{"id":"80ce5aa2-d8c9-4a4d-992f-6a5165ef71ab","arxiv_id":"2607.10238","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DAR provides 15k videos with 37k event-aligned viewer-emotion segments and causal chains, and DAR-R1 (SFT+GRPO) leads 10+ MLLMs on segmentation, emotion accuracy, and reasoning quality.","lead":"This paper builds DAR, a large video dataset that labels how a viewer's emotions shift across consecutive events, plus causal explanations for each shift. It also trains DAR-R1, an MLLM baseline that segments those shifts, classifies 27 emotions, and writes grounded reasons.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The load-bearing risk is that dual-MLLM verification may certify model-consensus affect stories rather than human viewer transitions, so SOTA on DAR may not transfer to genuine human affect.","rationale":"The paper’s strongest claim is a dual one: (1) DAR is the first large-scale, event-aligned, viewer-centric dynamic affective reasoning benchmark with dense causal chains, and (2) DAR-R1 (SFT+GRPO with the five-component reward) sets SOTA on that benchmark versus 10+ MLLMs. Both halves require that the 36,908 emotion–reason annotations faithfully track human viewer affect transitions. The construction pipeline (§3.3) is almost entirely MLLM-driven; dual-consistency (§3.4) uses other MLLMs as judges; human validation (Appendix B / Table 4) is a limited error audit against those outputs, not a primary human gold standard. The reader correctly flags this as the weakest assumption. It is not a mathematical contradiction and does not invalidate the engineering contribution (tasks, scale, reward design, ablations, OOD TSL numbers), so REJECT is too strong. ACCEPT would over-claim human-ground-truth status. CONDITIONAL remains the right verdict: treat DAR as a useful, large, model-assisted benchmark whose human fidelity still needs the independent multi-annotator check above before it is treated as definitive affect ground truth. No stronger independent load-bearing flaw (e.g., broken reward math, non-reproducible protocol, or internal inconsistency) is evident in the manuscript.","tokens_in":20107,"tokens_out":844,"duration_ms":10729,"concrete_test":"On a stratified sample of ≥200 test videos (covering multi-segment transitions and the top transition pairs such as Craving→Satisfaction, Anxiety→Fear), collect independent 3-annotator human labels for segment boundaries (±0.5s), 27-way emotion, and short free-form viewer-centric reasons, without showing model outputs. Report (i) human–human vs human–DAR agreement (κ / mIoU / exact-match), (ii) re-score of DAR-R1 and the strongest baseline under human gold, and (iii) fraction of dual-MLLM-accepted samples that a majority of humans reject as non-viewer-centric or causally unsupported. If human–DAR emotion κ < 0.4 or Emo-Acc under human gold drops >10 points relative to Table 2, the human-ground-truth premise fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that DAR is a large-scale human-viewer-centric ground-truth benchmark (and that DAR-R1 SOTA on it advances dynamic affective reasoning) rests on §3.3–3.4: Gemini proposes boundaries, Qwen3-VL generates incremental captions and four emotion–reason pairs from VCE Top-3 candidates, then InternVL3.5 + Qwen3-omni dual-score visual grounding, causal logic, viewer-centricity, temporal consistency, and answer consistency, with rewrite-on-fail. Table 4 reports only ~2–3% expert error rates on a full test split plus a training subset, but that audit is post-hoc agreement with already-filtered model outputs, not independent multi-annotator human labeling of the same clips. Because generators and judges share similar pretraining priors, dual-consistency can lock in coherent but non-human affect narratives (e.g., over-weighting cinematic tropes). Every supervised target and every GRPO reward (R_emo, R_reason, GPT-4o judge in Table 2) then optimizes toward that consensus. Absolute Emo-Acc remains 28.6% even for DAR-R1, consistent with either hard task difficulty or label noise. If the labels are primarily model-consensus rather than human affect, the “first large-scale viewer-centric” and “new SOTA” claims hold only inside a synthetic loop.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper reframes video emotion analysis as Dynamic Affective Reasoning (DAR): temporally grounded segmentation of viewer affect shifts, fine-grained classification over 27 categories, and causal explanation of those shifts. It releases a large benchmark (15,087 videos, 36,908 event-aligned segments) built from VCE via a coarse-to-fine pipeline (semantic boundaries + PySceneDetect snapping, incremental differential captioning, stream-of-affect reasoning under Affective Events Theory), with dual-MLLM consistency checks and a post-hoc expert audit (Table 4). It also proposes DAR-R1 (SFT cold-start + GRPO with structural, count, temporal IoU, emotion, and reasoning rewards) and reports gains over 10+ MLLMs on segment-count accuracy, mIoU, emotion accuracy, and GPT-judged reasoning (Table 2), with supporting human ratings (Table 3), fine-tuned emotion baselines (Table 5), reward ablations (Table 7), and a small OOD localization check on TSL (Table 6).","tokens_in":20530,"tokens_out":1494,"duration_ms":18614,"significance":"If the annotations are accepted as a useful proxy for viewer-centric affect dynamics, the work is a clear contribution: it defines a hierarchical task suite that existing static clip-label and character-centric benchmarks do not cover, ships a large event-aligned resource with dense causal chains, and shows that task-specific GRPO rewards improve both boundary localization and reasoning quality over strong open MLLMs. The construction pipeline, multi-dimensional reward design, ablations, fine-tuned emotion baselines, and expert preference study are concrete engineering assets for affective computing and video MLLMs. The main significance risk is not novelty of the task idea but whether SOTA on DAR generalizes beyond model-consensus affect narratives to genuine human viewer transitions.","major_comments":[{"comment":"§3.3–3.4 and Table 4: The central claim that DAR is a large-scale viewer-centric ground-truth resource rests on MLLM generation (Gemini boundary proposals; Qwen3-VL captions and four emotion–reason pairs from VCE Top-3) filtered by other MLLMs (InternVL3.5, Qwen3-omni) on the same axes later used for evaluation (visual grounding, causal logic, viewer centricity, temporal consistency, answer consistency). Table 4 reports low expert error rates (~2–3%) on the full test split and a training subset, but this is post-hoc agreement with already-filtered model outputs, not independent multi-annotator human labeling of the same clips with inter-annotator agreement. Because generators and judges share similar pretraining priors, dual-consistency can certify coherent model-consensus stories rather than human affect transitions. This underwrites every supervised target and the R_emo / R_reason / GP","section":null},{"comment":"§5.1 Evaluation Metrics and Table 2 GPT-Score: Reasoning quality is scored by GPT-4o on the same five dimensions used in dual-consistency verification during construction. Combined with MLLM-generated rationales as training targets, this creates a closed evaluation loop that can reward stylistic and prior alignment rather than human-validated causal correctness. Table 3’s expert ratings on 100 videos are a useful check and show agreement with GPT preferences, but N=100 and five experts are thin relative to 1,441 test videos. Strengthen the paper by expanding human evaluation (or releasing full expert scores), reporting correlation/disagreement between GPT-4o and experts per dimension, and reporting automatic text metrics (e.g., the ROUGE-L/SBERT already used in Tables 5 and 7) alongside GPT scores in the main comparison table.","section":null},{"comment":"Table 2 Emo-Acc (28.6% for DAR-R1) and §5.2: Absolute emotion accuracy remains low even after SFT+GRPO, while SC-Acc and mIoU are only moderately improved. The manuscript attributes this to task difficulty, but under the annotation pipeline of §3.3 the same numbers are also consistent with residual label noise or over-constrained Top-3 candidate conditioning. Please analyze error modes (confusion among related emotions, boundary-induced label mismatch, sustained vs. transition segments) and, if possible, estimate an upper bound via human agreement on emotion labels for matched segments. Without that, it is hard to interpret whether DAR-R1 advances genuine affective recognition or mainly format/localization compliance on synthetic targets.","section":null}],"minor_comments":[{"comment":"Table 1: “Dynamic Emotion” is marked only for DAR; briefly define the criterion in the caption so the comparison is falsifiable against MELD timestamps or continuous-affect datasets discussed in Appendix A.1.","section":null},{"comment":"§4.2 Eqs. (4)–(9): Reward weights (w1…w5), λ’s, and temperatures are stated in text/implementation details but not jointly tabulated; a single hyperparameter table would aid reproducibility.","section":null},{"comment":"Figure 3(c) transition matrix: report how transitions are normalized (row-stochastic?) and whether rare emotions are pooled; the two highlighted transitions alone do not establish “structured non-random” dynamics without a null baseline.","section":null},{"comment":"§5.1: Vision encoder is frozen for Qwen2.5-VL-3B; note whether this choice limits fine temporal boundary learning relative to unfreezing the aligner only, or provide a short ablation if available.","section":null},{"comment":"Appendix E Table 5: Fine-tuned emotion baselines still lag badly on Emo-Acc; clarify training budget parity (epochs, data mixture, output schema adaptation) so the comparison is fair.","section":null},{"comment":"Typos/consistency: “viewer-Centric” capitalization in §2.2; occasional missing spaces in compound words in the abstract/intro (“changeasaresultof”, “large-scalebenchmark”); arXiv id in header is fine but ensure camera-ready fixes.","section":null}],"recommendation":"major_revision","confidential_remarks":"The task formulation and engineering are solid for a top CV venue if the authors either strengthen human grounding or tone the “first large-scale viewer-centric ground truth” rhetoric. I would not reject solely for LLM-assisted annotation—common practice—but the current manuscript overclaims human affect fidelity relative to the evidence in Table 4. Fit is good for a dataset+benchmark track; main risk is synthetic-loop inflation of SOTA claims."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway: this is a real dataset-and-task paper, not a thin rebrand. They take VCE videos, cut them into event-aligned segments, attach 27-way viewer emotions plus causal rationales, and define three hierarchical tasks (segmentation, classification, reasoning). That package is cleaner than the static clip-label setup most of the field still uses.\n\nWhat is actually new is the joint product: dense temporal boundaries, fine-grained viewer labels, and explicit causal chains on ~15k videos / ~37k segments, plus a reproducible SFT+GRPO baseline (DAR-R1) with task-specific rewards for structure, count, IoU, emotion match, and rationale quality. Table 2 is honest about how badly off-the-shelf emotion models and general MLLMs do on the joint problem; SFT helps a lot, GRPO helps more on boundaries and causal scores. They also fine-tune emotion-focused baselines (Table 5), run reward ablations (Table 7), a small OOD localization check on TSL (Table 6), and a 100-video expert rating that tracks the GPT judge. Construction is documented carefully (coarse-to-fine cuts, differential captions, dual-MLLM rewrite loop). Citations to AET, VCE, StimuVAR, AffectGPT, etc. look fair; they do not pretend prior work never touched viewer affect.\n\nThe soft spot is real but proportionate: gold labels and reasons are mostly model-generated (Gemini/Qwen/InternVL) and dual-checked by other MLLMs on the same axes later used for scoring. Table 4’s ~2–3% expert error rates are post-hoc agreement with already-filtered outputs, not independent multi-annotator human labeling from scratch. So “human viewer ground truth” is only partially established; some of the SOTA is SOTA inside a synthetic consensus loop. Absolute Emo-Acc is still only 28.6%, which is either hard task or label noise. Free parameters in the reward design are many, as usual for GRPO recipes. None of that collapses the contribution; it just means treat DAR as a strong constructed benchmark, not definitive psychology.\n\nWho it is for: people building video MLLM evals or affective agents who need temporal + causal supervision. Worth a serious referee. I would engage, cite the dataset/tasks if I work in this area, and bring it to reading group with the label-origin caveat on the table.","headline":"Solid new benchmark package for event-aligned viewer affect; main soft spot is MLLM-mediated labels, not a broken method.","tokens_in":21159,"tokens_out":592,"would_cite":true,"duration_ms":6455,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Video emotion should be modeled as event-driven affect transitions, not one static clip label.","keywords":["dynamic affective reasoning","viewer-centric video emotion","affective segmentation","causal emotion reasoning","multimodal large language models","reinforcement learning","Affective Events Theory","video understanding"],"falsifier":"Have independent human viewers mark emotion-shift times, choose from the 27 categories, and write short causal reasons on the same test videos; if human boundaries, labels, and rationales systematically disagree with DAR annotations or reverse DAR-R1’s ranking against strong baselines, the central claim fails.","tokens_in":20968,"feed_emoji":"🎬","tokens_out":856,"duration_ms":8855,"temperature":0.7,"pith_summary":"This paper argues that standard video emotion analysis is too static: it labels whole clips and largely ignores how a viewer’s feeling changes when consecutive events accumulate. Drawing on Affective Events Theory, the authors reframe the problem as Dynamic Affective Reasoning—finding when emotion shifts, naming the fine-grained viewer emotion, and explaining why the visual event caused it. They release DAR, a large viewer-centric dataset of 15,087 videos and 36,908 event-aligned segments with 27 emotion categories and dense causal reasoning chains. They also define three linked tasks (affective segmentation, fine-grained classification, and causal reasoning) and train DAR-R1, a two-stage model that first learns the required output structure and then uses reinforcement learning with rewards for boundaries, emotion accuracy, and reasoning quality. Across more than ten multimodal models, DAR-R1 leads on temporal localization and affective reasoning, supporting the claim that emotion understanding needs event-level, viewer-centric dynamics.","feed_headline":"Video emotion shifts with events, not one static label","feed_subtitle":"A 15k-video benchmark and RL-tuned model track when feelings change and why","key_machinery":"DAR-R1: cold-start supervised fine-tuning for structured multi-segment outputs, followed by Group Relative Policy Optimization with rewards for schema validity, segment count, temporal IoU/boundary precision, emotion match under IoU>0.5, and non-repetitive reasoning quality.","core_discovery":"The paper claims that viewer emotion in video is better treated as a sequence of event-aligned affect phases with explicit causal explanations than as a single clip-level label, and that a large dataset built this way (DAR) plus a two-stage SFT-then-GRPO model (DAR-R1) yields new state-of-the-art results on emotional localization and causal reasoning.","pith_inferences":["If model-written labels dominate the ground truth, future gains may overfit annotation style unless human re-annotation or live viewer studies become the primary yardstick.","The same event-aligned formulation could transfer to domains where stimuli accumulate over time, such as interactive counseling videos or continuous user-response monitoring.","Reward design that separately pressures boundaries, counts, and rationales may be reusable for other multi-segment video reasoning tasks beyond emotion."],"forward_implications":["Emotion-aware video systems should output multi-phase timelines with reasons, not one global tag.","Training and evaluation should reward temporal boundary precision and causal grounding, not only category accuracy.","Viewer-centric 27-way taxonomies become a practical target for fine-grained affect modeling beyond basic facial-expression labels.","Affective agents can be scored on when emotion changes, what it becomes, and why the preceding visual event caused the shift."],"fun_headline_variants":["Video emotions track event sequences, not one static label","Viewer affect shifts with causal video events over time","DAR maps emotion phases to consecutive video events","Emotions evolve as event-aligned phases, not clip labels","Benchmark treats viewer feelings as dynamic event chains"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The load-bearing premise is that the largely model-generated, model-verified emotion labels and causal rationales truly track human viewers’ affect transitions rather than a machine consensus story about emotion.","fun_headline_variants_meta":{"raw":{"variants":["Video emotions track event sequences, not one static label","Viewer affect shifts with causal video events over time","DAR maps emotion phases to consecutive video events","Emotions evolve as event-aligned phases, not clip labels","Benchmark treats viewer feelings as dynamic event chains"]},"model":"grok-4.5","effort":"low","cost_usd":0.00557,"raw_usage":{"total_tokens":1477,"prompt_tokens":775,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":55700000,"prompt_tokens_details":{"text_tokens":775,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":647,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":775,"tokens_out":55,"duration_ms":5250,"temperature":1.0,"reasoning_tokens":647,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T13:18:14.398195+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Have independent human viewers mark emotion-shift times, choose from the 27 categories, and write short causal reasons on the same test videos; if human boundaries, labels, and rationales systematically disagree with DAR annotations or reverse DAR-R1’s ranking against strong baselines, the central claim fails.","supporting_citations":[],"review_version":1}