{"id":"b89eb654-6ebb-4a4f-9186-d3e7c5c4e1dd","arxiv_id":"2508.04017","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Large multimodal models mostly fail to proactively detect flawed textual premises, and their performance depends on error type and on how they weight text versus images.","lead":"This paper presents ISEval, a benchmark testing whether large multimodal models can detect faulty inputs on their own. It finds most models fail to spot textual errors unless explicitly asked to check, which matters for building AI systems that don't silently reason from wrong information.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline finding may conflate failure to volunteer scrutiny with failure to detect it; benchmark validity unverified.","rationale":"The reader's verdict is UNVERDICTED, and my analysis does not move it. I agree with the reader that the taxonomy and metrics are unproven, but I identify a sharper threat to the central claim: the 'without guidance' manipulation may measure response-format compliance rather than internal scrutiny. This is a specificity improvement over the reader's weakest assumption. The central claim is empirically plausible, and the availability of code is a credit, but because the full manuscript is unreadable in the provided input, the necessary details—prompt wording, scoring rules, controls, human baselines—cannot be verified from the paper itself. The proposed concrete test would settle whether the headline result survives the format confound. If it does not, the paper's main conclusion would need to be weakened to a statement about spontaneous reporting of detected errors, not about failure to detect them. No author-focused or theatrical criticism is intended; this is a standard construct-validity check for a behavioral evaluation.","tokens_in":13377,"tokens_out":2292,"duration_ms":34297,"concrete_test":"Using the public repository (https://github.com/MLGroupJLU/LMM_ISEval), run the ten models on a matched subset of the ISEval items in two conditions: (A) the original open-ended, no-guidance setting, and (B) a forced-choice setting that directly asks 'Is there an error in the premise? Yes/No' for the same items, counterbalanced for order. Compute per-model detection accuracy in both conditions, and also insert 100 valid (flaw-free) control premises to measure false-positive rates. If detection accuracy is high in (B) but low in (A), the headline conflates not volunteering with not detecting. Additionally, have two independent human annotators apply the three metrics to the same model outputs; if inter-annotator agreement is below a pre-specified threshold (e.g., Cohen's kappa < 0.7), the scoring rubric is not reliable enough to support the comparative claims.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that most LMMs 'struggle to actively detect flawed textual premises without guidance,' with error types and modality-trust differences as secondary findings. For that claim to hold, the evaluation must isolate the models' ability to recognize faulty inputs from their tendency to comment on them in open-ended generation. The abstract alone does not establish this. In particular, the contrast between 'without guidance' and 'with explicit prompts' may primarily reflect instruction-following and response-format conventions: a model that correctly identifies a premise error when asked a direct question but does not spontaneously mention it in a free-form answer would be scored as 'struggling to actively detect,' even though its internal detection is intact. The seven flawed-premise categories and three metrics are also asserted without evidence of human validation, false-positive control, or inter-annotator agreement. Because the full text supplied here is garbled, neither the prompt templates, scoring rubrics, nor control conditions can be inspected. The strongest load-bearing assumption is therefore that the ISEval task format measures scrutiny ability rather than response style, and that the scoring metrics reward genuine detection rather than superficial phrasing. If this assumption fails, the headline result—'most models struggle'—does not generalize to real-world active scrutiny.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ISEval, a benchmark framework for evaluating whether large multimodal models (LMMs) actively scrutinize faulty inputs. It proposes seven categories of flawed premises, three evaluation metrics, and reports an evaluation of ten LMMs. Based on the abstract, the main findings are: (i) most models fail to actively detect flawed textual premises unless explicitly prompted; (ii) performance varies by error type, with logical fallacies easier than surface-level linguistic errors and some conditional flaws; and (iii) models differ in how they weight visual versus textual evidence in conflicts. Code is released. My review is limited to the abstract because the supplied full-text file is corrupted and unreadable.","tokens_in":13663,"tokens_out":4856,"duration_ms":66969,"significance":"If the construct validity of the benchmark can be established, this is a timely and useful contribution: proactive input verification is distinct from passive acceptance and is important for reliable LMM deployment. Releasing code and evaluating ten models are concrete strengths. However, the contribution stands or falls on the validity of the metrics and on whether 'active' detection is measured as a distinct capability rather than as a response-format artifact.","major_comments":[{"comment":"The headline claim that models 'struggle to actively detect flawed textual premises without guidance' conflates failure to volunteer scrutiny with failure to detect it. If the evaluation scores only whether a model spontaneously mentions the error in an open-ended answer, a model that correctly identifies the error when asked directly would still be scored as failing. The paper needs a control condition: present the same flawed premises with an explicit prompt such as 'Is there any error?' and report the agreement/disagreement with the open-ended condition. Without this, the central claim is not established.","section":"Abstract (first finding)"},{"comment":"The abstract lists seven categories of flawed premises and three evaluation metrics but gives no definitions, examples, scoring rubrics, or validation. The reader cannot assess whether the taxonomy is comprehensive or whether the scoring rewards genuine detection rather than superficial phrasing. The modality-trust finding, in particular, requires that the conflict stimuli are unambiguously flawed; otherwise 'over-rely on text' could be an artifact of ambiguous items. The paper should report human annotation protocols, inter-annotator agreement, and false-positive rates on valid inputs.","section":"Abstract (ISEval description)"},{"comment":"All findings are stated qualitatively: 'most models struggle,' 'models excel at logical fallacies but struggle with surface-level errors,' 'aya-vision-8b over-rely on text.' No numbers, error bars, or statistical tests are reported in the abstract. Because the supplied full text is not decipherable, I cannot verify whether the body contains the necessary tables and significance tests. If those details are absent, the conclusions outrun the evidence; if they are present, the abstract should include representative numeric results.","section":"Abstract (quantitative claims)"}],"minor_comments":[{"comment":"The abstract should enumerate the ten evaluated models with exact version numbers; 'aya-vision-8b' is not a standard instantiation and needs a citation.","section":"Abstract"},{"comment":"The phrase 'reflects a strong reliance on explicit prompts' is causal; the design is observational. Rephrase to 'suggests' or 'is consistent with' to avoid overstatement.","section":"Abstract"},{"comment":"The provided full-text file is heavily corrupted (mojibake). If this reflects the arXiv PDF's text layer, please re-upload; otherwise, the submission package should include a readable version. I could not inspect the methods, tables, or appendices.","section":"Full text"}],"recommendation":"major_revision","confidential_remarks":"My review is necessarily based on the abstract because the full text was not decodable. The main stress-test concern—that 'active detection' may be confounded with response style—is real and load-bearing. If the full paper already includes a direct-question control condition and validation of the taxonomy, then the revision needed may be local; if not, the benchmark's headline claim is not yet supported. Given the potential significance, a major revision with an explicit control analysis and validator agreement statistics would resolve the issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the quick version: this paper is about whether LMMs catch bad inputs on their own, not just when prompted. That is a real question, and the proposed benchmark is a reasonable first cut. But the headline result may overstate what the data show, and I can't verify the data because my copy of the full text arrived garbled.\n\nWhat's new: the seven-type taxonomy of flawed premises and three metrics are a genuine attempt to measure input scrutiny, not just task performance. The finding that models are better at logical fallacies than surface-level linguistic errors is plausible and useful. Modality trust being model-specific is also a nice result. If the code is public and the scoring is sound, this becomes a useful eval for safety testing.\n\nThe soft spots: first, the distinction between 'struggling to detect' and 'not volunteering' is under-specified. The paper contrasts 'without guidance' vs 'with explicit prompts,' but if the 'without guidance' condition produces free-form answers, a model that knows there's an error but does not flag it will be scored as failing. That conflates ability with response style—exactly the stress-test concern, and it lands. The authors need to show that when directly asked about the premise, models that fail the active condition still fail; otherwise the conclusion is about proactive reporting, not detection. This is the load-bearing issue.\n\nSecond, the abstract contains no quantitative results: no table, no error bars, no agreement statistics. Maybe the body of the paper has all that, but I can't tell. The taxonomy's validity also needs human-validation or at least a reliable rubric and inter-annotator agreement.\n\nThird, the 'unexplored' claim is a bit strong; there is adjacent work on false-premise reasoning in language models, though the multimodal angle is less studied.\n\nWould a serious referee take it? Yes. The benchmark idea is worth engaging with, and the failure mode I'm worried about is testable. Send it out and ask for the active-versus-prompted details. If the result survives that check, it's solid; if it doesn't, the paper still reports useful model comparisons but the headline should be softened. If you can access the GitHub repo, read the scoring code first—that's the quickest check.","headline":"A useful benchmark for a real question, but the headline finding likely muddles 'didn't say anything' with 'didn't notice,' and I couldn't verify the full text.","tokens_in":629,"tokens_out":830,"would_cite":false,"duration_ms":49852,"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":"Large multimodal models mostly fail to actively detect flawed textual premises, and only explicit prompts trigger reliable scrutiny.","keywords":["input scrutiny","large multimodal models","benchmark","premise verification","hallucination","error detection","modality trust","ISEval"],"falsifier":"Take the ISEval items and add newly generated flaws outside the seven categories, presented both with and without explicit verification instructions. If models detect the new flaw types unguided at rates close to their guided rates, the claimed passive-acceptance failure is specific to the chosen taxonomy rather than general input scrutiny.","tokens_in":13323,"feed_emoji":"🤖","tokens_out":4089,"duration_ms":53712,"temperature":0.7,"pith_summary":"This paper tries to establish that large multimodal models do not proactively check whether the textual premises they are given are valid. It introduces ISEval, a benchmark with seven categories of flawed premises and three evaluation metrics, and reports results across ten models. The headline finding is a gap between guided and unguided performance: models detect errors when explicitly told to look for them, but much less often on their own. If true, this means reliance on explicit prompt instructions, not genuine vigilance, is what keeps these systems from reasoning on false premises.","feed_headline":"Most multimodal AI models accept bad text unless prompted to check","feed_subtitle":"Benchmarking ten large multimodal models shows explicit verification prompts are needed to catch faulty inputs.","key_machinery":"ISEval, the Input Scrutiny Ability Evaluation Framework: a set of seven categories of flawed premises and three evaluation metrics applied to multimodal inputs. Its defining move is to compare model responses with and without an explicit instruction to scrutinise the premise, so that 'active' scrutiny is measured separately from prompt-following.","core_discovery":"The paper's central claim is that large multimodal models are strong at following instructions to verify inputs but weak at spontaneously doing so. On ten models, unguided detection of flawed textual premises is generally low; performance rises when the prompt explicitly asks for verification. Error type modulates the result: logical fallacies are detected more often than surface-level linguistic errors and certain conditional flaws. Model behaviour also differs in how much it trusts text versus image evidence: some models balance both, while at least one model over-relies on textual content, missing contradictions in the image. The practical upshot is that current LMMs cannot be assumed to","pith_inferences":["The guided-versus-unguided gap suggests models may possess verification capability that is not deployed by default; a testable hypothesis is that decoding-time self-verification methods will close the gap without changing weights.","Because passive acceptance of faulty premises is a known contributor to hallucinated answers, ISEval scores could be checked against hallucination-benchmark scores to see whether input scrutiny is a separable skill.","The taxonomy is text-centric; a natural extension is to faulty visual premises, such as doctored or contradictory images, which would test whether the modality-trust findings hold when the image, not the text, is wrong."],"forward_implications":["Applications built on multimodal models should add an explicit verification stage rather than assume the model vets its own inputs.","Benchmarks should report per-category results, since aggregate scores hide the gap between logical-fallacy detection and surface-level language errors.","The finding that some models over-trust text over images implies conflict-resolution training is a needed target for improvement.","ISEval provides a common yardstick for comparing future models on input scrutiny, not just on output accuracy."],"supporting_citations":[],"fun_headline_variants":["Multimodal AI models rarely catch bad inputs on their own","AI models fail to flag bad text unless prompted to check","Study: LMMs rarely catch bad inputs without explicit prompts","Most multimodal models overlook bad inputs unless prompted"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The benchmark's seven categories and three metrics are assumed to be a representative and meaningful measure of 'input scrutiny ability'; if the taxonomy is arbitrary or the scoring rewards phrasing, the headline failure rates will not generalize to real faulty inputs.","fun_headline_variants_meta":{"raw":{"variants":["Multimodal AI models rarely catch bad inputs on their own","AI models fail to flag bad text unless prompted to check","Study: LMMs rarely catch bad inputs without explicit prompts","Most multimodal models overlook bad inputs unless prompted"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000712,"raw_usage":{"total_tokens":3041,"prompt_tokens":745,"completion_tokens":2296,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2230}},"tokens_in":489,"tokens_out":2296,"duration_ms":21196,"temperature":1.0,"reasoning_tokens":2230,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:55:57.239821+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the ISEval items and add newly generated flaws outside the seven categories, presented both with and without explicit verification instructions. If models detect the new flaw types unguided at rates close to their guided rates, the claimed passive-acceptance failure is specific to the chosen taxonomy rather than general input scrutiny.","supporting_citations":[],"review_version":1}