{"id":"0cfb63eb-d9aa-4b4c-aa91-ec6cd6c076fd","arxiv_id":"2502.07096","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Lotus combines AI-generated abstractive narration with extractive original clips to help creators repurpose long videos into short-form videos, with positive reception in a small user study.","lead":"Lotus is a video editing system that turns long videos into short-form clips by combining newly written AI narration with original footage. A small user study suggests creators found it faster and more enjoyable than their usual tools, though the quality difference was not statistically significant.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The initial abstractive draft's clip matching is unvalidated; if GPT-4o scoring is noisy, user-study success may reflect manual repair rather than system quality.","rationale":"Agree with the reader's conditional verdict. The system is well-motivated and the formative study is thoughtful, but the central claim rests on an unvalidated automatic pipeline and a small self-report study. The weakest link is the clip-transcript matching: it is the first place the pipeline commits to content, and it is exposed to LLM scoring bias. The user study does not isolate the initial draft's contribution; the interaction logs suggest heavy manual revision. A human-rating check of the matching, plus an analysis of initial-clip retention, would directly test whether the system's initial output or the user's repair effort drives the reported benefits. The paper's other limitations (non-significant mixed-method preference, small n, no released artifacts) reinforce, rather than replace, this concern. Therefore the conditional verdict should stand.","tokens_in":23998,"tokens_out":6121,"duration_ms":55454,"concrete_test":"Independently validate the clip-transcript matching: take the initial abstractive drafts for the six results-evaluation videos, extract the (clip, visual-concept) pairs that Lotus assigns, and have at least three human raters score each pair's visual-narration alignment on a 1-5 scale. Compare mean human scores against GPT-4o's scores using rank correlation (e.g., Spearman rho). Additionally, from the user-study logs, compute the proportion of initial abstractive clips that appear unchanged in each participant's final video; if fewer than half of initial clips survive without edits, the initial draft's contribution to the final product is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim implicitly requires that the automatically generated initial short-form video is a useful starting point that reduces creator effort. Section 4.3.3 selects long-form clips for each visual concept using a weighted score where GPT-4o rates roughly 25 candidate clips per concept based on keyframes and speech. These scores are not validated against human judgments, and the paper reports no measure of how often initial clips are retained or replaced during user editing. Table 1 shows participants performed many trim and delete operations (e.g., P1: 65 trims, 6 deletes; P7: 157 trims, 4 deletes), consistent with substantial manual repair. If GPT-4o's image-text similarity judgments are noisy or biased, the initial abstractive draft would be poorly matched, and the positive user-study results (all participants willing to use Lotus; final videos 'comparable in quality') could largely reflect participants' tolerance for fixing a weak initial draft rather than the system's summarization quality. This is load-bearing because the paper's design goal G1 ('Identifying Key Clips') and the claimed benefit of reduced effort depend on the initial draft being useful.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Lotus is a system that helps creators repurpose long-form videos into short-form videos by combining abstractive summarization (generated narration with matched visuals) and extractive summarization (original clips). The paper reports a formative study with professional editors, a two-stage generation pipeline (short-form transcript generation, visual concept extraction, clip-transcript matching with a weighted scoring function, and blending of abstractive and extractive clips), and an editing interface. Evaluation consists of a results evaluation (12 raters, 6 videos, three conditions: abstractive, extractive, and mixed) and a within-subjects user study (8 participants) comparing Lotus to participants' existing editing tools. The paper claims that the mixed method was more preferred than either pure method (though not significantly), that all participants would use Lotus in the future, and that participants created videos comparable in quality to those made with existing tools.","tokens_in":24337,"tokens_out":3229,"duration_ms":30692,"significance":"The paper addresses a genuinely important and under-served problem: the time-consuming work of repurposing long videos into short-form social media content. The proposed combination of abstractive and extractive summarization is a reasonable and timely design idea, and the system is described in unusual detail, including full prompts in the appendix, the use of publicly available models, and a clear interface description. If the effectiveness claims survive revision, Lotus would be a valuable contribution to the video-authoring and intelligent-user-interfaces literatures. The main strengths are the formative study that grounds the design goals, the explicit separation of abstractive and extractive capabilities with user control, and the open reporting of the pipeline. The main weaknesses are the lack of validation of the automatic clip-transcript matching and the over-interpretation of non-significant quantitative results.","major_comments":[{"comment":"The clip-transcript matching is a core load-bearing component that supports design goal G1 and the claim that the initial draft reduces creator effort, but the paper provides no evidence that GPT-4o's scoring of candidate clips agrees with human judgments of relevance. The user study interaction log in Table 1 shows very high numbers of trim operations (e.g., P7: 157 trims, P1: 65 trims), which is consistent with substantial manual repair of the initial draft. Without a measure of how often the initially matched clips were retained or replaced, the reader cannot tell whether the positive user reactions reflect the quality of the automatic draft or the participants' tolerance for fixing a weak draft. Please add either a human-rated validation of the matching component (e.g., relevance ratings for matched versus alternative clips) or a quantitative retention/replacement analysis from the logged user interactions.","section":"§4.3.3 and Table 1"},{"comment":"The headline result of the results evaluation, that the mixed method is 'more preferred' than the abstractive or extractive methods, is not statistically significant: with 12 raters, the mean ranks are 1.88, 2.06, and 2.04. The abstract and introduction state that the results evaluation 'demonstrates the benefit of flexibility,' but with no inferential test and no effect size, this is an overstatement. Please report a proper analysis (e.g., Friedman test with post-hoc comparisons, or a Bayesian equivalent), and adjust the wording of the conclusions to describe the pattern as suggestive and exploratory rather than demonstrative.","section":"§5, Figure 5"},{"comment":"The user study makes comparative claims, such as participants created videos of comparable quality 'without experiencing increased mental demand' and 'liked the editing process within Lotus over their existing tools,' but no statistical tests are reported for the TLX, Likert, or CSI ratings. With n=8, these claims are currently supported only by descriptive statistics and error bars. Please report paired non-parametric tests (e.g., Wilcoxon signed-rank) for the key comparisons, or explicitly limit those claims to qualitative observations from the interviews.","section":"§6.3, Figures 6 and 7"},{"comment":"The blending of abstractive and extractive segments depends on an 'empirically determined' threshold and a 'coherence score, determined by GPT-3 [26] loss,' but the paper does not describe the threshold value, the procedure used to set it, or how the GPT-3 loss is computed and applied. Because the mixed method is the central contribution, the reader cannot reproduce or independently assess the blending step. Please specify the threshold, the optimization or selection process, and the exact use of GPT-3 loss.","section":"§4.3.4"}],"minor_comments":[{"comment":"The sentence 'This pane consists consists of the Original Video Player' contains a duplicated word; please fix.","section":"§4.2.1"},{"comment":"The phrase 'Appendix Tableg 2' should be 'Appendix Table 2'.","section":"§5, Table 2"},{"comment":"The first sentence of 'Integration With Existing Tools' repeats the same idea twice ('Lotus is not as fully functional as traditional video editing tools... does not provide a comprehensive suite of features'); please consolidate.","section":"§7"},{"comment":"The text contains many instances of 'fexibility' and 'efect' that appear to be spelling errors; please proofread to replace these with 'flexibility' and 'effect'.","section":"Throughout"},{"comment":"The formula for position-based alignment is typeset in a corrupted format ('????? ??? = 1 − ??? (visual concept)− ??? (long-form clip)'); the actual equation needs to be rendered correctly.","section":"§4.3.3"},{"comment":"The prompts in the appendix are a valuable contribution, but the 'GPT-4o Scoring' prompt would benefit from a short sentence explaining how the visual concept is embedded in the speech window, since the user message references an '[embedded visual concept]' that is not defined in the prompt text.","section":"§9.1"}],"recommendation":"major_revision","confidential_remarks":"This is a solid IUI-style systems paper with a clear contribution and unusually transparent reporting of prompts and pipeline details. The two main issues—lack of validation of the automatic clip-matching component and over-interpretation of non-significant results—are fixable with additional analysis (clip retention statistics, appropriate statistical tests, and softened claims). The paper is within the scope of the journal and, with careful revision, could be a valuable addition. I do not see a fundamental flaw that would require rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Lotus combines abstractive and extractive summarization into one editing workflow, and that combination is genuinely new in the cited literature. The paper deserves a serious referee. The central preference result isn't significant, and the automatic clip-matching step is under-validated, but the system is clearly described and the user study gives real qualitative value.\n\nWhat's new: prior work does either extraction or abstraction. Lotus gives creators both modes per clip, with toggling, alternative clip suggestions, and automated blending. The formative study with five editors and 20 video pairs is done carefully and yields sensible design goals. The results evaluation compares three methods across six videos and shows preferences vary by video type; the discussion of why extractive works for talking heads and abstractive for travel is reasonable. The user study, though small, shows all eight participants wanted to use Lotus in the future and valued the initial generated draft as a starting point. That is meaningful evidence that the combination helps.\n\nSoft spots: the headline result that mixed was more preferred than abstractive or extractive is not statistically significant. With 12 raters and 6 videos, the error bars overlap. The user study has only 8 participants. No code or data are released, so reproduction is limited. The stress-test concern about GPT-4o scoring is fair: the clip-transcript matching in Section 4.3.3 depends on LLM judgments that are never validated against human ratings, and interaction logs show participants did a lot of trimming and deleting. The paper doesn't report how often initial clips were kept. That said, this is not a load-bearing flaw. The system is explicitly designed as a starting point for refinement, and participants confirmed they used it that way. The unvalidated matching would be a bigger problem if the paper claimed the automatic output was final, but it doesn't.\n\nWho is this for: HCI researchers in video authoring and summarization, and anyone building AI-assisted editing tools. It's a useful contribution to the subfield, not a paradigm shift. The writing is honest about limitations.\n\nRecommendation: send it to peer review. A good referee will ask for a larger preregistered study and some validation of the matching step, but the core idea and qualitative results justify a revision cycle.","headline":"A solid HCI systems paper that genuinely combines abstractive and extractive video summarization; the headline preference result isn't significant and the matching pipeline needs validation, but the work deserves review.","tokens_in":24722,"tokens_out":1855,"would_cite":true,"duration_ms":17619,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Lotus claims that combining AI-written narration with original footage lets creators make short videos from long ones in minutes, with results they rate comparable to their usual tools.","keywords":["video summarization","video editing","audio editing","authoring tools","generative AI","abstractive summarization","extractive summarization"],"falsifier":"Take a set of long videos, run Lotus's automatic clip-transcript matching, and have independent human annotators judge each assigned clip against its transcript sentence; if a substantial fraction of assignments are judged mismatched or the human agreement with GPT-4o's scores is low, the automatic draft's quality claim is falsified. A second test: run the full user study with a 'broken draft' condition in which the initial script is paired with randomly chosen clips, and measure whether satisfaction and final quality remain comparable.","tokens_in":23830,"feed_emoji":"🎬","tokens_out":4984,"duration_ms":41734,"temperature":0.7,"pith_summary":"Lotus is a system for turning a long video into a short one by combining two approaches that are usually kept separate: abstractive summarization (a new script is written and spoken by a synthetic voice) and extractive summarization (original clips are cut and stitched together). The paper claims that creators can use this mix to get the best of both: the flexibility to cover what matters in the shorter format, and the credibility of the original audio-visual connection. In an eight-person user study, every participant said they would use Lotus again, and they rated the videos they made with it comparable in quality to videos made with their regular tools. The paper also reports, from a ranking study, that the blended abstractive-extractive output was preferred over either pure method, though the difference was not statistically significant. If the claim holds, it suggests short-form repurposing can be substantially automated while leaving the human in control of the final cut.","feed_headline":"Creators accept AI-mixed short videos as equal to their own edits","feed_subtitle":"All eight editors in the study wanted to use Lotus again, and rated its output on par with their existing tools.","key_machinery":"The load-bearing object is the clip-transcript matching and blending pipeline. For matching, each long-form clip is scored against a visual concept from the short-form script using four signals: speech similarity via nomic-embed-text embeddings, keyframe similarity via CLIP, a GPT-4o perceptual score that rates roughly 25 candidate clips from their keyframes and speech, and position-based alignment that rewards clips whose relative position in the long video matches the concept's relative position in the script. For blending, abstractive and extractive segments are compared on speech similarity, the CLIP visual connection between a segment's speech and its keyframe, noun-phrase coverage, and position; when the best extractive match is close, the system generates permutations and picks the one with the highest coherence as measured by GPT-3 loss. This machinery turns an abstractive draft into a mixed video and gives the editing interface concrete alternatives to offer the user.","core_discovery":"Lotus's central claim is that an editing system can productively combine abstractive and extractive summarization in one workflow, and that creators will prefer working this way over starting from scratch. The pipeline first generates a short-form transcript from the long-form transcript with GPT-4o, extracts visual concepts, matches long-form clips to those concepts with a weighted scoring function, and synthesizes narration with ElevenLabs to produce an initial abstractive draft. It then scores abstractive segments against extractive segments and, where the difference is small, lets the user choose between versions so the final video can blend newly written narration with original footage. The user study found all eight participants wanted to use Lotus in the future, found the generated draft a useful starting point, and rated their Lotus-made videos comparable in quality to those from their existing tools; the results evaluation found the mixed method preferred, though not significantly, with preferences varying by video type.","pith_inferences":["The positive user-study results have not been separated from the quality of the automatic clip-transcript matching; a controlled study in which the initial draft is replaced by a random or mismatched draft would show how much of the value comes from the draft versus the interface.","The genre-dependence found in the ranking study suggests an adaptive default: the system could decide per video whether to start abstractive, extractive, or mixed.","Because participants wanted a horizontal timeline and finer trimming, the approach could be packaged as a module inside existing editors rather than as a standalone tool.","The reliance on a synthetic voice was the most-cited weakness; improvements in expressive speech generation would likely shift more creators toward abstractive clips."],"forward_implications":["Creators can produce a usable short-form draft in far less than the hours-to-days manual editors in the formative study reported.","The mixed approach is not uniformly best: for videos with a prominent on-camera narrator, extractive or mixed clips win, while for narrator-light footage abstractive summaries are preferred.","The initial abstractive draft serves as a scaffold: participants kept many of its clips and used alignment and search to add original footage around it.","Switching a clip between abstractive and extractive modes gives creators a way to fix narration they find unnatural without losing the visual.","If the preference for the mixed method holds beyond the study's sample, short-form tools should offer the blend as a default rather than forcing one summarization strategy."],"supporting_citations":[{"why":"GPT-4o is used to generate the short-form transcript, extract visual concepts, score candidate clips, and segment transcripts, making it the backbone of the abstractive side.","marker":"[8]"},{"why":"ROPE supplies the extractive baseline and the sentence-segmentation and combinatorial-optimization method that Lotus re-implements for extraction and blending.","marker":"[68]"},{"why":"CLIP computes keyframe-to-concept similarity in clip matching and the visual-connection score in blending.","marker":"[59]"},{"why":"nomic-embed-text provides the speech similarity between long-form clips and target visual concepts.","marker":"[51]"},{"why":"PySceneDetect performs the scene detection that segments the long-form video into clips for matching.","marker":"[28]"},{"why":"ElevenLabs synthesizes the narration for abstractive clips from the generated short-form transcript.","marker":"[5]"},{"why":"spaCy extracts noun phrases used in prompt construction and in the coverage score during blending.","marker":"[34]"},{"why":"GPT-3 loss is used to compute the coherence score that selects among blended clip permutations.","marker":"[26]"}],"fun_headline_variants":["AI video editor blends two summarization styles for better shorts","Creators rate AI-mixed shorts as good as their own edits","Lotus pairs abstractive and extractive video editing to streamline shorts","AI system helps creators mix narration and clips for short videos","Video editing tool combines two summarization approaches for shorter clips"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that GPT-4o's perceptual scoring of which long-form clips match the written narration is good enough that the initial abstractive draft is genuinely usable; if those judgments are unreliable, the draft would be mismatched and the positive user results would partly reflect participants' willingness to repair a weak starting point.","fun_headline_variants_meta":{"raw":{"variants":["AI video editor blends two summarization styles for better shorts","Creators rate AI-mixed shorts as good as their own edits","Lotus pairs abstractive and extractive video editing to streamline shorts","AI system helps creators mix narration and clips for short videos","Video editing tool combines two summarization approaches for shorter clips"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000934,"raw_usage":{"total_tokens":3994,"prompt_tokens":937,"completion_tokens":3057,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":2972}},"tokens_in":553,"tokens_out":3057,"duration_ms":19020,"temperature":1.0,"reasoning_tokens":2972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:48:14.908800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of long videos, run Lotus's automatic clip-transcript matching, and have independent human annotators judge each assigned clip against its transcript sentence; if a substantial fraction of assignments are judged mismatched or the human agreement with GPT-4o's scores is low, the automatic draft's quality claim is falsified. A second test: run the full user study with a 'broken draft' condition in which the initial script is paired with randomly chosen clips, and measure whether satisfaction and final quality remain comparable.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GPT-4o is used to generate the short-form transcript, extract visual concepts, score candidate clips, and segment transcripts, making it the backbone of the abstractive side."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ROPE supplies the extractive baseline and the sentence-segmentation and combinatorial-optimization method that Lotus re-implements for extraction and blending."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PySceneDetect performs the scene detection that segments the long-form video into clips for matching."},{"cited_title":"ElevenLabs","cited_arxiv_id":null,"evidence_quote":"ElevenLabs synthesizes the narration for abstractive clips from the generated short-form transcript."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"spaCy extracts noun phrases used in prompt construction and in the coverage score during blending."}],"review_version":1}