working on CLIP transfer, parameter-efficient video retrieval, or multi-objective training. The paper would benefit from a serious referee: the statistical concerns are fixable in revision, and the core method is likely useful regardless of the exact SOTA numbers. Recommendation: send to peer review, but make the authors add multiple seeds, a proper validation split for hyperparameters, and honest error bars. If they do that, this could be a solid accepted paper.","headline":"Solid and well-ablated parameter-efficient video-text retrieval method, but the SOTA claim needs error bars and a proper validation split before the 1.5% margin can be trusted.","tokens_in":18138,"tokens_out":4513,"would_cite":true,"duration_ms":45435,"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 argues that correcting three transfer gaps—vision, language, and alignment—lets a 0.56M-parameter CLIP adapter beat full fine-tuning on video-text retrieval.","keywords":["video-text retrieval","parameter-efficient transfer learning","CLIP adaptation","discrepancy reduction","vision-language alignment","pseudo caption generation","alignment distillation","temporal modeling"],"falsifier":"On a benchmark with ground-truth frame-level captions (or a human-annotated subset), compare the generated pseudo captions against the ground truth and check whether retrieval gains appear only when caption accuracy is high. Alternatively, corrupt the pseudo captions with random substitutions of nouns or verbs and measure the drop in R@1, which would reveal whether PImgAlign is learning from caption content or merely from the presence of extra training pairs.","tokens_in":17099,"feed_emoji":"🎬","tokens_out":6751,"duration_ms":64864,"temperature":0.7,"pith_summary":"DiscoVLA is a parameter-efficient method that adapts the frozen image-text model CLIP to video-text retrieval. It argues that transferring CLIP from images to videos suffers from three separate discrepancies: vision (missing temporal information), language (image captions versus video captions), and alignment (video-level matching is more complex than image-level matching). Existing parameter-efficient methods mostly address the vision gap alone, so DiscoVLA tackles all three at once with an image-video feature fusion, pseudo image-level alignment from generated captions, and image-to-video alignment distillation. With only 0.56M trainable parameters, it reaches 50.5% text-to-video R@1 on MSRVTT and outperforms prior methods on four benchmark datasets.","feed_headline":"0.56M parameters beat full fine-tuning on video-text retrieval","feed_subtitle":"DiscoVLA closes all three CLIP transfer gaps with 0.56M parameters, reaching 50.5% R@1 on MSRVTT.","key_machinery":"The load-bearing mechanism is IVFusion, a parameter-efficient attention variant that processes sampled frames with image-level attention and applies spatio-temporal attention only to the CLS tokens, merging the two branches with a lightweight adapter to keep complexity at $O(FN^2 + F^2N)$. PImgAlign supplies the fine-grained training signal: pseudo image captions produced from frames plus the video caption, matched to images through a max-over-row-and-column similarity, which tolerates one caption covering several frames. AlignDistill then distills the resulting image-level similarity distributions into the video-level similarity used for retrieval; both PImgAlign and AlignDistill are discarded at inference so the retrieval cost stays unchanged.","core_discovery":"The central claim is that the transfer from image-level CLIP to video-level retrieval can be decomposed into vision, language, and alignment discrepancies, and that jointly reducing all three yields state-of-the-art parameter-efficient video-text retrieval. On MSRVTT with CLIP (ViT-B/16), DiscoVLA attains 50.5% text-to-video R@1 and 49.2% video-to-text R@1, surpassing previous parameter-efficient methods and the full fine-tuning baseline while updating only 0.56M parameters. The gains come from three coupled modules: IVFusion merges image- and video-level features in both encoders; PImgAlign learns fine-grained image-level alignment from pseudo image captions generated by a multimodal language model; AlignDistill transfers that image-level alignment knowledge to video level via KL divergence.","pith_inferences":["A testable corollary not in the paper: the value of PImgAlign should scale with the quality of the pseudo captions; replacing the multimodal language model with a weaker captioner (or adding caption noise) would isolate how much of the improvement is due to caption fidelity rather than the alignment loss itself.","The max-over-row-and-column similarity acts as a soft assignment between frames and captions; in datasets with one-to-many frame-caption relations, other assignment strategies such as optimal transport might perform better, which the paper does not explore.","Because PImgAlign is training-only, the method effectively uses the video caption as a prior to re-caption frames; this suggests a general recipe for upgrading image-text backbones to video tasks without changing the inference-time architecture."],"forward_implications":["If DiscoVLA's gains hold, parameter-efficient video-text retrieval can match or beat full fine-tuning while updating two orders of magnitude fewer parameters.","Methods that focus only on temporal modeling are leaving language and alignment discrepancies on the table; addressing all three yields consistent improvements across MSRVTT, LSMDC, ActivityNet, and DiDeMo.","Image-level alignment learned from pseudo captions, combined with distillation, transfers CLIP's image-text matching strength to video-level matching without any inference-time caption generation.","The IVFusion complexity reduction, which applies spatio-temporal attention only to CLS tokens, makes video-level feature fusion computationally affordable in large pre-trained encoders."],"supporting_citations":[{"why":"Supplies the frozen image-text backbone that DiscoVLA adapts to video-text retrieval.","marker":"[42]"},{"why":"Provides the full fine-tuning baseline and the temporal average pooling strategy used by all compared methods.","marker":"[37]"},{"why":"A prior parameter-efficient method with BiLSTM-generated prompts; one of the key comparisons.","marker":"[22]"},{"why":"A prior method with dynamic global-local prompt tuning; another key comparison.","marker":"[58]"},{"why":"A prior method with shared bottleneck structures in both video and text branches; a key comparison.","marker":"[25]"},{"why":"A prior sparse-and-correlated adapter method for efficient text-video retrieval; a key comparison.","marker":"[5]"},{"why":"The multimodal large language model used to generate pseudo image captions for PImgAlign.","marker":"[35]"},{"why":"A recent temporal token merging baseline that DiscoVLA outperforms; supplies the strongest prior result on MSRVTT.","marker":"[44]"}],"fun_headline_variants":["0.56M params beat full fine-tuning on video-text retrieval","Three CLIP gaps closed with 0.56M parameters","DiscoVLA: tiny CLIP adaptation tops video-text retrieval","Video-text retrieval: 0.56M-param model outdoes full tuning","CLIP to video: closing vision, language, alignment gaps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"DiscoVLA's alignment gains depend on the pseudo image captions generated by a multimodal language model from a few frames and the video caption being accurate enough to serve as a training signal; the paper does not measure caption quality and only ablates the video-caption guidance.","fun_headline_variants_meta":{"raw":{"variants":["0.56M params beat full fine-tuning on video-text retrieval","Three CLIP gaps closed with 0.56M parameters","DiscoVLA: tiny CLIP adaptation tops video-text retrieval","Video-text retrieval: 0.56M-param model outdoes full tuning","CLIP to video: closing vision, language, alignment gaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000281,"raw_usage":{"total_tokens":1675,"prompt_tokens":970,"completion_tokens":705,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":613}},"tokens_in":586,"tokens_out":705,"duration_ms":8191,"temperature":1.0,"reasoning_tokens":613,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:58:48.432880+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a benchmark with ground-truth frame-level captions (or a human-annotated subset), compare the generated pseudo captions against the ground truth and check whether retrieval gains appear only when caption accuracy is high. Alternatively, corrupt the pseudo captions with random substitutions of nouns or verbs and measure the drop in R@1, which would reveal whether PImgAlign is learning from caption content or merely from the presence of extra training pairs.","supporting_citations":[{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen image-text backbone that DiscoVLA adapts to video-text retrieval."},{"cited_title":"Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning.Neu- rocomputing, 2022","cited_arxiv_id":null,"evidence_quote":"Provides the full fine-tuning baseline and the temporal average pooling strategy used by all compared methods."},{"cited_title":"V op: Text-video co- operative prompt tuning for cross-modal retrieval","cited_arxiv_id":null,"evidence_quote":"A prior parameter-efficient method with BiLSTM-generated prompts; one of the key comparisons."},{"cited_title":"Dgl: Dynamic global-local prompt tuning for text-video re- trieval.Proceedings of the AAAI Conference on Artificial Intelligence, 2024","cited_arxiv_id":null,"evidence_quote":"A prior method with dynamic global-local prompt tuning; another key comparison."},{"cited_title":"Mv-adapter: Multimodal video transfer learning for video text retrieval","cited_arxiv_id":null,"evidence_quote":"A prior method with shared bottleneck structures in both video and text branches; a key comparison."},{"cited_title":"RAP: Efficient text-video retrieval with sparse-and- correlated adapter","cited_arxiv_id":null,"evidence_quote":"A prior sparse-and-correlated adapter method for efficient text-video retrieval; a key comparison."},{"cited_title":"Llava-next: Im- proved reasoning, ocr, and world knowledge, 2024","cited_arxiv_id":null,"evidence_quote":"The multimodal large language model used to generate pseudo image captions for PImgAlign."}],"review_version":1}