{"id":"6e0b3c0d-e3e0-4f46-ba7c-8db622fa00e6","arxiv_id":"2506.18071","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"MUPA combines three ordering-based reasoning paths with a reflection agent that verifies and fuses answer-evidence pairs, reaching 30.3% and 47.4% grounded QA accuracy on NExT-GQA and DeVE-QA with a 7B model.","lead":"This paper proposes MUPA, a system that runs three video-question-answering paths in different orders (locate then answer, answer then locate, and both at once) and uses a reflection agent to select consistent answer-evidence pairs. The authors report state-of-the-art grounded question-answering scores on the NExT-GQA and DeVE-QA benchmarks, using 2B and 7B parameter models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DeVE-QA 'zero-shot' SOTA is not zero-shot: Table 5 lists 78K DeVE-QA samples in GQA training, so the 47.4% Acc@GQA and the '2B beats all 7B' claim are unsupported as stated.","rationale":"The central claim has two halves: new SOTA on NExT-GQA and new SOTA on DeVE-QA, plus the abstract's efficiency claim that 2B beats all 7B competitors. The NExT-GQA half is well supported: NExT-GQA does not appear in the Appendix C.1 training list, so the protocol is plausibly zero-shot. The load-bearing weakness is the DeVE-QA half. Table 5 explicitly includes DeVE-QA (78K of 179K samples) in the GQA agent's supervised fine-tuning data, yet Table 2 and the text call the evaluation 'zero-shot.' Because the GQA agent's heads are reused across all three paths, training on DeVE-QA means the 47.4% vs 44.2% SOTA gap over VideoMind-7B is an in-domain comparison, not evidence of zero-shot multi-path reasoning. The same table also disproves the abstract's blanket 2B-vs-7B claim: MUPA-2B is 43.9 on DeVE-QA, below VideoMind-7B's 44.2. The reader's weakest_assumption was the Verifier's circular training; that is a legitimate concern about the interpretability of the reflection mechanism, but I judge the training/evaluation overlap on DeVE-QA to be more directly load-bearing for the stated numeric SOTA. Because the reader's verdict was already CONDITIONAL, my independent concern does not move the verdict; it sharpens the reason for conditionality and the required revision: either re-run DeVE-QA in true zero-shot fashion, or re-scope the claims as in-domain fine-tuning and correct the abstract.","tokens_in":16642,"tokens_out":7418,"duration_ms":62033,"concrete_test":"Retrain MUPA's GQA agent on the identical 179K corpus after removing all 78K DeVE-QA samples, then re-run the DeVE-QA evaluation of Table 2 with all other settings fixed. If Acc@GQA falls to near VideoMind-7B's 44.2 or below, the claimed 47.4 zero-shot SOTA is an artifact of training on the evaluation benchmark.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The manuscript's headline state-of-the-art on DeVE-QA is not a zero-shot result. Table 2 is labeled 'Zero-shot' and Section 4.1 says 'zero-shot results on the challenging DEVE-QA benchmark'; the abstract's claim that 2B parameters outperform all 7B-scale competitors also depends on this comparison. However, Table 5 in Appendix C.1 lists the GQA agent's supervised fine-tuning data as 179K samples that explicitly include DeVE-QA (78K). The GQA agent is shared by all three reasoning paths: Path-3 uses it jointly, and Path-1 and Path-2 reuse its Grounder and Answerer heads. Every path therefore sees DeVE-QA question-answer pairs and ground-truth moments during training. The 47.4% Acc@GQA is an in-domain fine-tuned number, not zero-shot generalization. Furthermore, MUPA-2B scores 43.9 on DeVE-QA, below VideoMind-7B's 44.2, so the abstract's blanket 'outperforms all 7B-scale competitors' is false on one of the two benchmarks. The NExT-GQA result is not affected by this issue, but the DeVE-QA half of the central SOTA claim and the abstract's efficiency claim are unsupported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MUPA, a multi-path agentic framework for Grounded VideoQA. It combines a Grounder, an Answerer, a joint GQA agent, and a Reflective Agent along three reasoning paths (ground-first, answer-first, joint), then applies a verifier with product-of-experts rescoring for single-path consistency and mixture-of-experts clustering for multi-path fusion. The authors report state-of-the-art Acc@GQA results on NExT-GQA (30.3% for MUPA-7B) and DeVE-QA (47.4% for MUPA-7B), and claim that the 2B model outperforms all 7B-scale competitors. The paper includes ablations of the three paths and the reflection mechanism, qualitative examples, and additional moment-retrieval experiments on TACoS.","tokens_in":16930,"tokens_out":6458,"duration_ms":59287,"significance":"If the NExT-GQA result holds, MUPA provides a useful demonstration that orchestrating multiple lightweight LoRA-tuned agents with a post-hoc verifier and fusion module can improve grounded VideoQA over a single-path VideoMind baseline. The multi-path decomposition and the PoE/MoE pipeline are simple, modular, and the code is released. The ablation on NExT-GQA shows consistent, if modest, gains from multi-path reasoning and from the reflection mechanism. However, the DeVE-QA half of the central SOTA claim, the zero-shot description of those results, and the blanket '2B outperforms all 7B' claim are contradicted by the paper's own tables and training-data appendix; the verifier is trained on the Grounder's own pseudo-labels, so the claimed impartial auditing is not independently established.","major_comments":[{"comment":"The DeVE-QA results are not zero-shot. Table 5 lists DeVE-QA (78K samples) in the supervised fine-tuning data of the GQA agent, and Section 3.2.1 states that the GQA agent is shared by all three reasoning paths. Every path therefore sees DeVE-QA question-answer pairs and ground-truth moments during training. Consequently, the 47.4% Acc@GQA on DeVE-QA is an in-domain fine-tuned number, not a zero-shot result, and the 'Zero-shot' label in Table 2 and the phrase 'zero-shot results on the challenging DEVE-QA benchmark' in Section 4.1 are inaccurate. This also contradicts the Section 4.1 claim that MUPA relies on 'no task-specific fine-tuning.' The NExT-GQA result is not affected, but the DeVE-QA part of the SOTA claim must be reframed as a fine-tuned comparison.","section":"§4.1, Table 2, Appendix C.1"},{"comment":"The claim that 'Despite using only 2B parameters, our method outperforms all 7B-scale competitors' is false on DeVE-QA: MUPA-2B achieves 43.9 Acc@GQA while VideoMind-7B achieves 44.2. The 2B claim does hold on NExT-GQA (28.7 vs. 28.2 for VideoMind-7B), but the abstract's blanket statement and Section 4.1's 'outperforming all existing 7B-scale competitors' must be restricted to NExT-GQA or otherwise qualified.","section":"Abstract and Table 2"},{"comment":"The Verifier is trained on pseudo-labels generated by the pretrained Grounder, and at inference the PoE fusion multiplies the Grounder's confidence by the Verifier score. The paper presents the Reflective Agent as 'an impartial auditor,' but the Verifier's training signal is the Grounder's own output, so the verification may simply imitate the Grounder's biases rather than independently audit its evidence. The claimed suppression of hallucinated evidence is therefore not independently established. The authors should either train the Verifier with human-annotated or otherwise independent spans, or provide an analysis (e.g., per-span error breakdown or calibration against ground-truth moments) showing that the Verifier does not merely confirm the Grounder's mistakes.","section":"§3.3.1 and Appendix C.1"}],"minor_comments":[{"comment":"The text states that MUPA-2B achieves 29.0% Acc@GQA and a mIoP of 39.7%, but Table 1 reports 28.7 and 39.1; these numbers should be reconciled.","section":"§4.1 and Table 1"},{"comment":"The column header 'w w/o R1@0.5 mIoU R1@0.5 mIoP' is ambiguous as typeset, and the table does not clearly show a multi-path-with-reflection versus multi-path-without-reflection comparison. Section 4.3's claim that the Reflection Agent improves the full MUPA system would be better supported by an explicit row pair for the multi-path condition.","section":"Table 3"},{"comment":"Equation (3) defines answer consolidation using sum over i and k of p_ik, while Algorithm 2 uses sum over i of p_i; the paper should clarify which path-level or span-level confidence is intended in the final voting.","section":"§3.3.2 and Algorithm 2"},{"comment":"The reference to 'HiREST [? ]' is unresolved, and the appendix title 'Discussion on Entropy-based Measurement' does not match the actual content of the appendix.","section":"Appendix C.1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth a serious look, but the headline claims are looser than the numbers. The actual contribution is a clean empirical recipe: run three reasoning paths (ground-first, answer-first, joint GQA) over a shared Chain-of-LoRA backbone, then reflect and fuse with a Verifier plus PoE/MoE. Each component is borrowed from VideoMind, but the composition is new and the ablations show multi-path and reflection each help on NExT-GQA. That is a modest, real advance.\n\nWhere it gets soft is the DeVE-QA half of the story. Table 2 is labeled 'zero-shot', and the abstract's 'outperforms all 7B-scale competitors' leans on that. But Table 5 lists 78K DeVE-QA samples inside the GQA agent's supervised fine-tuning set. The GQA agent is shared across all three paths, so every path has seen DeVE-QA ground-truth moments at training time. The 47.4% Acc@GQA is an in-domain fine-tuned number, not zero-shot. And on that same table, MUPA-2B scores 43.9 against VideoMind-7B's 44.2, so the blanket 2B-vs-7B claim is false as stated. This is not a nitpick; it is the paper's own data contradicting its abstract.\n\nA second concern, softer but worth flagging: the Verifier is trained on pseudo-spans produced by the Grounder (Appendix C.1), then used at inference to score the Grounder's spans. That makes the 'impartial auditor' less independent than claimed. The PoE product of Grounder confidence and Verifier score can just reinforce the Grounder's biases. It does not sink the NExT-GQA numbers, but it weakens the interpretation that reflection independently suppresses hallucinations.\n\nThere are also no error bars, and several gains are around one point. That is a standard limitation for this kind of system paper, not a fatal flaw.\n\nCredit where due: the multi-path ablation is clean, the 7B SOTA on NExT-GQA (30.3% Acc@GQA, beating VideoMind-7B by a real margin) is supported, and the limitations section honestly admits higher inference cost. The citation pattern is fine, with self-citations to VideoMind being appropriate.\n\nWho is this for? Researchers working on grounded VideoQA and multi-agent reasoning pipelines. It deserves a serious referee, but the authors need to fix the zero-shot labeling, correct the abstract's efficiency claim, and clarify the Verifier's training dependence. I'd send it to peer review with a request for major revision, not desk reject.","headline":"A useful multi-path recipe with a solid NExT-GQA result, but the DeVE-QA 'zero-shot' SOTA and the 2B-beats-7B claim are contradicted by the paper's own tables.","tokens_in":17537,"tokens_out":1512,"would_cite":false,"duration_ms":15344,"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":"Running ground-first, answer-first, and joint paths, then fusing the verified evidence, sets a new state of the art in grounded video QA.","keywords":["grounded video question answering","multi-agent reasoning","temporal grounding","reflection agent","product of experts","mixture of experts","video moment retrieval","video-language models"],"falsifier":"Train the Verifier on human-annotated ground-truth spans from NExT-GQA or DeVE-QA instead of Grounder-produced pseudo-spans while holding everything else fixed; if MUPA's Acc@GQA and IoP scores do not change, the reflection stage is not the source of the gains, whereas a drop would show the current Verifier mostly imitates the Grounder's biases rather than auditing them.","tokens_in":16392,"feed_emoji":"🎬","tokens_out":8595,"duration_ms":81164,"temperature":0.7,"pith_summary":"MUPA sets out to prove that a video question-answering system can ground its answers in visible evidence by refusing to rely on a single reasoning path. It runs three paths in parallel—localize-then-answer, answer-then-localize, and jointly localize-and-answer—then subjects each answer–evidence pair to a verifier and fuses the surviving candidates into one answer with temporal boundaries. The paper reports that this design raises temporal grounding fidelity (IoP) without lowering answer accuracy, with MUPA-7B reaching 30.3% grounded accuracy on NExT-GQA and 47.4% on DeVE-QA, and MUPA-2B surpassing 7B-scale competitors on NExT-GQA. If true, the contribution is a parameter-shared recipe for trustworthy video question answering that adds no trainable parameters at the fusion stage.","feed_headline":"Three reasoning paths, one verifier, new grounded-video-QA records","feed_subtitle":"MUPA-7B reaches 30.3% and 47.4% grounded accuracy on NExT-GQA and DeVE-QA by verifying and fusing evidence from three paths.","key_machinery":"Three reasoning paths supply diversity: Path-1 grounds the question to moments, then answers; Path-2 answers first, then retrieves moments by combining answer and question; Path-3 decodes answer and evidence jointly. The reflection agent carries the argument: a Verifier trained with LoRA judges whether each zoomed-in, boundary-marked span contains the queried event, and the product-of-experts rule $p_{ik}=c_{ik}v_{ik}$ re-ranks spans so a span survives only if the Grounder and Verifier both believe it. Multi-path fusion then converts spans to (start, end) points, clusters them with confidence-weighted $k$-means, and refines cluster centers by weighted least squares, so the final top-$K$ evidence spans are a closed-form compromise across paths.","core_discovery":"On the paper's own terms, the discovery is that answer–evidence consistency, not extra capacity, is the bottleneck for grounded video question answering. The system uses four agents: a Grounder that localizes candidate moments, an Answerer that proposes answers, a GQA agent that does both at once, and a Reflective Agent that scores each answer–evidence pair, suppresses inconsistent pairs with a product-of-experts re-scoring rule, and fuses the rest by confidence-weighted clustering. The paper argues this makes answers driven primarily by video content rather than textual shortcuts. The headline empirical claim is state-of-the-art zero-shot grounded QA accuracy—30.3% on NExT-GQA and 47.4% on DeVE-QA at 7B scale—with grounding metrics improving across the board and answer accuracy staying on par with single-path models.","pith_inferences":["The Verifier is trained on spans produced by the same Grounder (Appendix C.1), so the reflection agent may be auditing the Grounder's own biases rather than an external notion of correctness; the paper does not isolate how much of the reported gain is independent verification.","The answer-first path exploits the assumption that answer prediction is more reliable than temporal localization; if a deployment's localization model is stronger than its QA head, the fixed three-path weighting would likely be suboptimal.","Inference cost scales with the number of paths, and the paper notes there is no planner to prune redundant trajectories; a learned gating module that skips low-value paths is a natural follow-up and can be tested on the same benchmarks.","A scaling curve across backbone sizes would clarify whether multi-path diversity remains beneficial as single-path models grow stronger."],"forward_implications":["If the central claim holds, a 2B-parameter grounded-QA model can beat larger systems by spending compute on multiple reasoning paths and verification rather than on raw scale.","The same frozen backbone with swappable LoRA adapters covers moment retrieval, video QA, and grounded video QA, so one deployment can serve all three tasks.","Because the fusion step (product-of-experts re-scoring plus weighted $k$-means) is parameter-free, the reflection stage's only extra trainable component is the Verifier itself.","Each final answer comes with a confidence-weighted consensus time interval, making errors more inspectable by a human or downstream model."],"supporting_citations":[{"why":"Supplies the NExT-GQA benchmark and the Acc@GQA grounded-accuracy metric that the paper's central comparison uses.","marker":"[4]"},{"why":"Supplies the DeVE-QA benchmark for dense-event grounded QA where MUPA reports 47.4% Acc@GQA.","marker":"[8]"},{"why":"Provides the VideoMind Chain-of-LoRA agent design, the Grounder and Verifier architecture, and the main per-size baseline.","marker":"[20]"},{"why":"Provides the Qwen2-VL backbone shared by all MUPA agents.","marker":"[2]"},{"why":"Supplies the SeViLA two-stage grounder-then-answer baseline and its related pipeline for comparison.","marker":"[62]"},{"why":"Supplies the FrozenBiLM zero-shot video QA baseline and its grounding evaluation.","marker":"[6]"},{"why":"Supplies the LLoVi long-video QA baseline that MUPA-2B is compared against.","marker":"[36]"},{"why":"Supplies the HawkEye 7B grounding baseline.","marker":"[37]"},{"why":"Supplies the VideoChat-TPO 7B baseline.","marker":"[38]"}],"fun_headline_variants":["Multi-path agentic reasoning improves grounded video QA","Answer–evidence consistency beats model scale in video QA","Reflective agents fuse three reasoning paths for video QA","MUPA-7B sets grounded-video-QA records with four agents","Verifying evidence across paths yields SOTA video QA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Verifier acts as an independent auditor of the Grounder's evidence, yet it is trained on pseudo-ground-truth spans produced by that same Grounder; if the Verifier has internalized the Grounder's mistakes, the claimed suppression of hallucinated evidence is not independently established.","fun_headline_variants_meta":{"raw":{"variants":["Multi-path agentic reasoning improves grounded video QA","Answer–evidence consistency beats model scale in video QA","Reflective agents fuse three reasoning paths for video QA","MUPA-7B sets grounded-video-QA records with four agents","Verifying evidence across paths yields SOTA video QA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1572,"prompt_tokens":940,"completion_tokens":632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":552}},"tokens_in":556,"tokens_out":632,"duration_ms":6617,"temperature":1.0,"reasoning_tokens":552,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:56:01.835960+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the Verifier on human-annotated ground-truth spans from NExT-GQA or DeVE-QA instead of Grounder-produced pseudo-spans while holding everything else fixed; if MUPA's Acc@GQA and IoP scores do not change, the reflection stage is not the source of the gains, whereas a drop would show the current Verifier mostly imitates the Grounder's biases rather than auditing them.","supporting_citations":[{"cited_title":"Can I trust your answer? Visually grounded video question answering,","cited_arxiv_id":null,"evidence_quote":"Supplies the NExT-GQA benchmark and the Acc@GQA grounded-accuracy metric that the paper's central comparison uses."},{"cited_title":"Self-chained image-language model for video localization and question answering,","cited_arxiv_id":null,"evidence_quote":"Supplies the SeViLA two-stage grounder-then-answer baseline and its related pipeline for comparison."},{"cited_title":"Zero-shot video question answering via frozen bidirectional language models,","cited_arxiv_id":null,"evidence_quote":"Supplies the FrozenBiLM zero-shot video QA baseline and its grounding evaluation."}],"review_version":2}