{"id":"8a0ed0eb-43c2-4a67-b200-010ac725c521","arxiv_id":"2605.09223","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"CREST selects video frames using curvature-adaptive non-maximum suppression on CLIP relevance scores, beating AKS by ~0.5% on two long-video QA benchmarks while using a fraction of MIRA's preprocessing cost.","lead":"The paper introduces CREST, a training-free frame-selection method that uses the local curvature of query-frame relevance to decide where to sample frames in long videos. It reports small accuracy gains over AKS on two benchmarks and near-parity with the stronger MIRA pipeline at a fraction of preprocessing cost, but the headline comparison mixes frame budgets and the evidence has weaknesses.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MIRA accuracy-retention claim is confounded by unequal frame budgets (M=32 vs 64), making the abstract's 93–95% figure not an apples-to-apples comparison; the AKS advantage is the only budget-controlled claim.","rationale":"The reader's verdict is CONDITIONAL, and I agree that conditions are necessary. However, I identify the most load-bearing concern as the MIRA frame-budget confound rather than the curvature assumption. The AKS comparison (CREST > AKS at fixed M=32) is budget-controlled and, if reproducible, supports the core algorithmic claim. The MIRA comparison is prominently advertised in the abstract yet is confounded by unequal frame budgets; the stated retention range is also numerically inconsistent for VideoMME. This is a direct threat to the headline efficiency claim, not an abstract generalization worry. The proposed concrete test—running CREST at M=64—would settle whether the retention claim holds under equal budgets. The reader already flagged this issue in the rationale but set 'weakest_assumption' to curvature; I partially diverge by prioritizing the confound. My recommendation is UNCHANGED because the reader's CONDITIONAL verdict already accommodates this concern as a condition (equal-budget rerun).","tokens_in":14307,"tokens_out":10680,"duration_ms":107172,"concrete_test":"Run CREST with M=64 on LongVideoBench and VideoMME using the same backbone (LLaVA-Video-7B) and CLIP scorer, and compare accuracy directly with MIRA at M=64. Compute the per-benchmark retention ratio; if it falls outside the claimed 93–95% range, the abstract's claim is misleading. Also report the exact retention for each benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim's second half—'retaining 93–95% of MIRA's accuracy at only 3–4% of its preprocessing cost'—is not a controlled comparison. In Table 1, MIRA is evaluated with M=64 frames while CREST uses M=32, a difference the authors explicitly acknowledge in the Limitations. The 93–95% retention is therefore a combined effect of selection quality and frame budget, not a pure accuracy/efficiency trade-off. The abstract's phrasing 'Under a fixed backbone and frame budget' implies both comparisons share a budget, which is false for the MIRA comparison. A quick arithmetic check also shows the retained accuracy is 93.4% on LVB but 98.3% on VideoMME, not the '93–95%' stated, adding imprecision. This is load-bearing because the efficiency-accuracy trade-off is the paper's key selling point; if the retention at M=64 were materially different, the headline claim would need substantial qualification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CREST, a training-free frame selection method for long-video question answering. CREST computes CLIP-based query–frame relevance scores, estimates local temporal curvature via second-order finite differences, and performs greedy selection with curvature-adaptive non-maximum suppression: high-curvature peaks receive narrow suppression radii to preserve dense evidence, while flat regions receive broader suppression. An exponential decay of suppression radii over selection steps allows previously suppressed regions to re-enter the candidate pool. Experiments on LongVideoBench, VideoMME, and a newly introduced TempRel diagnostic report that CREST outperforms AKS under a fixed backbone and frame budget, and retains 93–95% of MIRA's accuracy at 3–4% of MIRA's preprocessing cost. Pairwise LLM-as-a-judge evaluations also report higher win rates for CREST-selected frames.","tokens_in":14612,"tokens_out":6272,"duration_ms":67022,"significance":"If the central claims hold, the paper offers a simple, parameter-light geometric heuristic that improves over a relevance–coverage baseline at no training cost and with deterministic behavior. The method's reliance on an external CLIP scorer and a fixed decay parameter (ρ=0.5) is honest, and the authors make code and a diagnostic benchmark available — strengths worth noting. However, the headline MIRA retention claim is currently confounded by unequal frame budgets, and the ablation table contains an internal inconsistency that undermines the claimed complementary contributions of curvature and decay. The AKS advantage, while consistent across budgets and scorers, is numerically small and reported without significance measures. The paper is a plausible incremental contribution pending these corrections.","major_comments":[{"comment":"The headline claim that CREST 'retains 93–95% of the accuracy of MIRA' is not a controlled comparison. In Table 1, CREST uses M=32 frames while MIRA uses M=64, a difference the authors acknowledge in the Limitations. The abstract's phrase 'Under a fixed backbone and frame budget' does not apply to this comparison. Moreover, the stated range is arithmetically wrong: LVB retention is 60.21/64.5 = 93.3% and VideoMME retention is 65.04/66.2 = 98.2%, so the honest range is 93–98%. This is load-bearing because the accuracy/efficiency trade-off is the paper's key contribution. Please correct the abstract and Figure 1, and either run CREST at M=64 or explicitly qualify the retention as computed under unequal budgets.","section":"Abstract; Table 1; Limitations"},{"comment":"In the ablation study, the rows 'w/o curvature' and 'w/o both' are identical across all scorers and both datasets (e.g., 59.31, 58.12, 58.79, 63.67, 64.04, 63.89). This is inconsistent with the text's claim that 'curvature modulation and temporal decay contribute independently and complementarily.' If decay had an independent effect, removing both components should degrade performance relative to removing curvature only. The identical numbers suggest either an implementation bug, a copy-paste error in the table, or that temporal decay has no effect when curvature is removed. This directly affects the interpretation of the ablation and must be investigated and corrected.","section":"Table 5; §3.3"},{"comment":"The advantage of CREST over AKS is +0.45% on LongVideoBench and +0.56% on VideoMME. The paper states 'Results are deterministic across multiple runs,' which rules out run-to-run variance, but benchmark accuracy is still a sample of questions. These small margins could be driven by a few items. A per-category breakdown or a paired bootstrap over questions would demonstrate that the improvement is consistent rather than concentrated in a small subset. As the AKS comparison is the paper's only fully budget-controlled claim, the missing variability analysis is a substantive gap.","section":"§3.2, Table 1"},{"comment":"The LLM-as-a-judge protocol is biased toward verbosity and detail enumeration. The judge prompt explicitly instructs the judge to score higher for 'more exhaustive listing' and 'better readiness for follow-up questions,' and to 'assume every statement in both descriptions is already strictly limited to visible frames.' CREST's event-centric sampling may produce longer, more detailed descriptions by construction, so the win rates (60.58% and 54.50%) may reflect detail richness rather than factual grounding. The paper should validate the judge with human agreement or an analysis of whether the judged preferences correlate with factual accuracy (e.g., precision of statements) rather than only coverage.","section":"§3.5, Appendix B.2"}],"minor_comments":[{"comment":"The instruction 'restores t for t newly outside N(j, R_j^(n))' is ambiguous. It should specify that the restored score is the original normalized relevance value, and clarify the interaction with the zeroing operation in line 9, especially when a restored score has been re-zeroed by a later selected peak.","section":"Algorithm 1, line 13"},{"comment":"The caption states both methods are evaluated 'under the same frame budget,' but Table 1 shows MIRA at M=64 and CREST at M=32. If Table 2 indeed used the same budget, clarify which budget and why; otherwise, correct the caption to avoid confusion.","section":"Table 2 caption"},{"comment":"TempRel is introduced as a new diagnostic benchmark but the main text gives no details on its construction (number of videos/questions, source, annotation procedure, or the exact definition of the ER and HR regimes). Please add a description or point clearly to the supplementary material, since the TempRel results are part of the paper's contribution.","section":"§3.4, TempRel"},{"comment":"Several reference entries are incomplete, e.g., 'Cheng and 1 others', 'Liu and 1 others', 'Islam and 1 others'. These appear to be truncated author lists and should be completed for camera-ready.","section":"References"},{"comment":"The decay parameter ρ is fixed at 0.5 without tuning. The paper acknowledges this, but a small sensitivity table (e.g., ρ ∈ {0.3, 0.5, 0.7}) would make the 'no tuning' claim more credible and characterize the method's robustness, as the Limitations already call for.","section":"Limitations / ρ sensitivity"},{"comment":"The '93–95%' figure appears in both the abstract and Figure 1. As noted in Major Comment 1, the actual retention is 93.3% (LVB) and 98.2% (VideoMME). Please update all instances to reflect the correct range.","section":"Figure 1 and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper has a sound core idea and the AKS comparison is the strongest controlled result. The MIRA retention claim is overstated in the abstract, and the ablation inconsistency (Table 5) is the most concerning internal issue. I recommend major revision: the authors should fix the ablation, correct the headline numbers, and add significance or robustness analysis for the AKS margins. The LLM-judge evaluation should be validated or substantially caveated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely new and clean: instead of just taking score magnitude or global coverage, CREST computes the second difference of the query–frame relevance curve and uses it to shrink the suppression radius around sharp peaks. That is a simple, parameter-light way to concentrate a fixed frame budget on brief, decisive events. The paper also does a few things right: no fitted parameters beyond the fixed ρ=0.5, robustness checks across three relevance scorers, and an honest Limitations section that admits the MIRA comparison is not budget-controlled. That transparency counts for something.\n\nThe main AKS comparison is plausible. CREST beats AKS by about half a point on both benchmarks under the same backbone and budget. But half a point with no error bars or significance test is a thin margin, even if the runs are deterministic. Determinism just means the same code gives the same number; it does not tell you whether the difference would survive a different video sample or a different random seed in the scorer.\n\nThe bigger problem is the MIRA claim. CREST passes 32 frames to the MLLM; MIRA gets 64. The abstract says \"under a fixed backbone and frame budget\" and then reports retaining 93–95% of MIRA's accuracy, implying an apples-to-apples trade-off. The authors admit in Limitations that the 93–95% figure combines selection quality and budget difference, so as stated the headline is misleading. Also, the arithmetic does not support \"93–95%\": on VideoMME the retention is about 98%, not 95%. That should be corrected.\n\nThe ablation table has a red flag that needs explaining: \"w/o both\" is numerically identical to \"w/o curvature\" in every single cell, including all three scorers and both datasets. That should not happen unless removing curvature already removes the effect of decay, which would be surprising. The temporal decay also does not consistently help—removing it improves LVB/BLIP and V-MME/Sevila in Table 5. That inconsistency deserves a discussion, not just a sentence saying decay is beneficial.\n\nTwo smaller weaknesses. First, the main results table reports AKS at 59.76 on LVB, but Table 4 lists AKS with CLIP at 58.41 and BLIP at 59.76. If the main experiments use CLIP, the baseline number appears to have been swapped with the BLIP column. That needs to be checked. Second, the LLM-as-a-judge protocol instructs the judge to assume every statement in both descriptions is grounded in visible frames and to score for exhaustiveness. That rewards longer, more detailed descriptions, so the win rates are not strong evidence of better grounding.\n\nTempRel is a sensible idea, but the paper gives no details on how it was built and it is not released. That makes the 6.88% improvement hard to interpret.\n\nWho is this for? Researchers working on long-video frame selection, especially lightweight training-free methods. It deserves a serious referee—send it to review, but with a clear expectation that the authors fix the budget confound, rerun MIRA at M=32, add error bars, explain the ablation identity, and release code and TempRel before acceptance. If those are fixed, this could be a useful baseline for the subfield.","headline":"Simple, clean frame-selection heuristic with a plausible but small win over AKS under matched budgets; the flashier MIRA-retention claim is confounded by unequal frame budgets and the abstract overstates it.","tokens_in":15067,"tokens_out":2607,"would_cite":true,"duration_ms":31847,"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":"CREST claims that the local curvature of query–frame relevance scores, not their magnitude alone, determines where informative frames lie, enabling a training-free selection rule that rivals costlier retrieval pipelines.","keywords":["frame selection","long-video understanding","curvature","query-conditioned relevance","non-maximum suppression","training-free","temporal geometry","video question answering"],"falsifier":"Replace the CLIP relevance scores on a fixed video set with smoothed or high-frequency-noisy versions (keeping frames and queries identical) and compare CREST vs AKS accuracy: if CREST's advantage disappears or reverses when curvature no longer tracks genuine events, the mechanism's dependence on that premise is confirmed. A simpler check: on a video where the decisive information is a slow, low-curvature transition, CREST should miss it relative to top-K selection.","tokens_in":14171,"feed_emoji":"🎬","tokens_out":7391,"duration_ms":61173,"temperature":0.7,"pith_summary":"CREST is a training-free frame selection method for long-video understanding that treats query–frame relevance as a temporal signal and uses its local curvature to guide a greedy non-maximum-suppression loop. Sharp curvature marks brief decisive events, so the suppression radius contracts there to keep dense evidence; flat regions get broader suppression to avoid redundancy. Under a shared backbone and frame budget, the paper reports higher accuracy than the AKS relevance–coverage baseline on LongVideoBench and VideoMME, and 93–95% of the accuracy of the stronger MIRA pipeline at only 3–4% of its preprocessing cost. If these results hold, a simple geometric rule can replace expensive multi-stage retrieval for video question answering.","feed_headline":"Curvature-guided frame sampling rivals costly retrieval at 3% cost","feed_subtitle":"Curvature of relevance scores picks frames that beat a baseline and nearly match multi-stage retrieval.","key_machinery":"The central mechanism is curvature-regulated non-maximum suppression. Relevance scores are normalized and a discrete second-difference κ_t = |s_{t+1} − 2s_t + s_{t−1}| is computed. The baseline suppression radius R_base = T/M is divided by (1+κ_i) for each selected peak, so high-curvature event peaks keep a narrow suppression window and flat regions a wide one. An exponential decay with rate λ = ln 2/(ρM) progressively restores suppressed scores, letting the budget spread across the whole video.","core_discovery":"The paper's central claim is that local temporal geometry of query-conditioned relevance scores, not just score magnitude, determines where informative frames lie. CREST computes the discrete second derivative κ_t = |s_{t+1} − 2s_t + s_{t−1}| on normalized CLIP relevance scores, then performs greedy selection with suppression radius R_i = R_base/(1+κ_i), so high-curvature event peaks retain dense neighboring evidence and flat regions are suppressed more broadly. An exponential decay on past suppression radii lets suppressed regions re-enter the candidate pool, spreading selections across the full video. The reported result is that this simple rule outperforms the AKS relevance–coverage basel","pith_inferences":["The curvature heuristic may transfer beyond CLIP relevance scores to any temporal proxy—attention weights, object-detection confidence, optical-flow magnitude—where sharp transitions often mark state changes.","Because the paper compares CREST at 32 frames against MIRA at 64 frames, the reported accuracy gap is confounded by budget; a head-to-head at M=64 would likely close part of the gap and could also expose whether CREST's cost advantage grows with budget.","The decay parameter ρ is fixed at 0.5 without tuning; adapting it to video length or event density is a natural testable extension that could improve robustness on unusual temporal distributions.","The TempRel diagnostic—two controlled regimes of extended and hierarchical relevance—could serve as a reusable microbenchmark for frame-selection methods, isolating where different selection policies fail."],"forward_implications":["If CREST's claim holds, a training-free geometric rule can substitute for multi-stage retrieval pipelines, cutting preprocessing time by roughly 26–31× and peak memory by about 10.7×.","At a fixed frame budget, CREST extracts more usable evidence per frame than AKS, so the limiting factor in long-video QA shifts from frame count to selection policy.","The benefit is largest when relevant evidence is spread over long temporal spans (the paper's 'extended relevance' regime), suggesting the method suits videos with slowly developing narratives.","CREST's scoring mechanism is agnostic to the relevance model: gains persist across CLIP, BLIP, and Sevila scorers, so it can be layered on different backbones.","The query-conditioned design means the method applies directly to video question answering but not to query-agnostic summarization, as the paper itself notes."],"fun_headline_variants":["Curvature of relevance scores picks near-retrieval frames at 3% cost","Frame selection via curvature: 93-95% of retrieval quality at 3% cost","Relevance curvature trims frames: matches retrieval at 3% of the cost","Curvature-based frame selection nears retrieval accuracy at 3% preprocessing cost","Use relevance curvature to select video frames: near retrieval at 3% cost"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the discrete second difference of the query–frame relevance score is a reliable indicator of where informative events occur, and that the specific mapping R_i = R_base/(1+κ_i) is the right way to convert that local shape into a suppression radius.","fun_headline_variants_meta":{"raw":{"variants":["Curvature of relevance scores picks near-retrieval frames at 3% cost","Frame selection via curvature: 93-95% of retrieval quality at 3% cost","Relevance curvature trims frames: matches retrieval at 3% of the cost","Curvature-based frame selection nears retrieval accuracy at 3% preprocessing cost","Use relevance curvature to select video frames: near retrieval at 3% cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00109,"raw_usage":{"total_tokens":4423,"prompt_tokens":809,"completion_tokens":3614,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":3506}},"tokens_in":553,"tokens_out":3614,"duration_ms":26622,"temperature":1.0,"reasoning_tokens":3506,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T00:13:09.670416+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the CLIP relevance scores on a fixed video set with smoothed or high-frequency-noisy versions (keeping frames and queries identical) and compare CREST vs AKS accuracy: if CREST's advantage disappears or reverses when curvature no longer tracks genuine events, the mechanism's dependence on that premise is confirmed. A simpler check: on a video where the decisive information is a slow, low-curvature transition, CREST should miss it relative to top-K selection.","supporting_citations":[],"review_version":3}