{"id":"5828577b-3995-4987-8c74-57ef5dec6bbd","arxiv_id":"2507.14784","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LeAdQA improves video question answering by using LLM-rewritten causal queries to drive temporal grounding that selects relevant video segments for the answering model.","lead":"VideoQA systems often feed too many irrelevant frames into the answer model. LeAdQA uses an LLM to rewrite question-option pairs, then a temporal grounding model to pick only the moments that matter, improving accuracy on three benchmarks while using fewer frames.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 5 does not isolate causal rewriting: 'w/o Causal Rewriting' equals the no-grounding uniform-sampling baseline, so the claimed +1.3 points may come from grounding alone.","rationale":"The strongest evidence for the paper's central claim is Table 5. Reading that table closely, the 'w/o Causal Rewriting' condition is 79.3, exactly the 'Uniform Sampling' value, i.e., no grounding at all. The '+Causal Rewriting' row is 80.6, the full LeAdQA pipeline. Therefore the table cannot establish the marginal contribution of causal rewriting. The method (Eqs. 14-17) routes q'_i only into the grounding module; if visual input is uniform frames, rewriting cannot alter the MLLM output. This suggests either a mislabeled baseline or an undocumented path by which the rewritten queries affect the final prompt. The missing condition is grounding with original q_i=(Q,o_i). The reader's leakage concern is real and worth testing, but it is a contamination hypothesis; the ablation defect is visible in the paper itself and blocks attribution of the improvement to causal rewriting even in the absence of leakage. With this fixed, the grounding benefit remains credible (random 78.0 < uniform 79.3 < ground-truth segments 82.1), so the appropriate verdict stays conditional rather than reject. Acceptance should require the missing ablation, a non-causal paraphrase control, and release of the prompt templates and code so the check can be performed.","tokens_in":13360,"tokens_out":9838,"duration_ms":102663,"concrete_test":"Run a four-condition ablation on NExT-QA with Tarsier-34B, holding the visual input pipeline fixed: (1) uniform frame sampling, no rewriting (79.3); (2) UniVTG grounding with original q_i=(Q,o_i) as queries; (3) UniVTG grounding with non-causal paraphrases of q_i; (4) UniVTG grounding with causal rewrites (the paper's 80.6 condition). If (4) is not clearly above both (2) and (3) by more than about 0.5 points, the causal-rewriting effect is unsupported. Also inspect the deployed prompt template Pa to verify whether q'_i is ever included in the MLLM input; if it is, the method description must be corrected and that path ablated separately.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim assigns a load-bearing role to 'LLM-driven causal question-option rewriting' (Eq. 1). The only direct evidence is Table 5, 'Impact of QA Pair Rewriting with GPT-4.' In the 'Varying Rewriting (w/ Uniform Grounding)' block, 'w/o Causal Rewriting' is reported as 79.3, which is identical to the 'Uniform Sampling' row in the 'Varying Grounding' block (no grounding at all). The '+Causal Rewriting (ours)' row is 80.6, the full LeAdQA pipeline with grounding and rewriting. Per the method, the enriched queries q'_i are used only to produce grounded segments and are not included in the final MLLM prompt (Eqs. 14-17); if the visual input is truly uniform frames, toggling rewriting should change nothing. The table therefore appears to compare 'no grounding' against 'grounding + rewriting', not rewriting on versus off. The missing critical control is grounding with the original question-option tuples q_i=(Q,o_i) as UniVTG queries. Without that row, the reported +1.3 cannot be attributed to causal rewriting; it is the joint effect of adding temporal grounding and rewriting. There is also no non-causal paraphrase control, so the 'causal' component is confounded with converting options into natural-language sentences. This does not refute the grounding benefit shown in the same table (random 78.0 < uniform 79.3 < ground-truth segments 82.1), but it removes the paper's only isolation of the causal-rewriting mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LeAdQA, a VideoQA framework that combines LLM-driven rewriting of question-option pairs with a temporal grounding module (UniVTG) and an interval-fusion mechanism, then feeds the selected segments to an MLLM for answer generation. Experiments on NExT-GQA, NExT-QA, and IntentQA report consistent accuracy gains over base MLLMs, and an ablation shows that ground-truth segments outperform uniform sampling, which in turn outperforms random sampling. The central claim is that causal question-option rewriting improves temporal grounding and thereby improves VideoQA accuracy.","tokens_in":13720,"tokens_out":4123,"duration_ms":48298,"significance":"If the central claim is upheld, LeAdQA is a practically attractive, training-efficient wrapper that can be added to existing MLLMs such as Tarsier and Qwen2.5-VL, with reported gains of 0.5--1.6 points on three benchmarks. The paper explicitly credits the grounding benefit with a clean ordering across random, uniform, and ground-truth sampling, and it tests multiple MLLM backbones and frame budgets. However, the attribution of the gains to the causal-rewriting mechanism is not isolated by the current ablations, and the main NExT-QA comparison mixes validation and test splits. These issues are fixable but are load-bearing for the paper's headline claims.","major_comments":[{"comment":"The text states that LeAdQA is evaluated on the test sets of NExT-GQA and IntentQA and on the validation set of NExT-QA, yet Table 3 compares LeAdQA against baselines whose reported numbers in the cited works are test-set accuracies. Comparing a validation-set number with published test-set numbers is not apples-to-apples and can inflate the reported SOTA claim. Please either evaluate on the official NExT-QA test split or re-run all baselines on the same validation split, and clearly label the split in the table and caption.","section":"Results and Analysis, Table 3"},{"comment":"The 'Varying Rewriting (w/ Uniform Grounding)' block does not isolate causal rewriting. The row 'w/o Causal Rewriting' reports 79.3, identical to the 'Uniform Sampling' row in the 'Varying Grounding' block. Since the enriched queries q'_i are used only to produce grounded segments and are not included in the final MLLM prompt (Equations 14--17), toggling rewriting while keeping uniform sampling should change nothing; the comparison therefore conflates 'adding grounding' with 'adding causal rewriting'. The missing control is grounding with the original question-option tuples q_i=(Q,o_i) as UniVTG queries, with all other components identical. A second useful control is a non-causal paraphrase of the question-option pairs, to separate the effect of causal rewriting from the effect of converting options into natural-language sentences.","section":"Ablation Study, Table 5"},{"comment":"The LeAdQA-34B row in Table 1 reports only Acc@QA (77.0) and Acc@GQA (20.7); all grounding metrics (mIoP, IoP@0.3, IoP@0.5, mIoU, IoU@0.3, IoU@0.5) are missing for this strongest model. Since the paper's central claim is that temporal grounding drives the QA gains, the absence of grounding metrics for the best-performing configuration prevents the reader from verifying the grounding benefit at the largest scale. Please report the full grounding metric set for LeAdQA-34B.","section":"Table 1"},{"comment":"No error bars, multiple-seed runs, or significance tests are reported anywhere. Several headline gains are small (for example +0.5 on NExT-QA for Tarsier-7B and +0.8 on IntentQA for Tarsier-7B in Table 4), and the interval-fusion hyperparameters (Top-K, IoU threshold) appear to be tuned on the NExT-QA validation set. Without variance estimates or a significance test, it is difficult to judge whether these differences are robust. Please provide at least three seeds or a paired significance test for the main comparisons and ablations.","section":"In-depth Analysis and Tables 3--5"},{"comment":"There is a correctness-risk concern about leakage from GPT-4o rewriting. The prompt Pr asks GPT-4o to write a hypothetical video situation under which option o_i is true; for the correct option this description is likely to match the actual video, while for incorrect options it is not. Because NExT-QA, IntentQA, and NExT-GQA are public benchmarks, GPT-4o may have prior knowledge of the answers, which could make the grounded segments systematically favor the correct option and inflate the measured grounding and QA improvements. A concrete test would be to compare against a neutral rephrasing instruction that does not assume the truth of the option (e.g., 'rewrite the question-option pair while preserving its original semantics') and to measure grounding quality with the original q_i as the query. This would separate genuine reasoning improvement from answer leakage via the LLM's parametric memory.","section":"Method, Eq. (1), and Ablation Study, Table 5"}],"minor_comments":[{"comment":"The dataset name 'IntenQA' appears to be a typo for 'IntentQA', which is the name used elsewhere in the paper, including Table 4.","section":"Experimental Settings, Datasets"},{"comment":"The sentence 'VideoAgent (Wang et al. 2025) futher employs LLMs as iterative information extractors' contains the typo 'futher'; it should be 'further'.","section":"Related Works, Video Question Answering"},{"comment":"Table 2 labels the base models as 'Qwen-3B' and 'Qwen-7B', while the Implementation Details section says the authors use 'Qwen2.5-VL-3B-Instruct and Qwen2.5-VL-7B-Instruct'. Please unify the naming to avoid ambiguity about which Qwen version is evaluated.","section":"Implementation Details and Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful framework and a plausible grounding benefit, but the central causal-rewriting claim is not isolated, and the NExT-QA comparison mixes splits. These are fixable with additional experiments (the missing grounding-with-original-query control and a paraphrase control are essential). The leakage risk from GPT-4o rewriting on public benchmarks should also be addressed explicitly, as it could affect the interpretation of the reported gains. I recommend major revision rather than rejection because the remaining issues are addressable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on LeAdQA. The paper combines an LLM-based question-option rewriting step with a temporal grounding model (UniVTG) and an NMS-style interval fusion, then feeds grounded frames to an MLLM. It reports consistent gains of 1-4 points across NExT-QA, IntentQA, and NExT-GQA, and shows that using ground-truth segments beats uniform sampling by a wide margin. That part is solid: the grounding itself appears to help, and the system is a reasonable engineering recipe.\n\nThe soft spots are real, though. The central claim — that causal rewriting drives the improvement — is not actually isolated in Table 5. The 'w/o Causal Rewriting' row is identical to the uniform-sampling row (79.3), which makes sense because per the method the rewritten queries are only used to select segments, and if you don't ground (uniform), rewriting changes nothing. So the comparison is really 'no grounding' vs 'grounding + rewriting.' The missing control is grounding with the original question-option tuples as queries. Without that row, you cannot attribute the +1.3 points to causal rewriting; it could be grounding alone. There's also no non-causal paraphrase control, so the 'causal' component is confounded with simply converting options into sentences.\n\nSecond, the NExT-QA numbers are reported on the validation set while the baselines are test-set numbers. That's apples-to-oranges, and given that they tuned hyperparameters on that same validation set, the SOTA claim is inflated.\n\nThird, there's a plausible leakage risk: GPT-4o rewrites each option as a plausible scenario, and GPT-4o has likely seen these public benchmarks. If the rewrite of the correct option aligns better with the video, the grounding will systematically favor the correct answer. The paper doesn't test this, and it should.\n\nOn the mechanical side, Table 1 is missing grounding metrics for the best 34B model, and there are no error bars or significance tests anywhere. The gains are small (1-2 points) and could easily be noise.\n\nNone of this refutes the grounding benefit — the ablation shows random < uniform < ground-truth, and that's the core evidence. But it does mean the paper's headline 'causal rewriting' contribution is unproven.\n\nWho is this for? People working on long-video VideoQA will find the recipe useful and the negative-ish results about rewriting worth knowing. It deserves a serious referee, but it needs a revised ablation with the proper control, test-set numbers, and ideally a leakage check before the claims hold up. I'd send it to review, but with the expectation of substantial revision.","headline":"A plausible grounding-based VideoQA recipe with consistent gains, but the key ablation conflates grounding with rewriting and the NExT-QA comparison is val-vs-test; referee-worthy but needs major revision.","tokens_in":14255,"tokens_out":2620,"would_cite":false,"duration_ms":28074,"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":"LeAdQA rewrites each answer option into a causal scene description, then grounds the video to it before answering.","keywords":["Video Question Answering","temporal grounding","causal reasoning","LLM-driven question rewriting","multiple-choice video QA","long video understanding","multimodal large language models","NExT-QA"],"falsifier":"Hold out a fresh set of videos and questions written in the same style as NExT-QA that GPT-4o has not seen, run LeAdQA with and without the GPT-4o rewriting, and check whether the causal-rewriting gain persists; alternatively, replace GPT-4o with a paraphrase model trained only on synthetic data and compare. If the gain disappears on unseen questions, the reported improvement comes from answer leakage rather than from grounding.","tokens_in":13174,"feed_emoji":"🎬","tokens_out":7083,"duration_ms":74533,"temperature":0.7,"pith_summary":"The paper argues that a major bottleneck in VideoQA is not model capacity but deciding which frames to watch. Its central move is to rewrite each question-option pair into a causal event description—a hypothetical video situation in which that option is true—and use those rewritten queries to drive a temporal grounding model that localizes the relevant video segments. The grounded segments, fused across options, are then passed to a multimodal LLM for the final answer. On NExT-QA, IntentQA, and NExT-GQA, the pipeline improves accuracy over the same MLLMs used alone, and grounded sampling reaches uniform-sampling accuracy with fewer frames. If correct, this reframes VideoQA as an evidence-retrieval problem rather than a whole-video comprehension problem.","feed_headline":"Causal answer rewriting improves VideoQA across three benchmarks","feed_subtitle":"LeAdQA turns each answer option into a video-segment query, raising accuracy with fewer frames.","key_machinery":"The load-bearing component is the LLM-driven question-option rephrasing: a structured prompt asks GPT-4o to convert each pair $(Q, o_i)$ into a causal-enriched description of a video situation in which $o_i$ would be true. This turns answer selection into a set of retrieval queries, so the temporal grounding stage (foreground, boundary, and saliency heads over clip features, trained with contrastive and IoU losses) can localize evidence instead of relying on uniform sampling. A two-level interval fusion (intra-option then inter-option, with an IoU threshold of 0.3) then consolidates the evidence before the MLLM reads the sampled frames.","core_discovery":"LeAdQA treats each answer option as a candidate claim about the video and asks an LLM to spell out the situation that would make that claim true. The rephrased query $q'_i = M_r(P_r(Q, o_i))$ is then fed into a motion-aware temporal grounding model that predicts, per clip, a foreground flag, boundary offsets, and a saliency score. Candidate intervals are merged within and across options via an IoU-based fusion, and the surviving segments supply keyframes to an MLLM that generates the answer. The paper reports that this combination raises Tarsier-34B from 75.2 to 77.0 on NExT-GQA QA accuracy, from 79.3 to 80.6 on NExT-QA, and from 76.9 to 78.5 on IntentQA; the ablation attributes +1.3 points on NExT-QA to the causal rewriting itself, and shows QA accuracy improving as grounding quality improves (random 78.0, uniform 79.3, ground-truth segments 82.1).","pith_inferences":["If the leak-free assumption holds, the same rephrasing trick could be applied to other multiple-choice visual tasks, such as image QA, embodied QA, or video caption ranking, where options can be converted into hypothetical scene descriptions.","The paper's own numbers suggest an upper bound: if grounding were perfect, accuracy would rise considerably further (82.1 vs. 80.6 with causal rewriting), so investing in grounding quality may yield larger returns than scaling the MLLM.","One can test the causal-rewriting hypothesis directly by training a grounding model on original question-option pairs and on rewritten pairs with the same videos and comparing temporal IoU; the paper reports that higher tIoU correlates with QA accuracy but does not isolate grounding quality with and without rewriting.","Treating answer selection as retrieval opens a natural extension: feeding multiple fused intervals back into the MLLM with their timestamps so the model can reason about event order, not just about which content is present."],"forward_implications":["With LeAdQA, VideoQA accuracy becomes a function of temporal grounding quality: the paper shows accuracy climbing from 78.0 (random frames) to 79.3 (uniform sampling) to 82.1 (ground-truth segments) on NExT-QA, so improving grounding should transfer directly to QA accuracy.","Query-focused sampling within grounded intervals reaches 81.2% accuracy at 32 frames where uniform sampling needs 48, making the method a frame-efficiency lever for long-video MLLMs.","The gains are not tied to a single backbone: both Qwen2.5-VL and Tarsier models improve when LeAdQA is added, suggesting the rephrasing-plus-grounding recipe is model-agnostic.","Causal rewriting helps most on causal questions, including a +4.2 gain on Causal-How questions in IntentQA, indicating that rewriting supplies event-chain structure that the visuals alone do not provide."],"supporting_citations":[{"why":"Supplies GPT-4o, the LLM used for question-option rephrasing in Eq. 1.","marker":"(Achiam et al. 2023)"},{"why":"Provides UniVTG, the temporal grounding architecture with foreground, boundary, and saliency heads adopted by LeAdQA.","marker":"(Lin et al. 2023)"},{"why":"Defines Tarsier-7B and Tarsier-34B, the MLLM backbones whose standalone results serve as the main comparison baselines.","marker":"(Wang et al. 2024a)"},{"why":"Defines Qwen2.5-VL backbones used to show that LeAdQA improves different model architectures.","marker":"(Bai et al. 2025)"},{"why":"Provides the NExT-QA benchmark with its temporal, causal, and descriptive question categories.","marker":"(Xiao et al. 2021)"},{"why":"Provides the IntentQA benchmark with Causal-Why, Causal-How, and temporal reasoning splits.","marker":"(Li et al. 2023b)"},{"why":"Provides NExT-GQA and the IoP, tIoU, and grounded-QA metrics used to measure grounding quality.","marker":"(Xiao et al. 2024)"}],"fun_headline_variants":["LLM-rewritten options sharpen video QA grounding","LeAdQA: causal rewriting lifts VideoQA on 3 benchmarks","Causal option rewriting grounds VideoQA better","LeAdQA improves VideoQA via grounded answer queries","Rewriting answer options anchors video QA grounding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"That rewriting each option with GPT-4o on public benchmarks does not leak the correct answer; if GPT-4o already knows these benchmark questions, the measured grounding and accuracy gains may be inflated.","fun_headline_variants_meta":{"raw":{"variants":["LLM-rewritten options sharpen video QA grounding","LeAdQA: causal rewriting lifts VideoQA on 3 benchmarks","Causal option rewriting grounds VideoQA better","LeAdQA improves VideoQA via grounded answer queries","Rewriting answer options anchors video QA grounding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000799,"raw_usage":{"total_tokens":3541,"prompt_tokens":997,"completion_tokens":2544,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2469}},"tokens_in":613,"tokens_out":2544,"duration_ms":20268,"temperature":1.0,"reasoning_tokens":2469,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:47:24.525333+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out a fresh set of videos and questions written in the same style as NExT-QA that GPT-4o has not seen, run LeAdQA with and without the GPT-4o rewriting, and check whether the causal-rewriting gain persists; alternatively, replace GPT-4o with a paraphrase model trained only on synthetic data and compare. If the gain disappears on unseen questions, the reported improvement comes from answer leakage rather than from grounding.","supporting_citations":[],"review_version":1}