{"id":"b35986cb-c9a1-4fbd-b875-2ef82b04ea06","arxiv_id":"2608.01660","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GCR is a training-free frame-selection pipeline that binds subtitle text to its aligned frame, fills the budget with diverse context, then replaces weak context frames with medoids from omitted regions, improving long-video QA by 2 to 3 points.","lead":"Long-video question-answering models must pick a handful of frames from hours of footage. This paper introduces GCR, a training-free selector that grounds subtitles onto their source frames, covers diverse visual context, and then refines the selection by revisiting omitted parts of the video, yielding consistent accuracy gains on two benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (7)'s empirically determined d_q prior is unreleased and never swept; the reported gains could be benchmark-specific calibration rather than evidence-curation principle.","rationale":"The reader's weakest_assumption identifies the same risk: the empirically determined d_q prior is a hidden tunable knob, and the paper gives no evidence that GCR's advantage survives without benchmark-specific calibration. My reading of the full text confirms this is the least secure part of the central claim. The controlled comparisons in Table 3, the leave-one-stage-out ablations in Table 2, and the budget-ratio sensitivity in Figure 5 all support the pipeline's internal soundness, but they deliberately avoid varying d_q and tau_g. The manuscript openly states d_q rules are in the supplementary, which is not provided, and tau_g is never given a value. This is a reproducibility and generalizability weakness, not a demonstrated failure. The reader's CONDITIONAL verdict is therefore appropriate, and my stress test does not move it.","tokens_in":12809,"tokens_out":2637,"duration_ms":26034,"concrete_test":"Release the d_q keyword rules from the supplementary and rerun the full LLaVA-OV B=32 evaluation on both benchmarks with d_q fixed to 0, 0.5, and 1, keeping all other pipeline settings identical. If the 2.54 pp and 1.93 pp advantages over the strongest baselines collapse or reverse under any fixed value, then Eq. (7)'s empirically determined prior is load-bearing and the generalizability claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that GCR is a general, training-free evidence-curation framework delivering consistent gains. The most load-bearing condition is that Eq. (7) genuinely fuses textual and visual evidence in a general way. But the fusion weight d_q is 'empirically determined according to the visual and textual reliance implied by keywords within each question stem', with detailed rules only in the missing supplementary material. The Sensitivity and Efficiency section sweeps only the budget ratios alpha_g, alpha_v, alpha_r (Figure 5); it never varies d_q, and tau_g in Eq. (17) is named but its value is never reported. Thus the paper does not show that the 2.54 pp and 1.93 pp headline gains on LongVideoBench and Video-MME are robust to this prior. If d_q was hand-tuned to the question patterns of these two benchmarks, the reported advantage could reflect benchmark-specific calibration rather than a general evidence-curation principle. The pipeline itself may be sound, but the generality of the empirical claim is not secured without either releasing the rules or showing the method is insensitive to d_q.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GCR, a training-free, fixed-budget frame-selection framework for long-video QA. It operates in three stages: Ground (convert timestamped subtitles/ASR into events, score them by a text-visual fusion, and render selected event text onto the corresponding real frame), Cover (add top visual anchors and fill the remaining budget with global MMR-based diverse frames), and Refine (summarize omitted temporal regions by real-frame medoids and conditionally swap them for the weakest revisable context frame). The method uses frozen SigLIP features, keeps a chronological real-frame set of size B, and requires no VLM training or architectural change. Experiments on LongVideoBench and Video-MME with LLaVA-OV, Qwen2.5-VL-7B, and LLaVA-Video-7B under budgets of 8, 32, and 64 report consistent overall gains over reproduced baselines, with the headline LLaVA-OV/32-frame results of 64.25% and 62.15%. The paper also reports leave-one-stage-out ablations, an evidence-integration study with blank-band and shuffled-overlay controls, category-wise analysis, sensitivity to budget ratios, and efficiency measurements.","tokens_in":13175,"tokens_out":5545,"duration_ms":47517,"significance":"If the results hold, GCR is a practically valuable contribution: it improves long-video QA without training, is architecture-agnostic, and directly addresses the revisability and text-visual alignment limitations of prior selectors. The paper's controlled experimental design is a genuine strength: same budget, VLM, and decoding configuration for all methods; leave-one-stage-out ablations; a blank-band control that isolates the effect of adding image area; a shuffled-overlay control that isolates temporal alignment; and a budget-sensitivity sweep. The Refine stage's marginal but positive contribution (0.22/0.45 pp) and the low selector-side cost (0.141 s/QA) are reported honestly. The central limitation is that two important knobs of the method—the question-dependent fusion prior d_q (Eq. 7) and the refinement gate tau_g (Eq. 17)—are either described only by reference to the supplementary material or never given a value, and neither is subjected to sensitivity analysis. Because the paper's main empirical claim is generality across benchmarks and backbones, this gap is load-bearing.","major_comments":[{"comment":"The visual-demand prior d_q controls the fusion of textual and visual event scores in the Ground stage, yet the paper only says it is \"empirically determined according to the visual and textual reliance implied by keywords within each question stem\" and defers the rules to the supplementary material, which was not available for review. This is load-bearing because d_q directly determines which events are grounded; if the keyword rules were derived from the question patterns of LongVideoBench and Video-MME, the reported gains (e.g., 2.54 pp and 1.93 pp) could reflect benchmark-specific calibration rather than a general evidence-curation principle. Please include the full d_q rules in the main text or appendix, and add an analysis that either sweeps d_q or evaluates on held-out question categories, to demonstrate that the method is not brittle to this prior.","section":"Methodology, Eq. (7)"},{"comment":"The refinement gate threshold tau_g is named and used to decide whether refinement is activated, but its value is never reported or swept. Since the Refine stage is a claimed contribution and the ablation attributes only 0.22/0.45 pp to it, the reader cannot tell whether this small gain is robust or is the result of a particular threshold choice. Please report the default tau_g and include it in the sensitivity analysis; Figure 5 currently sweeps only alpha_g, alpha_v, and alpha_r.","section":"Refine stage, Eq. (17)"}],"minor_comments":[{"comment":"The claim that GCR \"consistently achieves the best overall performance\" is supported at the dataset level, but several subgroup cells are below the strongest baseline (e.g., LLaVA-OV, B=8, LongVideoBench Long: 51.77 vs. WFS-SB 53.90; Qwen2.5-VL-7B, B=8, Short: 67.87 vs. WFS-SB 68.70). Please add a note clarifying that the consistency claim refers to overall accuracy, not every duration split.","section":"Table 1"},{"comment":"The figure caption reports \"Overall span\" in percentage points, while the y-axis label reads \"Accuracy (%)\"; please clarify that the plotted values are LongVideoBench overall accuracy with LLaVA-OV at B=32, and state that the horizontal line is the strongest reproduced baseline.","section":"Figure 5"},{"comment":"In Table 2, the \"w/o Ground\" and \"Full GCR\" rows both report 51.77 for the Long/8-frame condition; if this is correct, it is consistent with the small overall contribution of the later stages at this budget, but it should be noted explicitly.","section":"Table 2"},{"comment":"The text reports a 1.593s total-latency increase for GCR over WFS-SB, but the displayed totals (29.487 vs. 27.893) give a difference of 1.594s; please align the rounding.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the experimental methodology is generally careful. The main uncertainty is whether d_q and tau_g were tuned on the evaluation benchmarks; I would ask the authors to release the rules and add sensitivity analysis for both. If they can do so without substantially changing the results, I would expect the paper to be acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: GCR is a real improvement—it consistently beats uniform sampling and several strong selectors across two benchmarks and three 7B backbones under fixed budgets. The core ideas are the pixel-level rendering of timestamped event text onto its source frame, and the Refine stage that revisits omitted temporal regions and swaps in a real-frame medoid only when it improves evidence value. Those are not just recombinations of old tricks; they work, and the ablations back them up.\n\nWhat's genuinely good: the controlled integration study (visual-only vs. blank band vs. shuffled overlay vs. aligned overlay) isolates the effect of aligned rendering cleanly. The leave-one-stage-out ablations show each stage contributes. The sensitivity analysis on budget ratios shows the default allocation sits in a broad stable region. That's the right way to do an engineering paper.\n\nWhere it gets soft: d_q in Eq. (7) is 'empirically determined' by keyword rules that live only in the supplementary, and tau_g in Eq. (17) is never instantiated. The paper never sweeps d_q or reports tau_g. So the headline 2.5-point gains could partly reflect benchmark-specific calibration of d_q. The stress-test note is right that this is a gap. It's not fatal—the pipeline could still be sound with a fixed d_q—but the paper as written does not secure the generality claim. There's also no statistical significance testing, and the supplementary is unavailable in this version, so reproducibility is incomplete.\n\nWho it's for: researchers working on long-video QA or frame selection. They'll get a practical method and a clean evaluation template. The d_q issue is addressable: release the rules, add a sensitivity sweep, or argue the default is principled.\n\nRecommendation: send it to peer review. It's a solid contribution with one murky hyperparameter; a good reviewer will push for transparency, but that's exactly what peer review is for.","headline":"GCR is a solid, well-ablated frame selector; the hidden d_q prior is a real reproducibility gap that should be fixed before acceptance.","tokens_in":13550,"tokens_out":3628,"would_cite":true,"duration_ms":27350,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GCR treats fixed-budget frame selection for long-video question answering as a joint evidence-curation problem, grounding timestamped text onto its source frames, covering visual gaps, and refining the selection by swapping in stronger…","keywords":["long-video question answering","frame selection","evidence curation","timestamped subtitles","maximal marginal relevance","training-free","vision-language models","fixed budget"],"falsifier":"Run GCR on a fresh long-video QA benchmark whose question stems are written to avoid the keyword vocabulary used to set the visual-demand prior, and compare against the strongest baseline; if the margin collapses or reverses, the prior rather than the grounded-overlay pipeline is responsible for the gains.","tokens_in":12648,"feed_emoji":"🎞️","tokens_out":7923,"duration_ms":64460,"temperature":0.7,"pith_summary":"The paper argues that fixed-budget frame selection for long-video question answering should be treated as a joint evidence-curation problem, not a single-pass top-K retrieval. It introduces GCR, a training-free pipeline that first grounds timestamped subtitles or ASR text onto the specific frames they describe, then adds query-relevant visual anchors and diverse context under a global diversity criterion, and finally revisits omitted temporal regions, swapping in a real-frame medoid only when it is genuinely more valuable than the weakest current frame. Across two long-video QA benchmarks, three 7B vision-language backbones, and frame budgets of 8, 32, and 64, GCR consistently beats uniform sampling and reproduced baselines, with the largest gains on long-duration videos. A sympathetic reader should care because the method improves downstream question answering without retraining a vision-language model, changing its architecture, or increasing the frame budget.","feed_headline":"Three-stage frame picker lifts long-video QA without retraining","feed_subtitle":"Grounding subtitles onto frames and revisiting skipped regions consistently beats baselines under a fixed frame budget.","key_machinery":"The load-bearing objects are: timestamped text events converted from subtitles or ASR, each scored by fused textual and visual relevance and anchored to the highest-scoring real frame in its interval; event-aligned rendering, which paints the event text onto a bottom band of its own anchor frame, binding language to vision in the same image; a protected evidence core of grounded and visual anchors, surrounded by revisable context frames chosen by global maximal marginal relevance, which balances query relevance against pairwise visual redundancy; and omitted-evidence recovery, which summarizes each skipped temporal gap by a real-frame medoid and swaps it in only when its normalized relevance, novelty, and local-change score exceeds the weakest context frame's. These mechanisms together enforce the fixed budget, chronology, and training-free constraint while making the selection revisable.","core_discovery":"The central discovery is that textual and visual evidence reinforce each other best when timestamped language is rendered directly into the image slot of its temporally aligned frame, rather than supplied as a separate prompt, and that a fixed-budget frame set can be made revisable after initial construction. The paper demonstrates through controlled ablations that aligned subtitle overlay beats visual-only input, global subtitle prompts, frame-tagged prompts, and shuffled overlay, and that removing any of the three stages lowers accuracy. It further claims that revisiting omitted temporal regions with real-frame medoids yields consistent, if modest, gains while keeping the downstream VLM input fixed at exactly B chronologically ordered real frames.","pith_inferences":["Beyond the paper, replacing the hand-coded visual-demand prior with a learned or fixed default on fresh question distributions would test how much of the gain is pipeline versus benchmark-specific fitting.","Beyond the paper, the pixel-level text-binding idea could extend to other aligned modalities, such as audio-event captions or detected-object labels, potentially covering videos without usable subtitles.","Beyond the paper, the one-for-one replacement rule is a general principle for any fixed-budget evidence set, applicable to retrieval-augmented generation or extractive summarization.","A quantitative stress test the paper does not run: questions deliberately contradicted by subtitle text, forcing the visual pathway to dominate; the reported category gains suggest reasoning and OCR benefit most, but this adversarial split remains untested."],"forward_implications":["With a 32-frame budget, GCR lifts a 7B backbone to 64.25% on LongVideoBench and 62.15% on Video-MME, surpassing the strongest reproduced baseline by 2.54 and 1.93 percentage points respectively.","Gains grow on long-duration videos: at 64 frames it adds 5.85 points over the baseline on LongVideoBench-Long and 6.33 points on Video-MME-Long.","The improvement is not from added text: aligned overlay outperforms both global subtitle prompts and frame-tagged prompts, and a blank-band control shows the gain is not a layout artifact.","The default budget allocation is stable across a broad range, with overall accuracy varying by at most about one percentage point under the tested ground, visual, and refine ratios.","Selector-side cost stays modest: the refine stage adds about 0.1 seconds per QA while giving a 0.45-point improvement."],"supporting_citations":[{"why":"Supplies global maximal marginal relevance, the diversity criterion that fills the revisable context set in the Cover stage.","marker":"Carbonell and Goldstein 1998"},{"why":"Supplies the frozen vision and text encoders used for frame-query relevance, visual similarity, and medoid computation.","marker":"Zhai et al. 2023"},{"why":"Provides LongVideoBench, one of the two evaluation benchmarks and the source of timestamped subtitle data.","marker":"Wu et al. 2024"},{"why":"Provides Video-MME, the second benchmark spanning short, medium, and long videos.","marker":"Fu et al. 2025"},{"why":"AKS, an adaptive keyframe-sampling baseline that GCR must outperform under identical budgets.","marker":"Tang et al. 2025"},{"why":"WFS-SB, the strongest reproduced baseline on the 32-frame LLaVA-OV setting, from which the paper reports its main margin.","marker":"Chen et al. 2026"},{"why":"VSI, the language-guided visual-subtitle selector that GCR contrasts with aligned overlay rendering.","marker":"He et al. 2026"},{"why":"Provides LLaVA-OneVision-7B-OV, one of the three frozen downstream VLMs used in the main experiments.","marker":"Li et al. 2024"}],"fun_headline_variants":["GCR: training-free frame selection boosts long-video QA","Revisiting skipped frames improves long-video QA","Ground, cover, refine: better frame choices for video QA","No retraining, just better frames: GCR for long-video QA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the empirically set visual-demand prior, which uses hand-coded keyword rules to decide how much each question leans on visual versus textual evidence, transfers to new questions; if those rules were tuned to the benchmark question patterns, the reported gains would not generalize.","fun_headline_variants_meta":{"raw":{"variants":["GCR: training-free frame selection boosts long-video QA","Revisiting skipped frames improves long-video QA","Ground, cover, refine: better frame choices for video QA","No retraining, just better frames: GCR for long-video QA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001017,"raw_usage":{"total_tokens":4302,"prompt_tokens":966,"completion_tokens":3336,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":3264}},"tokens_in":582,"tokens_out":3336,"duration_ms":19877,"temperature":1.0,"reasoning_tokens":3264,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:04:48.933368+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GCR on a fresh long-video QA benchmark whose question stems are written to avoid the keyword vocabulary used to set the visual-demand prior, and compare against the strongest baseline; if the margin collapses or reverses, the prior rather than the grounded-overlay pipeline is responsible for the gains.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"AKS, an adaptive keyframe-sampling baseline that GCR must outperform under identical budgets."}],"review_version":1}