Paper deep dive

Beyond Single Videos: Benchmarking and Active Evidence Seeking for E-Commerce Cross-Video Reasoning

Jinghan Zhao, Yiman Hu, Liang Wu, Jian Xu, Bo Zheng

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 59

Intelligence

Status: not_run | Model: - | Prompt: - | Confidence: 0%

Entities (0)

No extracted entities yet.

Relation Signals (0)

No relation signals yet.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:E-commerce videos are information-dense and frequently compared by consumers evaluating products and merchants assessing marketing strategies. However, existing multimodal models mainly focus on single-video understanding and have limited ability to compare information across videos. We introduce AdsCVR, the first e-commerce cross-video reasoning benchmark, containing 2,483 videos and 6,110 question-answer pairs across six reasoning dimensions. Cross- video reasoning requires models to locate fine-grained evidence among many redundant frames and integrate visual details, speech, and on-screen text. We therefore propose AdSeek, an agentic framework that dynamically selects visual and audio tools during multi-turn exploration, replacing static uniform sampling with active evidence acquisition. To address the sparse credit assignment of reinforcement learning, we develop an offline trajectory rectification mechanism that identifies reasoning errors and missing multimodal evidence in RL-generated trajectories. The corrected trajectories provide supervised fine-tuning signals that reduce biases learned during RL. This mechanism supports a rectified bootstrapping pipeline in which initial RL exposes reasoning bottlenecks, supervised fine-tuning corrects them, and a final RL stage further improves the policy. AdSeek achieves 74.30 percent accuracy on the AdsCVR test split, outperforming its Qwen3-VL-8B-Instruct backbone by 27.90 percentage points. It also generalizes to the open- domain CrossVid benchmark, demonstrating effective active evidence gathering.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

55,490 characters extracted from source content.

Expand or collapse full text

Beyond Single Videos: Benchmarking and Active Evidence Seeking for E-Commerce Cross-Video Reasoning Jinghan Zhao Yiman Hu Liang Wu Jian Xu Bo Zheng Alibaba Group Abstract E-commerce videos are information-dense and frequently compared in practice, as consumers com- pare products and merchants evaluate marketing tactics, making cross-video analysis highly valuable for real-world business analytics. Yet, existing multimodal models primarily focus on single video un- derstanding, with limited ability to perform the cross-video comparisons needed for e-commerce video analysis. To bridge this gap, we introduceAdsCVR, the first e-commerce cross-video reasoning bench- mark. After filtering, AdsCVR contains 2,483 videos and 6,110 QA pairs across six reasoning dimen- sions. Extending reasoning across multiple videos fundamentally magnifies the challenge of pinpointing fine-grained evidence amid massive redundant frames, especially in e-commerce videos where visual details, speech, and on-screen text are combined into densely packed content. We proposeAdSeek, an agentic framework that dynamically allocates visual and audio tools during multi-turn exploration, re- placing static uniform sampling with active evidence grounding. While reinforcement learning (RL) can optimize sequential tool use, sparse outcome rewards struggle with long-horizon credit assignment, providing limited guidance on where a trajectory goes wrong or how to correct it. To address this, we propose anoffline trajectory rectification mechanismthat identifies and corrects reasoning errors and missing multimodal evidence in RL-generated trajectories. It then uses them for SFT to strengthen the model’s evidence gathering and reasoning abilities, thereby mitigating biases learned during RL. This mechanism anchors ourrectified bootstrapping pipeline: an initial RL phase exposes reasoning bottle- necks, the subsequent SFT phase uses rectified data to calibrate model’s thinking and action biases, and a final RL phase achieves optimal convergence. AdSeek achieves 74.30% accuracy on the AdsCVR test split, improving upon its backbone (Qwen3-VL-8B-Instruct) by 27.90%. It also generalizes well to the open-domain CrossVid benchmark, validating its effectiveness in active evidence gathering. 1 Introduction Many similar products on e-commerce platforms are promoted through different videos, presenting differ- ent marketing strategies. These videos often pack product demonstrations, selling points, and promotional messages into just tens of seconds through visuals, speech, and on-screen text. Consumers and merchants often need to compare this dense multimodal information across videos to inform purchasing and marketing decisions. However, existing advertising benchmarks such as M-AU, AdsQA, and E-VAds [ Bose et al., 2023,Long et al.,2025,Liu et al.,2026] predominantly focus on single-video understanding and do not adequately capture this practical need or the challenges of identifying key differences across videos. Therefore, we proposeAdsCVR, the first e-commerce cross-video reasoning benchmark, evaluating models’ ability to understand complex e-commerce videos and draw effective comparisons across multi- ple videos. Unlike general-domain CVR benchmarks such as CrossVid and MVU-Eval [ Li et al.,2026,Peng et al.,2025], AdsCVR focuses on commercial reasoning over advertisements, where decisive evidence is scattered across videos, non-contiguous temporal intervals, and visual and audio modalities amid distracting dense content. 1 arXiv:2610.03099v1 [cs.CV] 2 Oct 2026 T MLLM Fixed Context Target Video single pass (a) Traditional Video Understanding Advice: ❌ 1. video quality is low 2. .... 3. ... Query:give some advice to this advertisement video ? Target Video Ref video1 Ref video2 (b) Comparative Video Understanding and Active Evidence Seeking frame_select(V1, 12–16s) → Display only frame_select(V2, 3–8s) → Applied to lips frame_select(V3, 0–4s) → Swatched on arm You can add more on-skin effect.... ✅ Query Observation Context Plan Tool More Precise GroundingRicher Reference Value “only $30” Figure 1:Single-videounderstandingversusactivecross-videounderstanding.(a) Fixed-context, single- pass analysis may produce generic advice. (b) AdSeek leverages richer cross-video references and produce more precise evidence grounding for more specific and actionable advice. To improve models’ ability to understand and compare e-commerce videos, we proposeAdSeek, an agentic framework for multi-turn active evidence acquisition. Unlike static uniform sampling [Wu et al., 2024], AdSeek dynamically selects which videos, temporal segments, modalities, and visual regions to in- spect based on the comparison objective. It filters out irrelevant content and aligns evidence across videos before generating an answer. Following tool-augmented language modeling [ Yao et al.,2023,Schick et al.,2023,Yang et al.,2023, Hao et al.,2023], we build a multi-turn interaction environment for AdSeek and train its evidence-search policy through reinforcement learning (RL) [Qian et al.,2025]. During RL training process, sparse outcome rewards struggle with long-horizon credit assignment [ Arjona-Medina et al.,2019], failing to provide explicit feedback for intermediate tool calls or reasoning decisions. So, error patterns that emerge during exploration may persist even with continued RL training. To address this challenge, we introduce anoffline trajectory rectification mechanism. It uses a teacher model to identify intermediate deviations that remain undetected by outcome-based RL feedback and rec- tify them by supplying missing evidence and correcting faulty reasoning. The corrected trajectories directly superviseintermediate evidenceacquisitionand reasoning, helping the policy model internalize these correc- tions via SFT. Building on this, ourRectified Bootstrapping Pipeline (RBP)uses initial RL exploration to expose bottlenecks, offline trajectory rectification to correct persistent biases, and subsequent RL to further refine the policy through online exploration. Our contributions include: •We introduceAdsCVR, an e-commerce cross-video reasoning benchmark comprising 2,483 videos and6,110QApairsspanning86productcategories(seeTable1foracomparison). Itevaluatesmodels’ understanding of complex commercial analysis and their ability to identify fine-grained differences across information-dense videos. •We proposeAdSeek, an agentic framework that actively acquires and aligns multimodal evidence through multi-turn reasoning and tool use, adapting its search across videos, temporal segments, modalities, and visual regions to each comparison objective. •WedeveloptheRectifiedBootstrappingPipeline(RBP),whichcombinesonlineRLexplorationwith 2 offline trajectory rectification. By correcting biased trajectories and enabling further exploration, this training paradigm strengthens the model’s ability to acquire multimodal evidence. •AdSeek yields a 74.30% accuracy on the AdsCVRtest split, outperforming the Qwen3-VL-8B-Instruct baseline by an absolute margin of 27.90%. Beyond e-commerce, it exhibits robust transferability on the open-domain CrossVid dataset, proving its broader capacity for proactive multimodal evidence collection. 2 Related Work Video Understanding Benchmarks Advances in MLLMs have driven video QA evaluations [Lin et al.,2024,Li et al.,2024] from single-video contentandcausalreasoning[ Yuetal.,2019,Xiaoetal.,2021]tolong-contextunderstanding[Fuetal.,2025, Zhouetal.,2025,Wuetal.,2024,Chengetal.,2025]andegocentrictasks[Huangetal.,2024]. Beyondsingle videos, benchmarks like CrossVid [Li et al.,2026], MVU-Eval [Peng et al.,2025], and CVBench [Zhu et al., 2025] evaluate joint understanding of video collections, requiring models to aggregate distributed evidence and compare entities across videos. E-commerce Advertisement Video Understanding Advertisement understanding requires connecting audiovisual content with persuasive intent. Early works focused on message symbolism [Hussain et al.,2017], temporal segmentation, and classification [Wang et al.,2021,Bose et al.,2023]. Recent QA benchmarks further evaluate visual reasoning in fast-paced ads [ Zhang et al.,2025], core messages and target audiences [Long et al.,2025], and conversion-oriented e- commerce elements [Liu et al.,2026]. Crucially, these benchmarks evaluate videos individually. In contrast, AdsCVR connects advertisement understanding with cross-video reasoning, requiring models to align cross- ad evidence to compare product presentations. Table 1:Comparison of video understanding benchmarks.Temporal and Intent denote explicit temporal and intent understanding coverage (✓); —: not explicitly targeted. MCQ: multiple-choice QA; Open: open- ended QA; Seg.: segmentation; Cls.: classification. BenchmarkDomain #Videos #QA Video Count Task format Temporal Intent MVU-Eval [Peng et al.,2025] General 4,959 1,824 MultipleMCQ✓— CrossVid [Li et al.,2026]General 5,331 9,015 Multiple MCQ / Open ✓— Tencent-AVS [Wang et al.,2021] Ads12,000 N/A Single Seg. + Cls.✓— M-AU [Bose et al.,2023]Ads8,399 N/A SingleCls.✓ VideoAds [Zhang et al.,2025]Ads200 1,100 SingleMCQ✓ AdsQA [Long et al.,2025]Ads1,544 7,859 SingleOpen—✓ E-VAds [Liu et al.,2026]E-commerce 3,961 19,785 SingleOpen—✓ AdsCVRE-commerce2,4836,110MultipleMCQ✓ MLLMs for Video Understanding Video MLLMs acquire temporal reasoning by training visual encoders and language models on multimodal instructions [ Lin et al.,2024,Li et al.,2024,Zhang et al.,2024,Cheng et al.,2024,Wang et al.,2024b]. 3 However, early systems rely on uniform frame sampling, risking the omission of brief, decisive events. To address this, recent agentic and tool-using approaches make perception adaptive, iteratively retrieving or selecting keyframes based on query relevance and reasoning feedback [Wang et al.,2024a,Jeoung et al., 2024,Wang et al.,2025b,He et al.,2026,Zhang et al.,2026]. AgentCVR extends this active acquisition to cross-video reasoning via script-simulated RL [Qiu et al.,2026]. Together, these approaches shift video understanding from fixed frame sampling toward adaptive evidence acquisition guided by the question and intermediate reasoning. 3 The AdsCVR Benchmark We introduce the AdsCVR benchmark to evaluate fine-grained e-commerce video understanding. We de- fine acomparable video groupas a set of videos sharing contextual dimensions (e.g., narrative subject or commercial intent) to expose subtle distinctions in advertising strategies. Figure2illustrates our three-stage construction pipeline. 3.1 Video Filtering and Multi-View Annotation Video filtering.To construct a diverse benchmark of manageable size, we first preprocess the videos to support subsequent filtering: We extract video metadata, speech transcripts (Whisper-large-v3), and shot boundaries(TransNetV2). Tofilteroutoutliers, weestimatetemporalandspatialcomplexitiesusingDINOv3 embeddingsandJPEGcompression(seeAppendix Cformetricdetails). Basedonthesemeasures, wediscard videos with extreme durations or low visual variation. Finally, we balance the dataset by dropping long-tail categories(fewerthantenvideos)anddownsamplingdominantones, yielding2,483videosacross86product categories (Figure3a). (a) Video Filter and Annotation DINOv3 Whisper Data Enrichment TransNet Events Objects Creative Commercial Multi-view Caption Metadata category, video title, ... ASR speech transcripts Shot segments split video by shots Multi-view Captions events, objects, creative, commercial Information density dcpx, jcpx Category-balanced Sampling Quality Filtering (b) Comparable Video Grouping(c) QA Construction & Verification ICODCSVCTSID Intra-class Comparsion Outlier Detection Creative Strategy Visual Composition Temporal Structure Information Density QA Generation f Which video use a more emotional appeal? A. Video 1 B. Video 2 C. Video 3 D. Video 4 Answer: A, with evidence in .... Question Blind-test Filtering Correct Answer Guess without video Question Evidence Verifiablity Evidence Incorrect Answer Evidence Insufficient Question Difficulty Filtering Too Easy difficulty control AdsCVR TrainTest comparable group Structured Perception Discrete Labels Numerical Statistics Rule-based Matching •Category •... •Shot Segment count •... Select * From pool Where category=’shoes’ AND creative_tag=’Scenario Simulation’ Video 2Video 3Video 4Video 1 Set-similarity Constraints •Creative tags •... Figure 2:AdsCVR construction pipeline.(a)Video Annotation:Raw videos undergo multimodal fea- ture extraction and multidimensional captioning, followed by quality and category-balanced filtering. (b) Comparable Grouping:Structured rules guide the assembly of 2–4 videos into comparable groups across six reasoning dimensions. (c)QA Generation & Filtering:LLMs generate multiple-choice questions for each group. A rigorous multi-stage filtering pipeline (blind-test, verifiability, and diﬀiculty) yields 6,110 high-quality QA pairs (3,110 train, 3,000 test). 4 Multi-view annotation.To go beyond sparse metadata and capture fine-grained semantics, we prompt Gemini to generate dense, multidimensional captions from four expert perspectives:events(actions and demonstrations),objects(products and scenes),creative structure(narratives and filming techniques), and commercial logic (audiences and advertising strategies; see Figure3b). The full multi-view annotation prompt is provided in AppendixA.2. These annotations provide the semantic foundation for the subsequent video grouping and QA generation. 3.2 Video Grouping Weformcomparablegroupsoftwotofourvideosusingdimension-specificsamplingrules(Table2),covering visual execution and commercial intent. Table 2:AdsCVR comparison dimensions and video-grouping rules. DimensionComparison objective and sampling rule Intra-class Comparison (IC)Fine-grained discrimination: sample unique video combinations within the same first-level category. Outlier Detection (OD)Anomaly identification: combineN−1reference videos with one outlier from a different second-level or first-level category, with a50%/50% allocation. Creative Strategy (CS)Commercial intent: include at least two distinct commercial-strategy buckets per group. Visual Composition (VC)Visual presentation: require creative-tag Jaccard similarity<0.5between candidate and seed videos. Temporal Structure (TS)Narrative progression: use segment count as a proxy and span at least two ranges (1–3, 4–6,≥7 segments) per group. Information Density (ID)Visual pacing: sample intra-category videos whose absolute shot counts differ by≥3, independent of duration. 3.3 QA Generation and Verification To construct informative and challenging multiple-choice QA pairs, we employ prompt-based generation followed by a rigorous multistage filtering pipeline. First, forQA generation, Gemini-3.5-Flash generates a single-choice question, options, the correct answer, and supporting textual clues based on aggregated video metadata, multidimensional captions, and dimension-specific grouping rationales. Next, we applyblind-test filteringto mitigate language-prior leakage, discarding any QA pair that Qwen3VL-8B answers correctly in three consecutive attempts using exclusively the text question and options (without video inputs). To ensure answer verifiabilityand grounding in video evidence, we then require GPT-5.5 to correctly predict the reference answer relying solely on the generated textual clues. Finally, fordiﬀiculty filtering, we evaluate the remaining pairs using Qwen3-VL-8B-Instruct with uniformly sampled frames, randomly discarding 50% of the questions it answers correctly to calibrate the overall benchmark diﬀiculty. AdsCVR contains 6,110 QA pairs from 2,483 e-commerce videos across 86 top-level product categories, with QA pairs randomly split into 3,110 training and 3,000 test examples. 4 AdSeek We introduce AdSeek , a framework for active multi-turn evidence acquisition trained through the Rectified Bootstrapping Pipeline, as illustrated in Figure4. 4.1 Task Definition Given a queryqand a video collectionV=V j N j=1 , the task is to generate an answerˆyby iteratively acquiring cross-video evidence via perception tools. 5 (a) Task-category statistics TS CS ID VC IC OD (b) Strategy label word cloud (c) video duration Figure 3:AdsCVR dataset statistics.(a) Associations between the six evaluation dimensions and 12 repre- sentative product categories, with flow width indicating video count. (b) Commercial-strategy word cloud. (c) Total video duration per QA pair across 6,110 examples. Interaction process.The policyΦinitializes with contexth 1 = (q,M), whereMcontains lightweight metadata (e.g., duration, resolution). At round i , the context expands to: h i = (q,M, t 1 , a 1 , o 1 , . . . , t i−1 , a i−1 , o i−1 ),(1) wheret k ,a k , ando k denote the reasoning trace, tool call, and returned textual observation at roundk. Con- ditioned onh i ,Φgeneratest i followed by either a final answerˆyor a tool calla i ∈A. Executinga i yields observationo i , updating the context as: h i+1 = (h i , t i , a i , o i ).(2) Perception tools.The action space comprises visual and audio tools: A=A vis ∪A aud , A vis =frame_select(j, s, e, f), A aud =audio_select(j, s, e), (3) wherej∈1, . . . , Nidentifies the video and0≤s < e≤D j specifies a temporal interval. The prompt f∈Fdefinesspecificvisualdetailstoinspect. Thevisualandaudiotoolsreturnafocusedtextualdescription and a speech transcript of the segment, respectively. Trajectory.This process repeats untilΦproduces an answer or hits the interaction limit, yielding a completedn-round trajectory: τ= (q,M, t 1 , a 1 , o 1 , . . . , t n−1 , a n−1 , o n−1 , t n ,ˆy).(4) 4.2 Exploration: Online Reinforcement Learning (RL) 4.2.1 Policy Optimization We optimize the policy using Group Relative Policy Optimization (GRPO) [Shao et al.,2024], implemented with ROLL [Wang et al.,2025a]. For each inputx, the policy samplesGtrajectoriesτ i G i=1 . Each trajectory 6 1 <think>I should start by looking for the first moment where the phone was mentioned, let’s check Video 1 to confirm </think> <tool_call> “tool_name”:“frame_select”, “parameter”:“video_id”:0,“start”:0, “end”:10,“focus”:“lipstick” </tool_call> Video 1 Video 2 Video 3 <think> Now I get the temporal window, I will check the ASR result to get price of this phone. </think> <tool_call> “tool_name”:“whisper”, “parameter”:“video_id”:0, “start”:5,“end”:20 </tool_call> <think> I get all the clues about question ,and the final result should be option A. </think> <answer> A. video 1 </answer> Question Whisper “this product is very beautiful, it’s match for the students....” video_id, start, end=0,10,20 Optimization Rollout Traj 1 Traj 1 Traj G ... Which video lacks any real-life product application or swatch on skin? Frame Select video_id,start,end,focus=0,0,3,”lipstick” A hand appears on screen picking up a red lipstick. filterRe-rollout (a) Multi-turn Inference Process: Agentic active evidence-seeking framework (b) Training Process: Rectified Bootstrapping Pipeline Rule2: Zero Variance Rule1: No Tool Exploration:Online RL Reward Model Group Computation SFT Data ... t 1 a 1 o 1 t 2 Observation Set o 1 o 2 ... Thinking set ... o 1 t 2 t 1 a 1 o 1 t 2 t 1 a 1 Rectify reasoning Append observation Exploitation: Offline Trajectory Rectification R 1 R 1 R G ... Reasoning Error Observation Insufficient Error Trajectory Trajectory Diagnosis & Calibration Figure 4:AdSeek overview.(a) AdSeek’s multi-turn inference process, where reasoning guides tool use to actively acquire relevant evidence. (b) The two modules of the Rectified Bootstrapping Pipeline: online RL exploration with trajectory filtering, and offline teacher-guided trajectory rectification, which corrects observation and reasoning errors and provides targeted SFT supervision for subsequent exploration. receives a reward combining final-answer correctness (w.r.t. the referencey ⋆ ) and action-format compliance: R(τ i ) =R acc (ˆy i , y ⋆ ) + 0.1R fmt (τ i ).(5) We compute the group-normalized advantage as b A i = (R(τ i )−μ R )/(σ R +ε A ), whereμ R andσ R are the group reward mean and standard deviation. For a generated tokenz i,k with preceding contextc i,k , we define the probability ratior i,k (θ) =π θ (z i,k |c i,k )/π θ old (z i,k |c i,k ). The token-level GRPO objective is then: L GRPO (θ) =−E i,k h min r i,k (θ) b A i ,clip(r i,k (θ),1−ε,1 +ε) b A i −βD KL π θ (·|c i,k )∥π ref (·|c i,k ) i , (6) whereεcontrols policy-ratio clipping andβweights the KL penalty. 4.2.2 Rollout Filtering and Resampling We apply two filtering rules to encourage evidence acquisition and retain informative reward comparisons: (1)No-tool trajectories.We discard trajectories with no effective tool calls,N action = 0. Since video contentisaccessibleonlythroughperceptiontools, thesetrajectoriesprovidenoevidenceofvideoinspection. (2)Zero-variance groups.We discard groupsGwith Var(R(τ) :τ∈G) = 0. Such groups yield zero normalized advantages and therefore no reward-driven policy-gradient signal. They can arise when all trajectories receive the same correctness and format rewards. After filtering, we collect additional rollouts to replenish the training batch, allowing up to three (MAX_TRY=3) resampling attempts. Implementation details are provided in Appendix B. 7 4.3 Exploitation: Offline Trajectory Rectification Outcome rewards provide limited guidance on intermediate failures [Lightman et al.,2024]. Building on feedback-guided refinement [Madaan et al.,2023,Shinn et al.,2023,Havrilla et al.,2024], we introduce offline trajectory rectification to diagnose and correct deviations in the policy’s exploration. This diagnosis targets two core issues: evidence suﬀiciency and reasoning validity. Candidate selection.We select failed training trajectories from the RL-trained policy where a stronger teacher (Φ teacher ) succeeds in the same environment. The teacher’s successful trajectory serves as a reference for identifying evidence gaps and guiding corrections. Observation suﬀiciency check.For a trajectory with observationsO τ =o 1 , . . . , o n−1 , the teacher at- tempts to answer the query using onlyO τ (without tool calls). An incorrect answer acts as a diagnostic signal indicating a potential evidence gap (e.g., missing videos, intervals, or modalities). Reasoning validity check.The teacher reviews each reasoning step against three criteria: (1)Factual consistency(claims must be grounded in observations, avoiding hallucinations), (2)Reasoning complete- ness(conclusions require suﬀicient evidence support), and (3)Historical consistency(alignment with prior rounds). Steps are flagged if any criterion’s risk score (1–5) exceeds 3. Targeted trajectory rectification.We apply corrections based on the diagnosis: (1)Evidence omission: We append missing observations from the teacher’s reference trajectory to the policy’s context, promptingΦ to regenerate intermediate reasoning. (2)Reasoning deviation:The teacher provides corrective comments on flagged steps to guideΦ’s reasoning regeneration. By targeting specific failures within the student’s own context (rather than replacing it entirely with the teacher’s rollout), we produce tailored corrective demon- strations. 4.3.1 Supervised Policy Optimization Following rationale bootstrapping [Zelikman et al.,2022], we constructD SFT from initially successful and newly rectified trajectories, helping the policy internalize corrected evidence acquisition and reasoning. For a tokenized trajectoryz 1:M , a maskm j ∈0,1isolates assistant tokens (reasoning traces, tool calls, and final answers, wherem j = 1) from the context (queries, metadata, and observations, wherem j = 0). The objective is: L SFT (θ) =−E τ∼D SFT " P M j=1 m j logπ θ (z j |z <j ) P M j=1 m j # .(7) The loss is averaged solely over assistant tokens. This applies direct supervision to intermediate steps on fixed trajectories, without invoking perception tools during training. 4.4 Training Paradigm: Rectified Bootstrapping Pipeline We integrate these components into theRectified Bootstrapping Pipeline(RBP), a three-stage paradigm: (1)Initial RLdevelops evidence acquisition and exposes trajectory errors; (2)Offline Rectification & SFT corrects missing evidence and faulty reasoning via teacher guidance, internalized through SFT; (3)Refined RLresumes online optimization. By complementing sparse outcome feedback with direct supervision of intermediate decisions, RBP strengthens the model’s ability to locate, integrate, and compare evidence scat- tered across videos, temporal segments, and modalities, supporting a deeper understanding of information- dense e-commerce videos. 8 5 Experiments Training and Evaluation Settings Our policy backbone is Qwen3-VL-8B-Instruct [Bai et al.,2025], using Qwen3-VL-4B-Instruct and Whisper large-v3 [Radford et al.,2023] for perception, and Qwen3.8-27B for offline attribution and evaluation. We use 16 PPUs for SFT via LLaMA-Factory [Zheng et al.,2024] and 32 PPUs for RL via ROLL [Wang et al., 2025a] (AppendixB). We evaluate accuracy on the AdsCVR test split (3,000 test / 3,110 train from 6,110 QA pairs) and CrossVid for out-of-domain generalization. Baselines include general MLLMs and AgentCVR [Qiu et al.,2026] 1 , a prior cross-video RL framework. 5.1 In-Domain Results on AdsCVR Table3highlights the diﬀiculty of AdsCVR: evaluated open-source single-pass models achieve average ac- curacies of 46.40%–75.15%, leaving substantial room for improvement. Trained only on AdsCVR, AdSeek reaches 73.57%, outperforming its Qwen3-VL-8B backbone by 27.17 percentage points and our AgentCVR reproductionby2.52points, whileapproachingQwen3.8-27B(75.15%). Incorporatingpublicout-of-domain training data further improves accuracy to 74.30%; we analyze mixed-source training in Section 5.3. Average accuracy improves across training stages, with the largest stage-to-stage gain occurring after RL resumes (+2.04 points). Relative to Stage 1, the final mixed-source model improves across all six dimen- sions, most notably Information Density (ID, +5.0 points) and Visual Composition (VC, +4.8 points), where answering requires locating fine-grained evidence in dense visual content or comparing details across shots. Together, these results suggest that AdSeek’s SFT stage preserves overall performance while enabling further improvement during subsequent RL, and that targeted trajectory rectification contributes to higher accuracy on challenging tasks involving dense visual content and cross-shot comparisons. Table 3:Accuracy (%) on the AdsCVR test split. All AdSeek variants use Qwen3-VL-8B. Bold indicates our best results. Average denotes the unweighted mean across dimensions. MethodTS CS ID IC OD VC Average General-purpose MLLM baselines Gemini 3.5 Flash79.3 75.0 76.2 75.6 78.8 81.5 77.73 GPT-5.585.2 80.2 84.3 79.6 82.0 83.1 82.40 Qwen3.8-27B79.7 74.3 74.9 77.1 72.2 72.7 75.15 Qwen3.5-4B64.8 59.0 56.6 61.0 60.6 61.9 60.65 Qwen3.5-9B68.9 64.6 67.1 64.0 64.9 64.5 65.67 Prior agentic baseline AgentCVR (2026)73.7 71.1 68.7 70.3 73.6 68.9 71.05 AdSeek with Qwen3-VL-8B backbone Qwen3-VL-8B (backbone)52.2 46.8 41.6 43.6 48.2 46.0 46.40 AdSeek (Stage 1)74.6 72.4 65.8 72.8 72.6 69.4 71.27 AdSeek (Stage 2)75.0 67.6 68.2 73.6 74.4 70.4 71.53 AdSeek (Stage 3, in-domain only) 75.0 74.4 70.6 72.8 74.6 74.0 73.57 AdSeek (Stage 3, mixed)77.0 73.4 70.8 75.6 74.8 74.2 74.30 Gain vs. backbone (p)+24.8+26.6+29.2+32.0+26.6+28.2+27.90 1 Reproduced in our own environment using the same infrastructure, based on the code released by the authors. 9 5.2 Out-of-Domain Results on CrossVid To assess cross-domain generalization, we evaluate AdSeek on ten tasks from CrossVid, a general-domain cross-video benchmark. As shown in Table4, AdSeek achieves an average accuracy of 31.55%, exceeding its Qwen3-VL-8B backbone by 2.08 percentage points and our AgentCVR reproduction by 3.31 points. It improves over the backbone on six of the ten tasks. These results suggest that AdSeek’s evidence-seeking capabilities transfer beyond e-commerce videos. Table 4:Out-of-domain performance (%) on CrossVid. Average is the unweighted mean over ten tasks; means and gains are computed before display rounding. Bold marks the highest average. MethodBU NC C PEA PI FSA MSR MOC PSS CCQA Average Qwen3-VL-8B23.5 27.4 40.7 34.9 62.5 5.71 34.5 27.2 7.22 31.2 29.47 AgentCVR (2026)27.6 37.5 48.9 30.1 50.2 11.0 22.3 28.4 8.14 18.2 28.24 AdSeek (Stage 3, mixed) 38.9 38.7 47.7 37.5 45.4 16.1 26.4 37.8 6.61 20.431.55 5.3 Ablations on Training Data and Action Space Training data.Table5shows that AdsCVR-only training outperforms MVU-Eval-only training by 9.50 percentage points (73.57% vs. 64.07%), supporting the value of domain-specific data for e-commerce video comparison. Our mixed-source dataset, AdSeek-train, supplements AdsCVR with MVU-Eval videos and script-simulated AgentCVR examples, further improving accuracy to 74.30%. The additional gains are con- centrated in IC (+2.8 points) and TS (+2.0 points), while CS decreases by 1.0 point. Thus, mixed-source training provides a modest overall benefit, with effects that vary across comparison dimensions. Action space.Restricting AdSeek toframe_selectoraudio_selectreduces average accuracy to 63.97% and 65.03%, respectively, corresponding to drops of 10.33 and 9.27 points from the full model. The two modalities exhibit different strengths: visual-only access performs better on IC (63.8% vs. 58.2%), whereas speech-only access performs better on the other five dimensions. Nevertheless, the full model out- performs both restricted variants across all six dimensions, supporting the benefit of combining visual and speech evidence for cross-video comparison. Table 5:Training-data and action-space ablations on AdsCVR test accuracy (%). Average denotes the un- weightedmeanacrossdimensions. MVU-Eval-onlyandrestricted-actionvariantsareevaluatedonthe3,000- question test split, with unparsed answers counted as incorrect. VariantTS CS ID IC OD VC Average Training data In-domain only (AdsCVR) 75.0 74.4 70.6 72.8 74.6 74.0 73.57 MVU-Eval only67.4 60.8 60.8 63.6 65.6 66.2 64.07 Mixed77.0 73.4 70.8 75.6 74.8 74.2 74.30 Action space frame_selectonly 66.2 63.8 61.2 63.8 65.8 63.0 63.97 audio_selectonly 67.4 65.6 65.6 58.2 66.2 67.2 65.03 AdSeek77.0 73.4 70.8 75.6 74.8 74.2 74.30 10 6 Limitations AdSeek involves computational trade-offs associated with multi-turn agentic reasoning. Sequential tool calls and repeated context processing can increase inference latency, while RL rollouts and teacher-guided tra- jectory rectification introduce additional training overhead. Future work could explore more eﬀicient tool scheduling and lightweight perception models to better balance reasoning quality and computational cost, particularly for latency-sensitive applications. 7 Conclusion We introduced AdsCVR, a comprehensive benchmark for fine-grained cross-video e-commerce comparison, alongside AdSeek, an agentic framework for active evidence acquisition. To optimize this framework, we developed the Rectified Bootstrapping Pipeline (RBP), which seamlessly integrates RL exploration with offline trajectory rectification via SFT. AdSeek achieves 74.30% accuracy on AdsCVR and demonstrates robust transferability to open-domain multimodal reasoning tasks. AI Use Statement We used generative AI models as components of the reported research workflow. Gemini models were used to produce multi-view video annotations and candidate AdsCVR questions, answer options, and supporting clues. Qwen3-VL-8B was used for prompt-only and frame-based filtering, GPT-5.5 was used for evidence- based answer verification, and Qwen3.8-27B was used as the teacher for offline trajectory diagnosis and as an evaluation judge where described in the paper. These model outputs were processed by deterministic filters and checked on representative examples by the authors; model-generated outputs were not treated as ground truth without the stated verification steps. The authors reviewed the final manuscript, experimental claims, and released code, and take responsibility for the accuracy, integrity, and originality of all submitted material. References Jose A Arjona-Medina, Michael Gillhofer, Michael Widrich, Thomas Unterthiner, Johannes Brandstetter, and Sepp Hochreiter. RUDDER: Return Decomposition for Delayed Rewards. InAdvances in Neural Information Processing Systems, volume 32, 2019. URL https://arxiv.org/abs/1806.07857. Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-VL Technical Report.arXiv preprint arXiv:2511.21631, 2025. URLhttps://arxiv.org/abs/2511.21631. Digbalay Bose, Rajat Hebbar, Tiantian Feng, Krishna Somandepalli, Anfeng Xu, and Shrikanth Narayanan. M-AU: Towards Multimodal Understanding of Advertisement Videos. InProceedings of the ACM In- ternational Conference on Multimedia, pages 86–95, 2023. URLhttps://arxiv.org/abs/2308.14052. Junhao Cheng, Yuying Ge, Teng Wang, Yixiao Ge, Jing Liao, and Ying Shan. Video-Holmes: Can MLLM Think Like Holmes for Complex Video Reasoning?arXiv preprint arXiv:2505.21374, 2025. URLhttps: //arxiv.org/abs/2505.21374. Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. VideoLLaMA 2: Advancing Spatiotemporal Modeling and Audio Under- standing in Video LLMs.arXiv preprint arXiv:2406.07476, 2024. URLhttps://arxiv.org/abs/2406.07476. 11 Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-MME: The First-Ever Compre- hensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 24108–24118, 2025. URLhttps://openaccess.thecvf.com/content/CVPR2025/html/Fu_Video-MME_The_First-Ever_ Comprehensive_Evaluation_Benchmark_of_Multi-modal_LLMs_in_CVPR_2025_paper.html. Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. ToolkenGPT: Augmenting Frozen Language Mod- els with Massive Tools via Tool Embeddings. InAdvances in Neural Information Processing Systems, volume 36, pages 45870–45894, 2023. URLhttps://proceedings.neurips.c/paper_files/paper/2023/hash/ 8fd1a81c882cd45f64958da6284f4a3f-Abstract-Conference.html. Alex Havrilla, Sharath Raparthy, Christoforus Nalmpantis, Jane Dwivedi-Yu, Maksym Zhuravinskyi, Eric Hambro, and Roberta Raileanu. GLoRe: When, Where, and How to Improve LLM Reasoning via Global and Local Refinements. InProceedings of the International Conference on Machine Learning, 2024. URL https://proceedings.mlr.press/v235/havrilla24a.html. Zefeng He, Xiaoye Qu, Yafu Li, Siyuan Huang, Daizong Liu, and Yu Cheng. FrameThinker: Learning to Think with Long Videos via Multi-Turn Frame Spotlighting. InInternational Conference on Learning Representations, 2026. URLhttps://openreview.net/forum?id=nsNpsCpVG1. Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, et al. EgoExoLearn: A Dataset for Bridging Asyn- chronous Ego- and Exo-Centric View of Procedural Activities in Real World. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22072–22086, 2024. URL https://openaccess.thecvf.com/content/CVPR2024/html/Huang_EgoExoLearn_A_Dataset_ for_Bridging_Asynchronous_Ego-_and_Exo-centric_View_CVPR_2024_paper.html . Zaeem Hussain, Mingda Zhang, Xiaozhong Zhang, Keren Ye, Christopher Thomas, Zuha Agha, Nathan Ong, and Adriana Kovashka. Automatic Understanding of Image and Video Advertisements. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1705–1715, 2017. URL https://openaccess.thecvf.com/content_cvpr_2017/html/Hussain_Automatic_Understanding_ of_CVPR_2017_paper.html. Sullam Jeoung, Goeric Huybrechts, Bhavana Ganesh, Aram Galstyan, and Sravan Bodapati. Adaptive Video Understanding Agent: Enhancing Eﬀiciency with Dynamic Frame Sampling and Feedback-Driven Rea- soning.arXiv preprint arXiv:2410.20252, 2024. URLhttps://arxiv.org/abs/2410.20252. Jingyao Li, Jingyun Wang, Molin Tan, Haochen Wang, Cilin Yan, Likun Shi, Jiayin Cai, Xiaolong Jiang, and Yao Hu. CrossVid: A Comprehensive Benchmark for Evaluating Cross-Video Reasoning in Multimodal Large Language Models. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 6244–6252, 2026. doi: 10.1609/aaai.v40i8.37550. Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. MVBench: A Comprehensive Multi-modal Video Understanding Benchmark. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195–22206, 2024. URL https://openaccess.thecvf.com/content/CVPR2024/html/Li_MVBench_A_Comprehensive_ Multi-modal_Video_Understanding_Benchmark_CVPR_2024_paper.html. Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s Verify Step by Step. InInternational Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=A7x2z0qFUD. 12 Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-LLaVA: Learning United Visual Representation by Alignment Before Projection. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 5971–5984, 2024. doi: 10.18653/v1/2024. emnlp-main.342. Xianjie Liu, Yiman Hu, Liang Wu, Ping Hu, YixiongZou, Jian Xu, and Bo Zheng. E-VAds: An E-Commerce Short Video Understanding Benchmark for MLLMs.arXiv preprint arXiv:2602.08355, 2026. URLhttps: //arxiv.org/abs/2602.08355. Xinwei Long, Kai Tian, Peng Xu, Guoli Jia, Jingxuan Li, Sa Yang, Yihua Shao, Kaiyan Zhang, Che Jiang, Hao Xu, et al. AdsQA: Towards Advertisement Video Understanding. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 23396– 23407, 2025. URLhttps://openaccess.thecvf.com/content/ICCV2025/html/Long_AdsQA_Towards_ Advertisement_Video_Understanding_ICCV_2025_paper.html. Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-Refine: Iterative Re- finement with Self-Feedback. InAdvances in Neural Information Processing Systems, vol- ume 36, pages 46534–46594, 2023. URLhttps://proceedings.neurips.c/paper_files/paper/2023/hash/ 91edff07232fb1b55a505a9e9f6c0f3-Abstract-Conference.html. Tianhao Peng, Haochen Wang, Yuanxing Zhang, Noah Wang, Zili Wang, Ge Zhang, Jian Yang, Shihao Li, Yanghai Wang, Xintao Wang, et al. MVU-Eval: Towards Multi-Video Understanding Evaluation for Multimodal LLMs. InAdvances in Neural Information Processing Systems, volume 38, 2025. doi: 10.52202/085713-1800. Cheng Qian, Emre Can Acikgoz, Qi He, Hongru Wang, Xiusi Chen, Dilek Hakkani-Tur, Gokhan Tur, and Heng Ji. ToolRL: Reward Is All Tool Learning Needs. InAdvances in Neural Information Processing Systems, volume 38, pages 105523–105553, 2025. doi: 10.52202/085713-3524. Yilun Qiu, Jiahe Wang, Cilin Yan, Jiayin Cai, Xiaolong Jiang, Yao Hu, and Chun Yuan. AgentCVR: Ac- tive Multi-Agent Cross-Video Reasoning via Script-Simulated Reinforcement Learning.arXiv preprint arXiv:2605.29643, 2026. URL https://arxiv.org/abs/2605.29643. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision. InProceedings of the 40th International Con- ference on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 28492– 28518. PMLR, 2023. URL https://proceedings.mlr.press/v202/radford23a.html. Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Ham- bro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language Mod- els Can Teach Themselves to Use Tools. InAdvances in Neural Information Processing Sys- tems, volume 36, pages 68539–68551, 2023. URL https://proceedings.neurips.c/paper/2023/hash/ d842425e4bf79ba039352da0f658a906-Abstract-Conference.html. Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models.arXiv preprint arXiv:2402.03300, 2024. URLhttps://arxiv.org/abs/2402.03300. Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Lan- guage Agents with Verbal Reinforcement Learning. InAdvances in Neural Information Processing Sys- 13 tems, volume 36, pages 8634–8652, 2023. URLhttps://proceedings.neurips.c/paper_files/paper/2023/ hash/1b44b878b782e6954cd888628510e90-Abstract-Conference.html. Weixun Wang, Shaopan Xiong, Gengru Chen, Wei Gao, Sheng Guo, Yancheng He, Ju Huang, Jiaheng Liu, Zhendong Li, Xiaoyang Li, et al. Reinforcement Learning Optimization for Large-Scale Learning: An Eﬀicient and User-Friendly Scaling Library.arXiv preprint arXiv:2506.06122, 2025a. URLhttps: //arxiv.org/abs/2506.06122. Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. VideoAgent: Long-Form Video Un- derstanding with Large Language Model as Agent. InEuropean Conference on Computer Vision, pages 58–76, 2024a. doi: 10.1007/978-3-031-72989-8_4. YiWang, KunchangLi, Xinhao Li, JiashuoYu, YinanHe, GuoChen, Baoqi Pei, RongkunZheng, Zun Wang, Yansong Shi, et al. InternVideo2: Scaling Foundation Models for Multimodal Video Understanding. In European Conference on Computer Vision, pages396–416, 2024b. URLhttps://arxiv.org/abs/2403.15377. Zhenzhi Wang, Liyu Wu, Zhimin Li, Jiangfeng Xiong, and Qinglin Lu. Overview of Tencent Multi-Modal Ads Video Understanding Challenge.arXiv preprint arXiv:2109.07951, 2021. URLhttps://arxiv.org/abs/ 2109.07951. Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. VideoTree: Adaptive Tree-Based Video Representation for LLM Reasoning on Long Videos. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3272– 3282, 2025b. URL https://openaccess.thecvf.com/content/CVPR2025/html/Wang_VideoTree_Adaptive_ Tree-based_Video_Representation_for_LLM_Reasoning_on_Long_CVPR_2025_paper.html. Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. LongVideoBench: A Benchmark for Long-Context Interleaved Video-Language Understanding. InAdvances in Neural Information Processing Systems, vol- ume 37, pages 28828–28857, 2024. doi: 10.52202/079017-0907. Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9772–9781, 2021. URLhttps://arxiv.org/abs/2105.08276. Rui Yang, Lin Song, Yanwei Li, Sijie Zhao, Yixiao Ge, Xiu Li, and Ying Shan. GPT4Tools: Teaching Large Language Models to Use Tools via Self-Instruction. InAdvances in Neural Information Processing Systems, volume 36, pages 71995–72007, 2023. URL https://proceedings.neurips.c/paper_files/paper/ 2023/hash/e393677793767624f2821cec8bdd02f1-Abstract-Conference.html. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Rep- resentations, 2023. URLhttps://openreview.net/forum?id=WE_vluYUL-X. Zhou Yu, Dejing Xu, Jun Yu, Ting Yu, Zhou Zhao, Yueting Zhuang, and Dacheng Tao. ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9127–9134, 2019. doi: 10.1609/aaai.v33i01. 33019127. Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. STaR: Bootstrapping Reasoning with Reasoning. InAdvances in Neural Information Processing Systems, volume 35, pages 15476–15488, 2022. URLhttps: //papers.nips.c/paper/2022/hash/639a9a172c044fbb64175b5fad42e9a5-Abstract-Conference.html . 14 Yaolun Zhang, Ruohui Wang, Jiahao Wang, Yepeng Tang, Xuanyu Zheng, Haonan Duan, Hao Lu, Hanming Deng, and Lewei Lu. EVA: Eﬀicient Reinforcement Learning for End-to-End Video Agent. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12289–12299, 2026. URL https://openaccess.thecvf.com/content/CVPR2026/html/Zhang_EVA_ Efficient_Reinforcement_Learning_for_End-to-End_Video_Agent_CVPR_2026_paper.html. Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. LLaVA-Video: Video Instruction Tuning with Synthetic Data.arXiv preprint arXiv:2410.02713, 2024. URLhttps://arxiv.org/ abs/2410.02713. Zheyuan Zhang, Wanying Dou, Linkai Peng, Hongyi Pan, Ulas Bagci, and Boqing Gong. VideoAds for Fast-Paced Video Understanding. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 21812–21821, 2025. URLhttps://openaccess.thecvf.com/content/ICCV2025/html/Zhang_ VideoAds_for_Fast-Paced_Video_Understanding_ICCV_2025_paper.html. Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, and Zheyan Luo. LLaMA-Factory: Unified Eﬀicient Fine-Tuning of 100+ Language Models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 400–410, 2024. doi: 10.18653/v1/2024.acl-demos.38. Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Zhengyang Liang, Shitao Xiao, Minghao Qin, Xi Yang, Yongping Xiong, Bo Zhang, et al. MLVU: Benchmarking Multi-Task Long Video Understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13691–13701, 2025. URL https://arxiv.org/abs/2406.04264. Nannan Zhu, Yonghao Dong, Teng Wang, Xueqian Li, Shengjun Deng, Yijia Wang, Zheng Hong, Tiantian Geng, Guo Niu, Hanyan Huang, et al. CVBench: Benchmarking Cross-Video Synergies for Complex Multimodal Reasoning.arXiv preprint arXiv:2508.19542, 2025. URL https://arxiv.org/abs/2508.19542. 15 A Annotation Guidelines AdsCVR annotations compare two to four e-commerce videos along six dimensions: TS, CS, ID, VC, IC, and OD. Data construction first applies category-balanced quality filtering and timestamped ASR/OCR extrac- tion, then uses Gemini 3.5 Flash to generate structured descriptions and salient tags. Candidate questions are generated as(question,options,answer,clues)tuples. A hierarchical verifier uses Qwen3-VL-8B for prompt-only shortcut and diﬀiculty checks, followed by GPT-5.5 checks for answer completeness, multi- video dependence, logical validity, and multimodal evidence requirements. A.1 Qualitative Examples Across Dimensions Figures5–10show selected cases for each of the six dimensions. Figure 5:Temporal Structure (TS) examples comparing how advertisements organize events and product claims. Figure 6:Creative Strategy (CS) examples comparing how advertisements attract attention and communicate product value. 16 Figure 7:Outlier Detection (OD) examples comparing advertisements across product categories. Figure 8:Intra-class Comparison (IC) examples comparing substantive similarities and differences among advertisements for related products. Figure 9:Information Density (ID) examples comparing the amount and pace of product-related spoken and visual information. 17 Figure 10:Visual Composition (VC) examples comparing framing, editing, transitions, and the functional role of individual shots. 18 A.2 Multi-view Annotation Prompt The following prompt is used to generate the four-view annotations described in Section3.3. The placehold- ers in the user prompt are filled with the corresponding metadata for each video. System prompt. SYSTEM_PROMPT_ALL = ( "You are a multi-role video synthesis analyst. Analyze the advertising video in depth by acting as each of the following four roles. " "[General requirements (strictly followed by all roles)] " "- Ground every fact: describe only content clearly visible in the video or reasonably inferable; do not invent dialogue, background stories, or off-screen elements " "- Segmentation: prioritize the [shot time segmentation] in the auxiliary information. If it is unavailable, segment by action changes, shot transitions, or scene changes; uniform time segmentation is prohibited " "- Caption length for each segment: 20-50 words " "[Role 1: Event analyst (temporal-event view)] " "Describe the main events occurring in each shot: " " Temporal dynamics: specific actions, motion trajectories, human interactions, object manipulation, and state changes " " Spatial context: scene environment and positions of key objects; describe them only when they change " "[Role 2: Scene analyst (object-and-space view)] " "Describe the people, objects, and scene in each shot: " " Spatial details: setting, spatial relations among objects (left/right, above/below, front/behind, near/far), people's positions and postures, color and material properties, and composition " " Key objects: objects relevant to the video topic or the current frame " "[Role 3: Creative analyst (cinematographic view)] " "Analyze the filming and narrative techniques of each shot: " " Cinematic language: camera movement (push/pull, pan, follow, or static), shot scale (close-up, medium, or wide), and camera angle (high, low, or eye-level) " " Narrative structure: the function of the shot in the overall script (introduction, demonstration, conflict, transition, or ending) " " Emotional tone: observable emotional style and pacing characteristics " "[Role 4: Commercial analyst (holistic view, no segmentation)] " "Make an overall commercial judgment about the video: " " audience: target audience, formatted as [price positioning + group], no more than 20 words and without a qualifier prefix " " Correct example: [mid-to-low-priced consumer goods, students and ordinary office workers] " " Incorrect example: [consumers who value the user experience and dislike a sticky feeling] " " ad_logic: core marketing logic, explaining in one sentence how the advertisement drives the purchase decision " "Output 3-5 additional tags for each dimension: short phrases, each no more than 10 words, representing the core category or key point of that dimension. " "Output all four dimensions at once, strictly in JSON format, with no content outside the JSON." ) User prompt. USER_PROMPT_ALL = """The following auxiliary information is provided for the video: ASR transcript: asr 19 Product title: item_title Video title: content_title Merchant name: shop_name Video description: summary Shot time segmentation: shot_detect Product category: item_cate_level1_name / item_cate_level1_name2 Watch the video frame by frame. Strictly follow the role requirements in the system instruction and use shot time segmentation as the basis for segmentation. If shot time segmentation is unavailable, use action changes, shot transitions, or scene changes. Output the analysis for all four dimensions at once. Output format (strict JSON, with no other content): "events":"segments":["start":0,"end":5,"caption":"Person A walks to the display, picks up the product, and the camera follows the hand movement"],"tags":["product demonstration","human interaction"],"objects":"segments":["start":0,"end":5, "caption":"A white display is centered, the product is on the left side of the surface, and a person stands in the right foreground"],"tags":["indoor setting", "product close-up"],"creative":"segments":["start":0,"end":5,"caption":"A static medium shot from an eye-level angle introduces the product"],"tags":["static camera","product introduction"],"commercial":"audience":"Mid-to-low-priced consumer goods for students and ordinary office workers","ad_logic":"A realistic usage scenario lowers the decision barrier and encourages impulse purchase","tags":["mass-market consumption","scenario-based promotion"] """ A.3 Dataset Split and Video-Overlap Audit AdsCVR is split randomly into training and test partitions. To obtain a stricter test subset for cross-video transfer, we separately extracted test examples and performed a video-overlap audit against the training par- tition before evaluation. The resulting strictly non-overlapping subset contains 1k samples. Table6aligns the audited test-subset accuracy with the mixed-source AdSeek result across the six dimensions. The high accuracy retained by the audited subset, together with the explicit overlap review, supports the robustness of our method to cross-video transfer. Table 6:Accuracy (%) on the strictly audited test subset and with mixed-source AdSeek across the six AdsCVR dimensions. AdsCVR is randomly split, and the separately extracted test subset is audited for video overlap with the training partition before evaluation. VariantTS CS ID IC OD VC Average Strictly audited test subset 86.7 78.5 78.5 75.6 74.6 78.5 78.73 AdSeek (mixed)77.0 73.4 70.8 75.6 74.8 74.2 74.30 B Implementation Details Table7summarizes the SFT hyperparameters used for AdsCVR, and Table8reports the corresponding RL configuration. The default SFT launch uses 16 PPUs; the RL run uses 32 PPUs. 20 Table 7:SFT hyperparameters for AdsCVR. HyperparameterValue Base modelQwen3-VL-8B-Instruct Fine-tuningFull-parameter SFT; FlashAttention 2; Liger kernel Templateqwen3_vl_nothink Cutoff length51,200 tokens Per-device batch size1 Gradient accumulation2 Learning rate1×10 −5 Scheduler / warmupCosine / warmup ratio 0.1 Epochs5 Precision / memoryBF16; gradient checkpointing CheckpointingSave every 10 steps; keep all checkpoints Distributed training16 PPUs Table 8:RL hyperparameters for AdsCVR. HyperparameterValue Base modelQwen3-VL-8B-Instruct Training steps / checkpoint interval500 / 20 steps Rollout / validation batch size512 / 256 Learning rate / scheduler5×10 −7 / constant Weight decay / warmup 1 × 10 −2 / 50 steps Per-device batch / accumulation1 / 16 Advantage estimatorGRPO; no advantage whitening PPO epochs1 KL regularizationKL loss coeﬀicient10 −2 ; initial KL coeﬀicient 0.001 Entropy coeﬀicient0.01 Reward weights Accuracy 1.0; format 0.1 Rollout groupsgroup size 8 Filtering / retriesMAX_TRY= 3 PrecisionBF16 Distributed training32 PPUs C Video Complexity Metrics We detail the temporal and spatial complexity proxies used for video filtering in Section 3.1. Frame sampling and temporal complexity.We sample frames at approximately 1 fps and uniformly retain at most 16 frames per video. For theTselected frames, letu t =z t /(∥z t ∥ 2 +δ)be the normalized DINOv3 feature, withδ= 10 −12 . We compute: c t = 1− P j∈N t w tj u ⊤ t u j P j∈N t w tj +δ , Dcpx(V) = T X t=1 c t ,Vcpx(V) = 100 T Dcpx(V), (8) 21 whereN t =j: 1≤j≤T,0<|j−t|≤d,w tj =exp(−|j−t|/(2d)), andd= 8. Both metrics are set to zero forT≤1. Higher Dcpx and Vcpx indicate greater cumulative and average visual variation, respectively. Spatial complexity.For the sameTselected framesI 1 , . . . , I T , we estimate spatial complexity using the average JPEG-encoded bytes per pixel: Jcpx(V) = 1 T T X t=1 B JPEG (I t ) H t W t ,(9) whereB JPEG (I t )is the encoded size in bytes andH t ×W t is the frame resolution. We use quality 95, 4:2:0 chroma subsampling, and optimized coding. Higher Jcpx indicates lower spatial compressibility under these fixed encoding settings. 22

Instant research discovery