{"id":"b4bd1577-ffbe-49ad-b3f5-f813bff35841","arxiv_id":"2501.02135","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 600K-question audio-visual trustworthiness benchmark reveals that current AVLLMs are brittle under mismatches and missing modalities, and CAVPref, a calibrated preference-optimization method, improves their accuracy.","lead":"AVTrustBench adds three trust tests for audio-visual AI: catching mismatched audio and video, refusing answers when choices are wrong, and spotting swapped event order. Across 600K questions, 13 leading models often score near chance, while a proposed training method, CAVPref, improves them substantially.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Benchmark ground truth is validated on only ~0.25% of samples; residual label errors would directly corrupt model accuracies, rankings, and the CAVPref gains.","rationale":"I agree with the reader's weakest_assumption. I considered alternative concerns: CircularEval makes the 'near chance' statement technically wrong, and the CAVPref train/test split is underspecified, but these weaken framing or are secondary to label validity. The most load-bearing issue remains that both the benchmark's descriptive claims and the CAVPref training/evaluation depend on labels generated by unvalidated heuristics, and the validation evidence is too thin and uses the wrong statistic. The paper already includes external transfer results and a human study, so I would not reject or call it unverdictable; I would keep the CONDITIONAL verdict and require release plus independent label verification before acceptance.","tokens_in":35674,"tokens_out":6785,"duration_ms":70018,"concrete_test":"Release the benchmark and have independent annotators verify a stratified random sample of 1,000 items per task (9,000 total), marking for each item whether the intended answer is correct and unique; for MVT/MAT also whether both modalities are truly necessary, and for MCIT/ICIT/MVIT/MAIT whether the expected 'None/mismatch' label is justified. Recompute per-task model accuracies and CAVPref deltas on the verified subset. If per-task label accuracy is below ~98%, or if any model ranking or the direction/size of CAVPref's improvements changes materially, the headline findings require revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claims—that AVLLMs are near-chance on trust tasks and that CAVPref yields up to 30.19% gains—rest on the correctness of the semi-automatic QA pipeline for all ~600K samples (Sec. 3.2, Supp. A.1-A.2). The pipeline is validated on only 500 randomly chosen samples per axis, and the reported 0.979 Spearman correlation is not a per-item label-accuracy rate and is not disaggregated by task or by construction step (Supp. N.1). For MVT and MAT, the intended answer 'the video/audio is missing' presupposes that the removed modality is genuinely required; the manual filter that excluded ~50% of MUSIC-AVQA samples is not audited at scale, so residual single-modality questions would make modality-dependency scores uninterpretable. For MCIT/ICIT/MVIT/MAIT, the expected 'None of the above'/mismatch response requires that the lookup-table replacements are unambiguously wrong; any wrong or ambiguous replacement changes the ground truth. Because reported accuracies are 20-50%, even a few percent of corrupted labels can alter model rankings and exaggerate both the model-human gap and CAVPref's advantage. The human study reports 91.27% but was run on a hand-picked 50-success/50-failure sample per task, so it does not validate the full label set. The external transfer results (Supp. D.6) are helpful but only cover one model and do not rescue the benchmark's internal validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces AVTrustBench, a 600K-sample multiple-choice benchmark for evaluating audio-visual LLMs (AVLLMs) across three axes: adversarial attack (MCIT, ICIT, MVIT, MAIT), compositional reasoning (COT-Stitch, COT-Swap, CAT), and modality-specific dependency (MVT, MAT). The authors report zero-shot evaluations of 13 AVLLMs, finding that most perform near or only slightly above chance. To improve robustness, they propose CAVPref, a preference optimization method with modality-conditioned losses and a distributionally robust optimization (DRO) component, and report gains up to 30.19% across all nine tasks on three AVLLMs (and seven more in the supplement). The paper also includes a human study estimating 91.27% accuracy on a selected subset.","tokens_in":35935,"tokens_out":6320,"duration_ms":56413,"significance":"If the benchmark labels are reliable, AVTrustBench fills a real gap: no existing benchmark systematically tests audio-visual trustworthiness, and the reported failure modes are important. The CAVPref method is a sensible extension of DPO that conditions on audio and visual correctness and uses DRO to mitigate tail-category underperformance; the broader applicability is evidenced by transfer results to VideoBench and MVBench (Supp. D.6) for one model. The paper provides the community with a large, multi-task evaluation suite and a clear mitigation strategy. The main risks are the limited validation of the semi-automatic ground-truth generation and the tuning of CAVPref hyperparameters on the benchmark itself.","major_comments":[{"comment":"The benchmark's core claims rest on the correctness of the semi-automatic QA labels. The validation reported in Supp. N.1 samples only 500 examples per axis (0.25% of the 600K corpus) and reports a Spearman correlation of 0.979, which is not a per-item label accuracy. Because reported accuracies are 20–50%, even a small fraction of erroneous labels can change model rankings and the model-human gap. I request a per-item human accuracy on a random sample of the AVTrustBench-test set, stratified by the nine tasks, plus agreement rates for each construction step (look-up table replacement, LLM distractor generation, modality-dependency filtering).","section":"Sec. 3.2 / Supp. N.1"},{"comment":"Modality-specific dependency tasks (MVT, MAT) assume that both audio and video are required to answer the question. The pipeline removes ~50% of MUSIC-AVQA samples for this suite via a manual 'careful inspection' (Supp. A.1), yet no inter-annotator agreement or audit is reported. Residual single-modality questions would make the expected answer ('the video is missing' / 'the audio is missing') ambiguous, rendering the MVT/MAT scores uninterpretable. Please quantify the reliability of the exclusion filter, e.g., by having independent annotators re-label a random sample of the retained and excluded questions.","section":"Supp. A.1"},{"comment":"The hyperparameters λy, λV, λA are chosen by ablation on MCIT, MVIT, and MAIT from the same benchmark used to report the headline improvement (Fig. 6). The paper does not specify whether these tasks overlap with the evaluation split. If the same tasks (or a subset) are used for both hyperparameter selection and final reporting, the 30.19% gain is optimistically biased. Please use a separate validation split for hyperparameter selection, or show that the performance gains persist across a range of λ values (e.g., λ ∈ {0.2, 0.5, 1.0, 2.0}) on all nine tasks.","section":"Sec. 5.3 / Eq. (5)"},{"comment":"The human study in Supp. N.1 manually selects 50 successful and 50 failed GPT-4o cases per task, yielding a 91.27% average accuracy. This sample is deliberately biased and cannot estimate the actual human accuracy on the full benchmark; the claim that tasks are 'straightforward for human subjects' is therefore not established by this protocol. Please provide a human evaluation on a random sample (e.g., 50–100 per task) drawn from the full test set, or at least report accuracy separately for the two halves and justify the selection.","section":"Supp. N.1"}],"minor_comments":[{"comment":"The abstract says '600K samples' but the test set is 181K; please clarify the train/validation/test split and the source of training samples for CAVPref.","section":"Sec. 3.2"},{"comment":"The 'Temporal?' column uses checkmarks for several benchmarks, but the meaning is not always clear (e.g., does 'Temporal' mean the tasks require temporal reasoning or that the benchmark uses videos?); consider a footnote.","section":"Table 1"},{"comment":"Eq. (4) introduces a general D_f divergence, but the closed-form solution in Eq. (5) is derived for KL divergence (Supp. D.2); please unify the notation or state when D_f is restricted to KL.","section":"Eqs. (4)–(5)"},{"comment":"The reference list contains duplicate entries ([35] and [36] are the same Silkie paper); please consolidate.","section":"References"},{"comment":"The right panel's dimension-wise averages would be easier to read if the three dimensions were labeled on the bars or in the caption.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an important gap and the benchmark scale is impressive, but the validation of the semi-automatic label generation is too thin for the conclusions drawn. The CAVPref improvement claim also needs a cleaner evaluation protocol. These issues are fixable with additional experiments, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The benchmark part of this paper is genuinely useful; the training part is more fragile than the headline suggests. AVTrustBench is the first benchmark I know that puts audio-visual LLMs through adversarial mismatches, compositional order/attribute binding, and modality ablation under one multiple-choice protocol. The zero-shot evaluation of 13 models is extensive, and the CircularEval protocol is a good safeguard against positional bias. The human study (91.27% on 900 selected samples) shows the tasks are answerable, which makes the near-chance model scores meaningful. That part deserves a serious referee.\n\nThe soft spots are real but not fatal. Ground truth is validated on only 500 samples per axis (roughly 0.25% of the 600K), and a Spearman correlation of 0.979 is not a per-item label accuracy. For MVT/MAT, the claim that both modalities are essential depends on a manual filter that removed half of MUSIC-AVQA; residual single-modality questions would make those numbers uninterpretable. For the adversarial tasks, wrong or ambiguous lookup-table replacements would corrupt the answer keys. Since reported accuracies sit at 20-50%, even a few percent label noise can shift rankings. The human study is on a hand-picked 50-success/50-failure sample per task, so it validates solvability, not the whole label set.\n\nCAVPref is a legitimate extension of DPO with modality-conditioned terms and a DRO component, and the gains are consistent across ten models. But the claim of up to 30.19% improvement is a percentage-point gain in-distribution, not a relative one. The hyperparameters are chosen by ablation on the benchmark itself, the preference data construction is underspecified, and the beta calibration leans on the authors' earlier AVSM. The transfer results on VideoBench and MVBench are a useful external check, but they cover one model, so they don't rescue the internal validity.\n\nWho benefits: anyone building or evaluating audio-visual LLMs. The zero-shot leaderboard alone is worth publishing. For the training claims to hold up, the authors need to release data and code, specify the train/test split, report error bars and relative gains, and audit the label filters at scale. That is revision work, not a rejection. Send it out; a good referee will ask for exactly those things.","headline":"Useful benchmark, fragile training claims; the zero-shot evaluation deserves peer review, but CAVPref's headline gains need far more scrutiny before being taken at face value.","tokens_in":36551,"tokens_out":2601,"would_cite":true,"duration_ms":23173,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Current audio-visual LLMs score near chance when inputs are mismatched or missing a modality, and a proposed preference-training method lifts accuracy by up to 30.19% across nine tasks.","keywords":["audio-visual large language models","trustworthiness benchmark","adversarial attacks","compositional reasoning","modality dependency","preference optimization","distributionally robust optimization","multiple-choice QA"],"falsifier":"Take 2,000 held-out AVTrustBench questions that were not part of the 500-per-axis validation set and have a small human panel answer them independently; if human-model agreement drops well below the claimed Spearman 0.979, or if humans cannot reach near the 91% ceiling on the full test set, the benchmark's scores would not measure what the paper says. A second check: for a random subset of MVT and MAT questions, have annotators judge whether both modalities are genuinely required; if many are answerable from one modality alone, the modality-dependency results collapse.","tokens_in":35431,"feed_emoji":"🎧","tokens_out":7966,"duration_ms":71663,"temperature":0.7,"pith_summary":"AVTrustBench is a 600,000-sample multiple-choice benchmark that probes audio-visual large language models under three conditions previous benchmarks did not cover: adversarial inputs where the answer is missing or irrelevant or a modality is swapped, compositional inputs where the order or attribute binding of events is changed, and inputs where one modality is deleted. The paper evaluates 13 models and finds that on the adversarial and compositional suites most of them perform near chance, and that all models are better at noticing a bad video than a bad audio track. The authors then propose CAVPref, a model-agnostic training loss built from three preference terms -- text, video, and audio -- each wrapped in a distributionally robust average, and report gains up to 30.19% over zero-shot accuracy across all nine tasks. A sympathetic reader would take away that trust in audio-visual LLMs should not be assumed out of the box, and that a training signal explicitly conditioned on both modalities measurably improves it.","feed_headline":"Benchmark: audio-visual LLMs score near chance when inputs clash","feed_subtitle":"A 600K-sample suite tests 13 models; a preference-training fix lifts accuracy by up to 30 percent.","key_machinery":"The load-bearing mechanism is a calibrated preference-optimization objective with three additive terms: $L_y$ for text preference between a winning and losing response, $L_V$ for visual conditioning (the winning response under the correct versus a mismatched video), and $L_A$ for audio conditioning (the winning response under the correct versus a mismatched audio). Each term is scaled by a $\\beta$ that grows with the normalized similarity difference between the winning and losing scenarios, using CLAP score differences for $L_y$ and an audio-visual similarity metric for $L_V$ and $L_A$. Every term is then passed through a distributionally robust transform $-\\lambda \\log \\mathbb{E}_P[e^{L/\\lambda}]$, which replaces the average loss with a worst-case expected loss over distributions that stay close to the training data. The final loss is $L_{\\mathrm{CAVPref}} = L_y^R + \\eta L_V^R + \\gamma L_A^R$, where the binary switches $\\eta$ and $\\gamma$ turn on only for tasks in which the video or audio stream, respectively, is the perturbed one. This structure is what lets the model learn to flag inconsistencies in each modality rather than merely imitate text preferences.","core_discovery":"On its own terms, the paper's central claim is that current audio-visual large language models do not genuinely understand audio-visual consistency. When offered multiple-choice answers with no correct option, when the video or audio is swapped for an unrelated event, when two events are reordered, or when the same words are bound to different attributes, most of the 13 evaluated models pick an offered option instead of saying that the input is inconsistent, and their accuracy hovers around chance on the adversarial and compositional suites. The paper further claims that this failure is fixable by training: CAVPref, which adds visual- and audio-conditioned preference terms to a standard DPO loss and applies a distributionally robust transform, improves accuracy by up to 30.19 percentage points across all nine tasks, with the largest gains on exactly the tasks that require detecting a missing or mismatched modality.","pith_inferences":["Editorial inference: the same three-term construction could be applied to image-only or audio-only LLMs by dropping the unused modality term, suggesting a general recipe for modality-consistency fine-tuning that the paper does not state.","Editorial inference: because the evaluation adopts circular assessment (a question is solved only if answered correctly under all rotations of the answer choices), the reported accuracies are lower-bound estimates; real open-ended reliability may be even worse than the tables show.","Editorial inference: a testable extension would be to measure per-category accuracy before and after CAVPref on a held-out long-tail audio-visual set, checking whether the robustness module transfers beyond the categories in this benchmark.","Editorial inference: another testable extension is to regenerate the preference data with a different large language model and rerun CAVPref, which would separate the method's contribution from the particular choice of distractor generation."],"forward_implications":["If the benchmark scores are right, no current audio-visual LLM is reliable enough to deploy in settings where a wrong answer can cause harm; near-chance accuracy on adversarial and compositional tasks means the models are effectively guessing.","The up-to-30.19% gains on three open models suggest a model-agnostic training recipe that can be applied to any audio-visual LLM without changing its architecture.","The large gap between video-removal and audio-removal accuracy implies that existing models lean on video; CAVPref narrows that gap, so modality-balanced training is an achievable goal.","The compositional suite's near-chance performance, together with the paper's 7B-versus-13B comparison, implies that simply scaling up the language backbone will not fix compositional understanding.","The release of the 600K-sample suite provides a standardized testbed that future audio-visual LLMs can be measured against for trustworthiness."],"supporting_citations":[{"why":"AVQA dataset supplies the video-audio QA pairs and category labels used to build the adversarial attack tasks.","marker":"[74]"},{"why":"MUSIC-AVQA dataset supplies the instrument-heavy QA pairs for adversarial attack and modality-dependency tasks and their category labels.","marker":"[31]"},{"why":"AudioSet supplies the real-world audio events used to construct the compositional reasoning tasks.","marker":"[19]"},{"why":"Direct preference optimization (DPO) is the baseline preference-optimization objective that CAVPref extends and is compared against.","marker":"[56]"},{"why":"MMBench's CircularEval strategy and GPT-4-based answer extraction are adopted as the evaluation protocol.","marker":"[42]"},{"why":"Winoground's attribute-binding formulation inspires the CAT compositional attribute task.","marker":"[65]"},{"why":"CLAP text-audio similarity scores are used to select AudioSet samples with strong audio-text alignment.","marker":"[18]"},{"why":"CLIP visual-text scores combine with CLAP to filter audio-visual samples for strong cross-modal alignment.","marker":"[55]"}],"fun_headline_variants":["AV LLMs near chance on mismatched audio-video inputs","New benchmark: AV LLMs can't spot audio-video mismatches","CAVPref training boosts AV LLM accuracy by up to 30%","13 AV LLMs flunk adversarial audio-video tests; fix up to 30%","Audio-visual trust gap: 13 models fail mismatch, fix gains 30%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark's scores stand on the assumption that the semi-automatically generated answer keys are correct for the roughly 99 percent of samples not checked by humans; if the look-up tables, LLM-generated distractors, or the single-modality filtering introduced errors, every accuracy number shifts.","fun_headline_variants_meta":{"raw":{"variants":["AV LLMs near chance on mismatched audio-video inputs","New benchmark: AV LLMs can't spot audio-video mismatches","CAVPref training boosts AV LLM accuracy by up to 30%","13 AV LLMs flunk adversarial audio-video tests; fix up to 30%","Audio-visual trust gap: 13 models fail mismatch, fix gains 30%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001586,"raw_usage":{"total_tokens":6327,"prompt_tokens":948,"completion_tokens":5379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":5277}},"tokens_in":564,"tokens_out":5379,"duration_ms":34670,"temperature":1.0,"reasoning_tokens":5277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:14:14.642528+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take 2,000 held-out AVTrustBench questions that were not part of the 500-per-axis validation set and have a small human panel answer them independently; if human-model agreement drops well below the claimed Spearman 0.979, or if humans cannot reach near the 91% ceiling on the full test set, the benchmark's scores would not measure what the paper says. A second check: for a random subset of MVT and MAT questions, have annotators judge whether both modalities are genuinely required; if many are answerable from one modality alone, the modality-dependency results collapse.","supporting_citations":[{"cited_title":"Avqa: A dataset for audio- visual question answering on videos","cited_arxiv_id":null,"evidence_quote":"AVQA dataset supplies the video-audio QA pairs and category labels used to build the adversarial attack tasks."},{"cited_title":"Learning to answer questions in dynamic audio-visual scenarios","cited_arxiv_id":null,"evidence_quote":"MUSIC-AVQA dataset supplies the instrument-heavy QA pairs for adversarial attack and modality-dependency tasks and their category labels."},{"cited_title":"Direct prefer- ence optimization: Your language model is secretly a reward model","cited_arxiv_id":null,"evidence_quote":"Direct preference optimization (DPO) is the baseline preference-optimization objective that CAVPref extends and is compared against."},{"cited_title":"Winoground: Probing vision and language models for visio- linguistic compositionality","cited_arxiv_id":null,"evidence_quote":"Winoground's attribute-binding formulation inspires the CAT compositional attribute task."}],"review_version":1}