{"id":"90a1dcf3-35ca-4b8b-9473-8b6a20124410","arxiv_id":"2505.11405","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Multimodal LLMs frequently accept hallucinated emotion claims on a new adversarial benchmark, with the worst failures on image, audio, and video perception rather than on textbook emotion knowledge.","lead":"This paper introduces EmotionHallucer, a benchmark of paired yes/no questions that tests whether multimodal AI models can tell true statements about emotions from deliberately distorted ones. The authors report that most current models fail, especially on real-world image, audio, and video emotion perception, and that a simple explain-then-predict prompt pipeline improves scores on their benchmark.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark ground-truth validity is the load-bearing assumption: hallucinated QA variants are taken as unambiguously false with no reported inter-annotator agreement or human baseline.","rationale":"The reader's conditional verdict is well supported, and the reader's weakest assumption is also the one I would stress-test. I considered whether the PEP-MEK average improvement mismatch or the missing data release should take priority; both are auditability problems, but the benchmark's central finding would survive neither if the ground-truth labels are contestable. The ground-truth assumption is load-bearing because the headline is empirical: models score around 50% on EmotionHallucer-NoAudio, and the paper concludes that \"most current models exhibit substantial issues.\" That claim is only interpretable if incorrect answers are known to be incorrect. The paper's Appendix A describes two-stage cross-review but gives no disagreement counts, no kappa, and no comparison to human performance; Appendix E concedes annotation noise. The perception categories are especially vulnerable because intensity and reasoning-cue judgments are graded and context-dependent. The proposed test directly measures label reliability and the human ceiling. If the test passes, the benchmark's main claims are much stronger; if it fails, the paper should be revised either to soften the strong wording or to recalibrate the reported scores. The verdict therefore remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":36471,"tokens_out":3342,"duration_ms":35732,"concrete_test":"Recruit three independent annotators fluent in English and familiar with affective science to answer a stratified random sample of at least 200 basic and 200 hallucinated items, with access to the source materials used during construction. Compute per-item accuracy against the published labels and Fleiss' kappa across annotators, separately for the emotion-knowledge and multimodal-perception subsets. If human accuracy on hallucinated items falls below roughly 95%, or kappa below about 0.7, the labels are too contestable to support the claim that models are hallucinating rather than facing ambiguous items. Until the benchmark data are released, a minimal alternative is to require the authors to provide complete annotation disagreement statistics and the exact source passages or clips underlying every item.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that \"most current models exhibit substantial issues with emotion hallucinations\" requires EmotionHallucer's labels to be a valid measure of hallucination. In particular, every basic item must be true and every hallucinated item must be unambiguously false. The paper's construction (Appendix A) uses trained annotators to generate hallucinated variants and a second annotator to cross-review, but it reports no inter-annotator agreement and no human accuracy baseline. Appendix E explicitly concedes that \"there is noise introduced by human annotations.\" The risk is concentrated in the perception categories: intensity items (e.g., \"normal\" versus \"strong\" sadness) and reasoning-cue items (e.g., whether a narrowed eye or quickened breath is a reliable cue) are plausibly contestable even for human raters. Without a human ceiling, low model accuracy could reflect benchmark ambiguity or label noise rather than emotion hallucination. Under the paired scoring rule, a contestable hallucinated item depresses the overall score for any model that answers the basic item correctly, so the headline finding is directly sensitive to this assumption. This is a correctness risk in the measurement, not an objection based on disagreement with current consensus.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces EmotionHallucer, a benchmark for detecting emotion hallucinations in large language models and multimodal large language models. The benchmark has two dimensions: emotion psychology knowledge (theory, definition, and finding subcategories) and real-world multimodal perception (category, intensity, reasoning result, and reasoning cue subcategories), built from a psychology textbook and from the SOUL, Twitter15/17, RAVDESS, MER 2023, and Social-IQ 2.0 datasets. Evaluation uses adversarial paired yes/no questions, where a model must answer both a basic (true) and a hallucinated (false) question correctly for the pair to count as correct. The authors evaluate 38 models, report overall accuracies far below 100% and often at or below the 25% chance level, and propose a Predict-Explain-Predict with Modality and Emotion Knowledge (PEP-MEK) framework that they report improves hallucination detection by an average of 9.90%.","tokens_in":36639,"tokens_out":7868,"duration_ms":71261,"significance":"If the benchmark's ground-truth labels are valid, EmotionHallucer would fill a genuine gap in hallucination evaluation, as existing benchmarks focus on objects, attributes, and general facts rather than affective content. The adversarial paired design and the coverage of four modalities are useful contributions, and the evaluation spans a broad set of current open- and closed-source models. The paper also proposes a concrete mitigation framework, PEP-MEK, and provides qualitative examples. However, the load-bearing assumption is that every basic statement is true and every hallucinated statement is unambiguously false. The paper does not report inter-annotator agreement or a human baseline, and it explicitly concedes annotation noise in Appendix E. The internal statistics also do not reconcile across sections, so the benchmark cannot currently be audited from the manuscript. These issues raise a correctness risk in the measurement that the revisions must address.","major_comments":[{"comment":"The benchmark statistics are internally inconsistent. Section 3.1.1 reports 81 + 133 + 178 = 392 pairs for the emotion knowledge dimension, while Appendix A.1.1 reports 784 basic and hallucination QA pairs for the textbook source and 200 for SOUL, 300 for image, 736 for audio, 360 for MER 2023, and 402 for Social-IQ 2.0. Summing the Appendix numbers gives 2,782 questions, not the 2,742 reported in Table 4. If the Section 3.1.1 numbers are interpreted as pairs and the Appendix numbers are interpreted as questions, the sum is 2,390 questions; if all are pairs, the sum is 4,780 questions. The phrase \"QA pair\" is used ambiguously, and no per-category breakdown reconciles with the stated total. The authors must clarify whether each reported number denotes one question or one basic-hallucinated pair, and provide a per-category count that sums exactly to the total in Table 4.","section":"Section 3.1.1, Appendix A, Table 4 (Section 3.2)"},{"comment":"The ground-truth validity of the hallucinated items is not established. The annotation pipeline relies on two trained annotators with cross-review, but the paper reports no inter-annotator agreement statistics, and Appendix E concedes that \"there is noise introduced by human annotations.\" Under the paired scoring rule described in Section 3.3, a contestable hallucinated item directly depresses the overall score of any model that correctly answers the paired basic item. The reasoning-cue example in Figure 2 illustrates the risk: the basic and hallucinated versions differ only in \"eyes momentarily widen\" versus \"eyes momentarily narrow,\" while both conclude \"a mix of amusement and fond recollection,\" so the hallucinated variant is plausibly contestable even for human raters. Without a human ceiling or per-category agreement measures, low model accuracy on such items may reflect label ambiguity rather than emotion hallucination. The authors should report human performance and inter-annotator agreement, at least for the perception subcategories.","section":"Appendix A and Section 3.3"},{"comment":"The headline comparison lacks statistical support. Many models score at or below the 25% random-chance level for paired accuracy, but the paper reports no confidence intervals, standard errors, or significance tests. For example, in Table 2 the difference between Gemini-2.5-Pro (51.58) and Gemini-2.5-Flash (50.56) is less than one percentage point, and the claim that \"most current models exhibit substantial issues\" needs an explicit comparison against chance that accounts for the finite number of items. The authors should report bootstrap or McNemar-style confidence intervals for the overall and per-category accuracies.","section":"Section 4.1, Tables 1 and 2"},{"comment":"The average improvement of 9.90% attributed to PEP-MEK is not reproducible from the reported numbers. Using the overall accuracies in Table 3, the gains are 20.15 - 10.49 = 9.66 for Qwen2.5-Omni, 26.03 - 9.65 = 16.38 for Emotion-LLaMA, and 37.84 - 33.44 = 4.40 for Gemini-2.5-Flash, which average to 10.15 percentage points, not 9.90. The authors should clarify how the 9.90% figure is computed (for example, whether it is weighted across subsets or computed over a different set of models) and report per-category gains and significance.","section":"Section 5, Table 3"}],"minor_comments":[{"comment":"The word \"mian\" appears instead of \"main\"; this typo should be corrected.","section":"Section 1"},{"comment":"The caption reads \"Emotion Konwledge\"; it should be \"Emotion Knowledge.\"","section":"Table 5 caption"},{"comment":"The text refers to \"VideoHallucer\" in the qualitative analysis; this should be \"EmotionHallucer.\" The same issue appears in the dataset statistics table, where \"Avg Short Vieo resolution\" and \"Avg Long Vieo resolution\" contain spelling errors.","section":"Figure 3 and Figure 4"},{"comment":"The phrase \"basic and hallucination QA pairs\" is ambiguous throughout the appendix; the authors should consistently state whether a \"QA pair\" contains one question or two complementary questions.","section":"Appendix A.1"},{"comment":"The text quotes GPT-4o-Audio's refusal response, but GPT-4o-Audio does not appear in Table 8; the authors should either include it in the table or remove the quote.","section":"Appendix C.1.4"},{"comment":"The GitHub repository is currently listed as a placeholder URL; providing the data or a detailed example bank in the supplement would support independent auditing of the benchmark.","section":"Benchmark release"}],"recommendation":"major_revision","confidential_remarks":"The paper claims to be the first emotion hallucination benchmark, and while the related work cites VideoHallucer, the novelty framing should be strengthened by clearly contrasting the emotion-specific design with existing general hallucination benchmarks. The main concerns for the editor are the internal arithmetic inconsistency in the benchmark size and the absence of human validation; these are fixable within a revision but are load-bearing for the benchmark's credibility. The evaluation is in principle reproducible given the detailed model/API list, but the missing data release and code prevent independent verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a real first: a dedicated benchmark for emotion hallucinations in MLLMs, with a seven-way taxonomy, paired adversarial QA, and unusually broad model coverage. If the labels are trustworthy, the main finding—current models, including Gemini-2.5-Pro, sit around 50% overall on the audio-free subset—is important. The knowledge-vs-perception gap also looks consistent across models.\n\nThe taxonomy is the genuine contribution. The paired design is borrowed from POPE and VideoHallucer, and the authors say so. But the paper would be far more useful if the data and annotation files were actually released; the abstract says they \"will be available,\" which is not enough for a benchmark paper.\n\nThe soft spots are concrete and addressable. The question count does not add up: Table 4 says 2,742 questions, while Section 3.1.1 and Appendix A sum to 2,390 pairs, which should mean 4,780 questions if each pair contains two. Something is off. Section 3.2's qualitative analysis refers to \"VideoHallucer,\" a clear sign of template reuse. The PEP-MEK improvement is also slightly misreported: the average from Table 3 is about 10.15%, not 9.90%, and \"16.38%t\" is a typo. These are minor, but in a benchmark paper they matter.\n\nThe load-bearing issue is label validity. The hallucinated variants are treated as unambiguously false, but there is no inter-annotator agreement and no human baseline. For intensity items (\"normal\" vs \"strong\" sadness) and reasoning-cue items (e.g., a narrowed eye vs a widened eye), the correct answer is plausibly contestable even for human raters. Since the paired scoring requires both answers correct, one contestable hallucinated variant depresses the overall score of any model that answered the basic item correctly. The paper itself admits in Appendix E that there is \"noise introduced by human annotations.\" Without a human ceiling, the headline \"most current models exhibit substantial issues with emotion hallucinations\" could partly be an artifact of benchmark ambiguity rather than genuine model failure. That is a measurement risk, not a philosophical disagreement.\n\nWho this is for: anyone working on MLLM hallucination evaluation or affective computing. The paper deserves a serious referee—the design is thoughtful, the gaps are fixable, and the contribution is real. I would want the data released, the counts corrected, and at least a small human-accuracy baseline before trusting the numbers, but this is a conditional-accept path, not a reject.\n\nSend it to review, but push for the missing evidence.","headline":"First emotion-hallucination benchmark with a genuinely useful taxonomy, but the validity of the labels is the load-bearing assumption and the paper is not yet auditable.","tokens_in":37196,"tokens_out":2938,"would_cite":true,"duration_ms":29688,"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":"Most multimodal AI models hallucinate emotions at high rates, according to the first benchmark built to measure it.","keywords":["emotion hallucination","multimodal large language models","benchmark","adversarial question answering","emotion psychology knowledge","multimodal emotion perception","hallucination evaluation","PEP-MEK"],"falsifier":"Randomly sample about 200 EmotionHallucer pairs, have independent expert annotators relabel them following the paper's own instructions without seeing model outputs, and measure inter-annotator agreement with the published ground truth; if agreement is low (e.g., Cohen's kappa below 0.6) or if experts frequently judge the 'hallucinated' variant as plausible or correct, then the models' scores are being measured against contestable labels rather than true emotion hallucination.","tokens_in":36255,"feed_emoji":"🎭","tokens_out":8680,"duration_ms":77738,"temperature":0.7,"pith_summary":"EmotionHallucer is claimed to be the first benchmark dedicated to measuring emotion hallucinations in multimodal large language models (MLLMs). The paper's central claim is that most current models hallucinate emotions at substantial rates: on the audio-free subset, even the best closed-source model scores only about 52 percent overall, where 25 percent is random chance for the paired question design. The benchmark splits emotion hallucination into two dimensions—emotion psychology knowledge (theory, definitions, empirical findings) and real-world multimodal perception (category, intensity, reasoning result, reasoning cue)—and uses adversarial binary question pairs to test whether models accept or reject subtle distortions. A sympathetic reading cares because emotion hallucinations matter for deployed affective systems in mental health, education, and human–computer interaction, where confidently wrong emotional readings could cause real harm.","feed_headline":"Top AI models score near chance at detecting emotion hallucinations","feed_subtitle":"A new benchmark tests 38 models on emotion psychology and multimodal perception; best scores hover near 50 percent.","key_machinery":"The load-bearing instrument is the adversarial binary question–answer pair. Each pair consists of a basic question, grounded in a trusted source (textbook statement or dataset label), and a hallucinated counterpart that deliberately distorts, exaggerates, or inverts the emotional content; both questions are yes/no with balanced answers, and a correct pair response requires the model to answer both in accordance with ground truth. This design converts the slippery concept of emotion hallucination into a measurable tendency to accept fabricated content or reject valid content. The secondary mechanism is the PEP-MEK prompting framework, which extracts modality-specific and emotion knowledge, makes an initial prediction, produces an explanation, verifies it, and then gives a final binary answer; it demonstrates that the measured weakness can be partially mitigated without retraining.","core_discovery":"On its own terms, the paper establishes that hallucination in emotion understanding is pervasive and measurably distinct from object-level hallucination. EmotionHallucer presents 2,742 questions organized as 1,371 basic/hallucinated pairs across text, image, audio, short video, and long video, with knowledge-based items drawn from a standard emotion psychology textbook and perception items inheriting labels from existing affective datasets. A model is scored correct on a pair only when it answers both the basic and the deliberately distorted question correctly. Evaluated across 38 LLMs and MLLMs, the paper reports that open-source models generally fail to exceed random 25 percent pair accuracy on the full multimodal benchmark, that closed-source models perform better, that reasoning capability confers an additional advantage, and that models perform better on emotion knowledge than on multimodal perception. The paper also demonstrates that a simple prompting framework, PEP-MEK, which adds modality and emotion knowledge extraction plus an explanation-before-final-answer step, improves hallucination detection by an average of 9.90 percentage points on selected models.","pith_inferences":["Because the paper reports no human baseline or inter-annotator agreement, the absolute accuracy numbers should be read as upper bounds on clean measurement; a natural follow-up is to measure expert agreement on a sample of items to validate the ground truth.","The adversarial pair design could transfer to other subjective domains where ground truth is contestable (e.g., social inference, humor, aesthetic taste), since the paired format compares relative faithfulness rather than requiring an absolute label.","The particularly low audio performance suggests that paralinguistic emotion perception is the next bottleneck; a testable prediction is that models with dedicated prosody training objectives will close the gap more than generic ASR-oriented audio models.","A cautious reading of the 'around 50 percent' headline number is that it is twice the 25 percent chance baseline but still far from reliable; readers should compare accuracy against the pair-chance baseline rather than the 50 percent yes/no baseline when interpreting the benchmark."],"forward_implications":["If the results hold, emotion-aware MLLMs cannot yet be trusted for high-stakes affective applications such as mental-health screening, empathetic dialogue, or automated customer care, since the best models still fail on roughly half of adversarial pairs.","The pair-based scoring shows that the failure mode is specifically hallucination detection, not basic understanding: many models answer basic questions well but accept distorted variants, so the bottleneck is robustness to subtle misinformation.","The knowledge-versus-perception gap implies that progress depends on better multimodal perception training and emotional data, not simply on adding more factual emotion knowledge.","The benefit of reasoning models and of PEP-MEK's explain-before-answer step suggests that reasoning-augmented prompting is a practical, low-cost lever for reducing emotion hallucination."],"supporting_citations":[{"why":"Single authoritative textbook source for all emotion psychology knowledge items (theory, definition, finding).","marker":"[69]"},{"why":"Binary question-answer evaluation framework adapted to detect hallucination; supplies the yes/no pair design.","marker":"[42]"},{"why":"Adversarial binary question construction that pairs a basic question with a hallucinated counterpart.","marker":"[75]"},{"why":"Hallucination evaluation methodology and bias metrics (percent difference, false positive ratio) applied in the benchmark.","marker":"[77]"},{"why":"Source dataset for real-world text perception items (sentiment and opinion understanding).","marker":"[15]"},{"why":"Source dataset for image perception items, with tweet sentiment labels.","marker":"[93]"},{"why":"Source dataset for audio perception items with intensity-labeled emotional speech.","marker":"[53]"},{"why":"Source dataset for short-video perception items with emotion labels.","marker":"[44]"},{"why":"Source dataset for long-video social emotion perception items.","marker":"[87, 80]"}],"fun_headline_variants":["AI near chance at spotting emotion hallucinations","New benchmark: best models barely detect emotion hallucinations","MLLMs stumble on emotion hallucination detection","EmotionHallucer: most AI fail at detecting emotion hallucinations","Models near chance in emotion hallucination test"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire benchmark rests on the assumption that its ground-truth labels are correct—knowledge items from one psychology textbook and perception labels inherited from existing datasets—so if any of those labels are contestable, the reported accuracy numbers do not cleanly measure emotion hallucination.","fun_headline_variants_meta":{"raw":{"variants":["AI near chance at spotting emotion hallucinations","New benchmark: best models barely detect emotion hallucinations","MLLMs stumble on emotion hallucination detection","EmotionHallucer: most AI fail at detecting emotion hallucinations","Models near chance in emotion hallucination test"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1367,"prompt_tokens":1064,"completion_tokens":303,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":680,"completion_tokens_details":{"reasoning_tokens":232}},"tokens_in":680,"tokens_out":303,"duration_ms":3014,"temperature":1.0,"reasoning_tokens":232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:53:50.921443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Randomly sample about 200 EmotionHallucer pairs, have independent expert annotators relabel them following the paper's own instructions without seeing model outputs, and measure inter-annotator agreement with the published ground truth; if agreement is low (e.g., Cohen's kappa below 0.6) or if experts frequently judge the 'hallucinated' variant as plausible or correct, then the models' scores are being measured against contestable labels rather than true emotion hallucination.","supporting_citations":[{"cited_title":"Shiota and James W","cited_arxiv_id":null,"evidence_quote":"Single authoritative textbook source for all emotion psychology knowledge items (theory, definition, finding)."},{"cited_title":"Adaptive co-attention network for named entity recognition in tweets","cited_arxiv_id":null,"evidence_quote":"Source dataset for image perception items, with tweet sentiment labels."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Source dataset for audio perception items with intensity-labeled emotional speech."},{"cited_title":"Mer 2023: Multi-label learning, modality robustness, and semi- supervised learning","cited_arxiv_id":null,"evidence_quote":"Source dataset for short-video perception items with emotion labels."}],"review_version":1}