acceptodds
Under review as a conference paper at ICLR 2027

Long-Term Composed Video Retrieval: Benchmarking and Adaptive History Refocusing

Abstract

The rapid growth of video data has made Composed Video Retrieval (CVR) an emerging video retrieval paradigm. Existing CVR methods have achieved strong performance on short-term, i.e., single-turn, benchmarks, but they remain limited in real-world long-term retrieval. In practice, users often refine their requests iteratively based on intermediate retrieval results, rather than expressing the full intent in a single modification text. To systematically evaluate this challenge, we introduce LTCVR, a novel benchmark for Long-Term Composed Video Retrieval, featuring five representative history-dependency patterns: progressive refinement, rollback, cross-turn combination, local correction, and preference change. Our analysis on LTCVR shows that existing methods degrade markedly when the retrieval intent depends on earlier interaction context, and attention analysis further reveals that representative methods fail to focus on the key historical turns. To address this limitation, we propose InAHR, an Intent-Driven Adaptive History Refocusing framework. InAHR first infers the current retrieval intent and then selects a compact subset of historical visual evidence that is intent-relevant and representative. The selected evidence is re-injected into the generation process to produce retrieval-oriented target descriptions for video retrieval. Extensive experiments demonstrate that InAHR consistently outperforms strong baselines across diverse long-horizon CVR scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Composed Video Retrieval, Long-Term Interaction, Multi-Turn Retrieval

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_36BhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies long-term composed video retrieval (CVR), where users iteratively refine their search requests through reference videos and textual feedback. Existing CVR benchmarks mainly focus on single-turn retrieval and do not adequately capture how user intent changes across interactions. To address this gap, the authors introduce LTCVR, a benchmark with 4,000 five-turn dialogues and 3,160 candidate videos. It covers five scenarios: progressive refinement, rollback, cross-turn combination, local correction, and preference change, which test whether models can retain useful historical information, revisit earlier results, combine constraints across turns, and discard outdated preferences. The authors also propose InAHR, a framework that selectively revisits interaction history. It first uses one representative frame from each reference video, together with the corresponding feedback, to infer the current retrieval intent and identify relevant turns. It then selects a compact set of historical frames by combining turn-level and frame-level relevance with a facility-location objective to reduce redundancy and preserve complementary evidence. Finally, a multimodal large language model uses the selected evidence to generate multiple target descriptions, whose text embeddings are averaged for retrieval. On LTCVR, InAHR achieves 41.9% Recall@1 and 66.7% Recall@5, exceeding the strongest baseline, with improvements across all five scenarios. The paper also includes component ablations and evaluations with different model backbones and more histories to further examine the method’s effectiveness.

Critical strengths

The benchmark provides a structured test of history-dependent video retrieval. LTCVR distinguishes between retaining, revisiting, combining, and discarding historical information, enabling a more targeted assessment than single-turn retrieval benchmarks. Its documented human review and quality-control process strengthen its value as a diagnostic resource, although its relevance to real user behavior still needs validation.

Critical weaknesses
  1. InAHR combines established techniques: MLLM-based intent inference, CLIP-based relevance scoring, facility-location selection, and multi-description aggregation. The pipeline is reasonable, but the paper does not establish a substantial technical advance beyond adapting these components to the proposed task. The main refocusing ablation also changes the available visual evidence, making it difficult to separate the benefit of adaptive selection from that of additional frames. This weakens the method’s contribution to the case for acceptance.
  2. LTCVR is built from synthetically generated trajectories and scenario-specific rewrites, without real user sessions or a user study demonstrating that these patterns reflect actual video-search behavior. Human verification supports sample correctness and coherence, but does not establish the prevalence of rollback or cross-turn combination. Evaluation also uses fixed histories, leaving the effectiveness of the task in a closed-loop search setting unclear. Furthermore, the paper does not report video-duration statistics or quantify how many queries require temporal reasoning. Given the method’s reliance on static frame features and mean-pooled video representations, a temporally demanding subset or frame-order perturbation study is needed to distinguish video understanding from static appearance matching.
Questions for authors
  1. What motivates the fixed five-turn setting, and which realistic video-search needs require history-dependent interaction rather than a single, sufficiently detailed query? Please clarify when users would benefit from progressively expressing their requirements, and whether the proposed setup reflects such needs or primarily serves as a controlled task construction.
  2. What are the average duration and duration distribution of LTCVR videos, and what proportion of queries require understanding action order or temporal relationships? How does performance change when only a single frame is provided or frame order is shuffled?
  3. In Table 3, InAHR’s Recall@1/Recall@5 on Preference Change drops from 39.8/68.3 with last-turn input to 27.2/51.8 with full history, falling below even the last-turn results of CoTMR and ImageScope. What causes this substantial degradation despite the intent-driven history selection? These results suggest that historical interference is reduced rather than effectively eliminated; could the authors clarify the scope of their claim and whether the method can learn when to discard history entirely?
  4. Under the same budget of eight evidence frames, two MLLM calls, and three target descriptions, how does InAHR compare with random frame selection, uniform sampling, and simple relevance-based selection? This comparison would help isolate the contribution of adaptive history refocusing from the benefits of additional visual evidence and inference computation.
  5. LTCVR constructs queries around known target videos, so every evaluated request has a matching answer by design. In real-world retrieval, however, users may specify modifications or combine constraints that no video in the gallery satisfies. How would the proposed framework handle such requests? Can the benchmark include no-match queries or partially satisfiable constraints to evaluate whether the system recognizes the absence of a suitable result rather than simply returning the closest video?
1 comment
AuthorsOpenReview ↗

We added equal-budget ablations, clarified the five-turn protocol and temporal limitations, and examined historical interference and behavior on no-match queries.

Dear Reviewer 36Bh,

  • Thank you for your thoughtful feedback. We address your concerns below.

1. W1 & Q4: Methodological Novelty and Equal-Budget Ablation

  • We acknowledge that the individual components are established techniques. InAHR's contribution lies in hierarchical, intent-conditioned history refocusing, which integrates soft turn priors, frame-level relevance, and representative selection to recover relevant historical evidence under evolving retrieval intents.

  • We also recognize the unequal frame budget in the original w/o Refocusing ablation. We therefore conducted equal-budget comparisons using eight frames, two MLLM calls, and three target descriptions, with all other settings unchanged. Relevance-8 corresponds to w/o Representative Selection in Table 3 of the revised manuscript.

    MethodR@1R@5R@10Avg.
    Random-832.856.466.852.0
    Uniform-838.662.973.158.2
    Relevance-840.564.774.459.9
    InAHR-841.966.776.561.7
  • InAHR improves Avg. by 9.7, 3.5, and 1.8 points over Random-8, Uniform-8, and Relevance-8, respectively, supporting the benefits of intent-guided, representative evidence selection beyond additional frames or relevance alone.

2. W2 & Q1: Five-Turn Protocol and Realistic Interactions

  • We adopt a standardized five-turn protocol to systematically evaluate diverse intent transitions under controlled conditions, particularly Rollback and Cross-turn Combination. This setting provides sufficient historical context to construct challenging interaction patterns while ensuring fair comparisons across methods.

  • Importantly, multi-turn retrieval is not simply splitting a complete query into multiple instructions. For example, when searching for travel videos, a user may prefer the beach scene in an earlier result and the sunset atmosphere in a later one, then request a video combining both. Such preferences emerge through interaction, requiring models to recover and integrate historical visual evidence rather than relying solely on the latest query.

  • While our benchmark adopts five turns for controlled evaluation, InAHR can also handle longer interactions. Appendix F.4 evaluates 7- and 9-turn extensions (600 sessions each), achieving 33.50/55.33 and 30.33/53.67 R@1/R@5, respectively, with gains of 7.50/10.00 and 6.66/12.17 points over the strongest baselines. We acknowledge that variable-length interactions better reflect real-world retrieval and will explore them in future benchmark extensions.

3. W2 & Q2: Video Duration and Temporal Understanding

  • We agree that distinguishing temporal understanding from static appearance matching is important. Our multimodal ablation shows that replacing historical videos with captions reduces overall R@1 from 41.9% to 37.4%, highlighting the value of historical visual evidence. We further analyzed video durations and identified temporally dependent queries involving state changes, action order, or temporal relationships.

    StatisticValue
    Mean Duration (s)18.44
    Median Duration (s)15.83
    Min / Max Duration (s)0.32 / 387.40
    Temporally Dependent Queries49.93%
  • These results indicate that LTCVR includes substantial temporal dependencies beyond static appearance matching. However, we acknowledge that InAHR primarily leverages multi-frame visual evidence without explicitly modeling fine-grained temporal order, and will clarify this limitation in the revised manuscript.

4. Q3: Historical Interference on Preference Change

  • We acknowledge that Soft Prior may retain obsolete history. We therefore evaluated a Conditional Hard Evidence Switch that selects all eight frames from the latest turn when Stage 1 identifies it as the only relevant turn; otherwise, Soft Prior is retained. This requires no ground-truth scenario labels and preserves the inference pipeline.

    MethodPref. R@1Pref. R@5
    InAHR (Soft Prior)27.251.8
    Conditional Hard Switch28.551.8
  • The switch improves R@1 from 27.2 to 28.5, while R@5 remains unchanged. In a separate ablation, Soft Prior outperforms Hard Turn Mask overall (61.7 vs. 60.3 Avg.), suggesting that hard masking may discard useful evidence. We therefore retain Soft Prior for broader applicability across different scenarios.

5. Q5: No-Match Queries

  • We agree that real-world retrieval may involve queries without fully matching videos. LTCVR adopts a closed-gallery setting to evaluate history-dependent intent understanding and ranking. Without a matching target, InAHR would return the highest-ranked video rather than abstain. Recognizing the absence of a suitable result would require additional answerability assessment and abstention mechanisms, which we consider valuable directions for future work.
Reviewer_7sTAOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces LTCVR, a benchmark of 4,000 five-turn composed-video-retrieval sessions covering progressive refinement, rollback, cross-turn combination, local correction, and preference change. Sessions are constructed from WebVid videos using target-conditioned synthetic interactions, filtering, and human review. Evaluation uses fixed histories and a gallery of 3,160 target videos. The proposed InAHR framework first infers the current intent and relevant turns, selects historical frames using hierarchical relevance and a weighted facility-location objective, and generates complementary target descriptions for retrieval. It improves over the evaluated baselines, with additional analyses of relevant-turn identification, longer histories, and alternative model backbones.

Critical strengths
  1. LTCVR goes beyond monotonic refinement by organizing 4,000 five-turn sessions into five explicit history-dependency patterns (progressive refinement, rollback, cross-turn combination, local correction, and preference change) and backing them with substantial human quality control: Appendix B.4 documents independent 3-reviewer ratings for all 4,000 sessions (Fleiss' kappa = 0.48–0.53 in Table 6), 1,185 manual revisions, 532 replacements, and re-verification by non-revising annotators.

  2. InAHR improves overall R@1 from 37.8% (ImageScope) to 41.9% and R@5 from 61.5% to 66.7% while sharing the exact same OpenCLIP retrieval encoder and precomputed gallery index as the text-generation baselines (Table 2; Appendix D.1). The gains hold across alternative MLLM backbones (Qwen3-VL-8B, InternVL3-8B, GPT-4o in Table 9), stronger retrieval encoders (Table 10), 7- and 9-turn extensions (Table 12), and component ablations covering soft vs. hard turn selection and facility-location selection (Table 4).

Critical weaknesses
  1. About Dataset Construction. While Appendix B documents the LTCVR construction pipeline, three key gaps remain:

(i) Appendix B.2 (Eq. 13) omits the exact retrieval model (Psi, Phi) used to retrieve intermediate reference videos v_{n+1} during seed trajectory generation;

(ii) both synthesis and human review (Appendix B.4) only verify consistency against a single predefined target video v*, without checking whether other videos in the 3,160-video gallery also satisfy the multi-turn instructions (false negatives);

(iii) prior to human review, GPT-4o both generates the 40,000 seed dialogues and filters out 59.5% of them (retaining 16,190), creating a potential single-model survivor bias.

Please give point-to-point to show the soundness of the proposed benchmark construction.

  1. In Table 3, feeding the full history on Preference Change still drops InAHR from 39.8% (last-turn only) to 27.2% R@1, showing that soft turn weighting still fails to suppress obsolete history when prior turns should be discarded. In addition, the "w/o Refocusing" ablation in Table 4 uses the 5-frame sparse overview instead of 8 frames (M=8), confounding the selection strategy with a smaller frame budget, and the strongest reasoning baseline in Table 11 (CoVR-R [R3], along with CoVR [R2]) is omitted from the main multi-turn comparison in Table 2. Please add an equal-budget 8-frame uniform sampling baseline, compare against an intent-conditioned hard switch to the last turn on Preference Change, and evaluate CoVR-R [R3] on LTCVR (or test on an external multi-turn benchmark such as IVCR-200K [R1]) to better demonstrate the effectiveness of your proposed method.

References:

[R1] Ning Han et al. IVCR-200K: A Large-Scale Multi-turn Dialogue Benchmark for Interactive Video Corpus Retrieval. SIGIR 2025. https://arxiv.org/abs/2512.01312

[R2] Lucas Ventura et al. CoVR: Learning Composed Video Retrieval from Web Video Captions. AAAI 2024. https://imagine.enpc.fr/~ventural/covr/

[R3] Omkar Thawakar et al. CoVR-R: Reason-Aware Composed Video Retrieval. 2026. https://arxiv.org/abs/2603.20190

Questions for authors
  1. Did you check how often top-ranked non-target videos in the 3,160-video gallery are also valid matches for the session history, and were any LTCVR videos or seed trajectories shared with the WebVid-CoVR training split or used for hyperparameter tuning (M, tau, beta)?

  2. Under the exact same two-stage prompts and an equal 8-frame budget, how does uniform 8-frame sampling across the session compare with InAHR in Table 4? On Preference Change (Table 3), what happens if Stage 1's predicted scenario or turn set triggers a hard fallback to the last turn?

  3. Can you clarify the exact retriever used to generate the progressive seed trajectories in Eq. 13 (Appendix B.2), confirm whether human ratings were compared against the Gemini 3 Flash judge in Table 5, and provide the stage-wise latency breakdown for Figure 5?

2 comments
AuthorsOpenReview ↗

We clarified benchmark reliability and evaluation settings, added controlled ablations and CoVR-R comparisons, and reported human–judge agreement and stage-wise latency.

Dear Reviewer 7sTA,

  • Thank you for your constructive feedback. We address your concerns below.

1. W1: Benchmark Reliability

  • W1.1: We use OpenCLIP ViT-G/14 (laion2b_s34b_b88k) in Eq. (13). We mean-pool 10 sampled frame features, L2-normalize video/text embeddings, then sum and re-normalize them to retrieve the next reference by cosine similarity.

  • W1.2: We audited 200 randomly sampled sessions (40 per scenario), merging and deduplicating the Top-5 results from InAHR, CoTMR, and ImageScope after excluding the predefined GT. Three annotators independently judged candidates against the full history and video frames, using majority voting. 59/1,948 (3.03%) were valid alternatives, with 21.50% of sessions containing at least one. Thus, false negatives were relatively uncommon among audited high-ranking candidates.

  • W1.3: We acknowledge potential bias from using GPT-4o for both generation and screening. To mitigate this, we use distinct prompts and task roles, followed by independent human review. Among 4,000 category-balanced sessions, 2,283 (57.1%) passed unchanged, 1,185 (29.6%) were revised, and 532 (13.3%) were replaced, with independent re-evaluation. Thus, 42.9% required human intervention, providing substantial quality control. Nevertheless, potential bias from initial automatic filtering remains a limitation, which we will clarify in the revised manuscript.

2. W2: Historical Interference and Fair Comparisons

  • W2.1: We evaluated a Conditional Hard Evidence Switch on Preference Change: when Stage 1 identifies only the latest turn as relevant, Stage 2 selects all eight frames from that turn; otherwise, it retains the soft prior.

    MethodR@1R@5R@10R@50R@100
    InAHR (Soft Prior)27.251.861.080.290.5
    Conditional Hard Switch28.551.861.081.590.8

    The switch improves R@1 from 27.2 to 28.5 and R@50 from 80.2 to 81.5. In a separate ablation, Soft Prior outperforms Hard Turn Mask overall (61.7 vs. 60.3 Avg.), suggesting that hard masking may discard useful evidence. We therefore adopt Soft Prior to ensure general applicability across different history-dependency scenarios, rather than optimizing for a specific scenario.

  • W2.2: We agree that the original w/o Refocusing ablation uses five frames instead of eight, making it difficult to isolate the effect of evidence selection. We added Random-8 and Uniform-8, alongside Relevance-8, keeping Stage 1, two MLLM calls, and all other settings unchanged.

    MethodR@1R@5R@10Avg.
    Random-832.856.466.852.0
    Uniform-838.662.973.158.2
    Relevance-840.564.774.459.9
    InAHR-841.966.776.561.7

    InAHR improves Avg. by 3.5 points over Uniform-8 and 1.8 over Relevance-8, supporting adaptive, representative selection beyond frame count or relevance alone.

  • W2.3: Following your suggestion, we adapted CoVR-R to LTCVR's fixed five-turn protocol.

    MethodALLCombinationCorrectionPref. ChangeProgressiveRollback
    CoVR-R34.0 / 52.030.9 / 53.836.2 / 57.218.5 / 30.039.1 / 51.234.4 / 58.8
    InAHR41.9 / 66.740.9 / 65.446.7 / 72.727.2 / 51.845.2 / 67.641.1 / 69.5

    Each cell reports R@1 / R@5 (%).

    InAHR outperforms CoVR-R by 7.9/14.7 points in overall R@1/R@5, with consistent gains across all five scenarios. This comparison further supports the importance of explicitly modeling evolving user intent and selectively using historical visual evidence.

3. Q1: Data Overlap and Hyperparameter Selection

  • Data Overlap. We acknowledge that LTCVR and the WebVid-CoVR training split share the same video collection, resulting in video-level overlap. However, InAHR is training-free and performs no task-specific training on WebVid-CoVR. Thus, this overlap does not itself imply direct task-specific training–test leakage for InAHR.

  • Hyperparameter Selection. We used a separate 200-session development set, rather than the 4,000 evaluation sessions, to select M = 8 and τ = 0.08, which were then fixed for evaluation.

4. Q3: Human Evaluation and Efficiency

  • Human–Judge Agreement. We conducted an independent human evaluation with two annotators using the same scoring criteria. Across 400 outputs, the macro-averaged Spearman correlation reached 0.8724, with an MAE of 0.1693, indicating strong agreement between human and MLLM judgments.

  • Stage-Wise Latency. The latency breakdown (s/session) is: Intent-State Induction 5.426; Adaptive History Refocusing 0.012; Target-State Reconstruction 11.298; Final Encoding & Retrieval 0.014; total 16.75.

Thanks for the detailed rebuttal. Specifying the seed retriever and development set, checking the automatic judge against human ratings, and breaking down latency clarify the evaluation. The equal-budget ablation strengthens the case for adaptive history selection. With eight frames, InAHR reaches 41.9% R@1 versus 38.6% for uniform sampling, and the adapted CoVR-R comparison shows gains across all five scenarios. The latest-turn fallback also shows a small R@1 gain on Preference Change.

The gallery audit found at least one valid non-target match among the audited results for 21.5% of sampled sessions, so single-target recall can penalize valid matches. GPT-4o’s role in both seed generation and initial screening remains a potential source of selection bias. I am keeping my rating at 3 (weak acceptance) because the curated five-turn benchmark and the controlled retrieval gains are useful contributions within this setting.

Reviewer_yTKSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a benchmark (Long-Term Composed Video Retrieval) to evaluate multi-turn video search interactions, where user's search intents evolve across queries. The authors introduce LTCVR, a 4,000-session five-turn benchmark constructed from WebVid-2M that spans five distinct history-dependency patterns: progressive refinement, rollback, cross-turn combination, local correction, and preference change, paired with multi-level intent understanding metrics. To overcome the performance degradation and diffuse historical attention in existing models, the paper proposes InAHR (Intent-Driven Adaptive History Refocusing), a three-stage training-free framework that induces current user intent and relevant turn priors from a sparse interaction overview. It applies hierarchical relevance scoring and facility-location submodular maximization to select a compact coreset of historical video frames, and generates multiple complementary target descriptions for retrieval.

Critical strengths

I found the proposed problem very valuable given that increasingly more video search queries are performed in multi-turn discussion. Understanding the change in user intent during the conversation offers new insights and can potentially improve the quality of video search.

Critical weaknesses
  • Missing important baselines related to dedicated multi-turn composed retrieval baselines. All reported baselines are single-turn models adapted to multi-turn sessions via a fixed linear recency-weighted average or full-history concatenation. Please consider stronger baselines with dedicated multi-turn architecture [1, 2, 3, 4]
  • Unexplained evaluation protocol for first and third turn in figures 2(a)-(b). The vertical axis of Figures 2(a)-(b) is not labeled. And the evaluation protocol of of 1-turn and 3-turn is not specified.

[1] Chen, Yanzhe, et al. "MAI: A multi-turn aggregation-iteration model for composed image retrieval." International Conference on Learning Representations. Vol. 2025. 2025. [2] Cao, Shengcao, et al. "CoCo-IR: Contextual Composed Image Retrieval." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026. [3] Zhang, Bingqing, et al. "ReCoVR: Closing the Loop in Interactive Composed Video Retrieval." arXiv preprint arXiv:2605.09836 (2026). [4] Tang, Mingqiang, et al. "Diverse-Intent Multi-Turn Fashion Image Retrieval." arXiv preprint arXiv:2607.20291 (2026).

Questions for authors

My main questions to the authors are regarding the missing baselines explained in the weakness questions.

1 comment
AuthorsOpenReview ↗

We added comparisons with three dedicated multi-turn retrieval methods, with InAHR outperforming the strongest adapted baseline by 5.5/5.3 percentage points in overall R@1/R@5. We also labeled Figure 2’s vertical axes and clarified the 1-, 3-, and 5-turn protocols, including their shared targets and gallery and differences in intent expression.

Dear Reviewer yTKS,

  • Thank you for recognizing the value of long-term composed video retrieval and for your constructive suggestions. In the following, we carefully address your concerns.

1. W1: Comparisons with Dedicated Multi-Turn Retrieval Methods

  • Thank you for recommending these relevant studies. We conducted additional experiments with three representative multi-turn interactive retrieval methods: ReCoVR [1], UMIVR [2], and ADEPT [3]. ReCoVR is designed for multi-turn composed video retrieval, while UMIVR and ADEPT employ uncertainty-guided query refinement and dynamic interaction strategies, respectively. We adapted these methods to LTCVR's fixed five-turn evaluation protocol and evaluated them on the same 4,000 sessions, 3,160-video gallery, and final retrieval targets.

    MethodALLCombinationCorrectionPref. ChangeProgressiveRollback
    ReCoVR-Adapted36.4 / 61.439.8 / 64.540.3 / 63.712.5 / 35.539.1 / 66.437.5 / 61.0
    UMIVR-Adapted35.5 / 59.638.9 / 62.639.8 / 64.815.5 / 32.237.5 / 63.335.1 / 59.9
    ADEPT-Adapted33.1 / 58.535.3 / 63.837.5 / 62.510.0 / 22.236.7 / 64.632.9 / 57.8
    InAHR (Ours)41.9 / 66.740.9 / 65.446.7 / 72.727.2 / 51.845.2 / 67.641.1 / 69.5

    Each cell reports R@1 / R@5 (%).

  • InAHR consistently outperforms the additional multi-turn retrieval baselines across all scenarios. Compared with the strongest adapted baseline, ReCoVR-Adapted, InAHR improves overall R@1/R@5 by 5.5/5.3 percentage points. Notably, on Preference Change, InAHR achieves a 14.7-point improvement in R@1. These results suggest that accumulating or iteratively refining historical retrieval information remains insufficient for handling complex intent transitions, whereas explicitly identifying relevant historical turns and refocusing on their visual evidence provides additional benefits.

  • We appreciate the reviewer’s recommendation of MAI [4], CoCo-IR [5], and Diverse-Intent Multi-Turn Fashion Image Retrieval [6]. These methods primarily target multi-turn image retrieval, and their complete training resources and pretrained weights were not publicly available to us. Adapting them to video retrieval would also require architectural modifications and additional training. We therefore prioritized more directly applicable interactive video retrieval methods.

  • We will add these comparisons and clarify how LTCVR differs from existing methods in task formulation and evaluation.

2. W2: Clarification of the Evaluation Protocol in Figure 2

  • Thank you for pointing out the insufficient description of Figures 2(a)–(b). We apologize for this lack of clarity. The vertical axis represents Recall@1 (%). We will add the corresponding labels and explicitly clarify the construction and evaluation protocols for the 1-turn, 3-turn, and 5-turn settings.

  • All three settings are derived from the same LTCVR sessions, with the final target videos and the 3,160-video retrieval gallery kept unchanged. They primarily differ in how the final retrieval intent is expressed and distributed across the interaction history.

    • 1-turn: We retain the final reference video from the original LTCVR session. During offline dataset construction, an MLLM generates a relatively complete and self-contained modification instruction based on the visual differences between this reference video and the final target video, forming a standard single-turn CVR query.
    • 3-turn: We retain three representative turns from the original five-turn session, including the final turn and key historical turns relevant to the final retrieval intent. For example, in Rollback, both the final turn and the earlier referenced turn are preserved to maintain the necessary history dependency.
    • 5-turn: We use the complete five-turn interaction history from LTCVR, corresponding to our standard long-term CVR evaluation setting.

References

[1] ReCoVR: Closing the Loop in Interactive Composed Video Retrieval, arXiv:2605.09836, 2026.

[2] Quantifying and Narrowing the Unknown: Interactive Text-to-Video Retrieval via Uncertainty Minimization, ICCV 2025.

[3] ADEPT: An Entropy-Driven Dual-Strategy Agent for Interactive Video Retrieval, ICASSP 2026.

[4] MAI: A Multi-turn Aggregation-Iteration Model for Composed Image Retrieval, ICLR 2025.

[5] CoCo-IR: Contextual Composed Image Retrieval, ECCV 2026.

[6] Diverse-Intent Multi-Turn Fashion Image Retrieval, arXiv:2607.20291, 2026.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.