{"id":"1a447036-c38c-43f8-93c2-3a6fdf89b2bb","arxiv_id":"2505.19509","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"MMKC-Bench provides a human-verified benchmark of multimodal knowledge conflicts and shows that current LMMs prefer internal parametric knowledge over external evidence.","lead":"The paper introduces MMKC-Bench, a new benchmark of 1,573 multimodal knowledge conflict instances to test how large multimodal models behave when external context contradicts what they already know. It finds that current models often stick to their internal knowledge rather than follow conflicting external evidence, which matters for building reliable retrieval-augmented systems.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"OAR baseline uses the same image in non-conflict and conflict settings, so high OAR on entity recognition and visual semantic conflicts may reflect following visual evidence, not parametric memory; the central claim is unsupported on the clean entity-knowledge subset.","rationale":"The reader's weakest-assumption analysis identifies the same central confound: the non-conflict baseline used to define OAR includes the same image that appears in the conflict setting, so OAR on recognition-based tasks may measure image-following rather than parametric-knowledge reliance. This is the most load-bearing concern because the paper's headline contribution is the empirical behavioral claim, not merely the benchmark resource. The paper provides useful human-verified data and a reproducible protocol, but the central claim needs re-analysis with a cleaner control. The concrete test I propose uses the paper's own entity-knowledge subset, where the answer cannot be read from the image; the reported numbers already suggest the effect reverses there, which would directly undercut the abstract's conclusion. I agree with the reader's conditional verdict: the concern is serious but addressable by revising the interpretation or the evaluation protocol, so no change in verdict is needed.","tokens_in":14006,"tokens_out":3291,"duration_ms":30035,"concrete_test":"Reproduce the analysis restricted to the entity knowledge conflict subset: for Qwen2.5-VL-7B, InternVL3-8B, and GPT-4o mini, compute OAR versus CAR under context-memory and inter-context settings, in both MCQ and open-ended formats. If OAR does not exceed CAR in a majority of these 12 cells, the claim that LMMs favor parametric knowledge over external evidence is not supported on the clean subset; additionally, run a text-only ablation for entity recognition and visual semantic conflicts, removing the image from the non-conflict baseline, to quantify how much of the original answer is driven by the image itself rather than by memory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline finding (Section 5.2, observation 1) is that LMMs favor internal parametric knowledge over external evidence because OAR exceeds CAR. OAR is defined (Section 5.1) as consistency with the model's answer in the non-conflict setting, where the non-conflict setting includes both the correct image and the correct text. For entity recognition and visual semantic conflicts, the correct image itself determines the original answer: an image of the Empire State Building, or of an OK gesture, carries the answer visually. In the conflict setting the same image is paired with a conflicting text, so a model that simply trusts the visual input will give the original answer and be counted as OAR. This conflates following external visual evidence with relying on parametric memory. The confound is not limited to entity recognition and visual semantics: averaging across conflict types amplifies the artifact. The clean test is the entity knowledge subset, where the image does not contain the queried fact (e.g., birth year). Table 3 (open-ended, context-memory) shows CAR exceeds OAR for Qwen2.5-VL-7B (0.62 vs 0.26) and InternVL3-8B (0.53 vs 0.27); GPT-4o mini also has CAR > OAR (0.47 vs 0.36). In Table 2 (MCQ), EK OAR is only roughly equal to CAR, and for InternVL3-8B context-memory CAR slightly exceeds OAR. Thus the central empirical claim is either reversed or unsupported on exactly the subset that isolates parametric knowledge from image evidence. The conclusion that LMMs favor internal knowledge may be an artifact of the measurement protocol rather than a property of the models.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces MMKC-Bench, a multimodal knowledge conflict benchmark covering context-memory and inter-context conflicts across three conflict types (entity recognition, entity knowledge, and visual semantic), with 1,573 knowledge instances and 3,381 images constructed through an automated pipeline with human verification. Nine LMMs from three families are evaluated on model-behavior analysis and conflict-detection tasks. The headline empirical claim, stated in the abstract and Section 5.2, is that while current LMMs can recognize knowledge conflicts, they tend to favor internal parametric knowledge over external evidence.","tokens_in":14299,"tokens_out":3966,"duration_ms":30175,"significance":"The benchmark itself addresses a real gap: existing multimodal conflict datasets mostly cover intra-memory conflicts, and the paper's pipeline for generating context-memory and inter-context conflicts with human verification is useful for the multimodal RAG community. The paper ships a public code repository and evaluates nine models across two question formats and two granularities of conflict detection, which is a solid engineering contribution. However, the central behavioral finding is not established because the OAR baseline is confounded, and the paper's own fine-grained results contradict the headline claim on the subset that actually isolates parametric knowledge. If the central claim were corrected, the benchmark could still be a valuable resource, but the current manuscript overstates what the experiments demonstrate.","major_comments":[{"comment":"The OAR baseline is confounded. OAR is defined as consistency with the model's answer in the non-conflict setting, where the non-conflict setting contains the correct image and the correct text. In entity recognition and visual semantic conflicts, the correct image itself determines the answer (e.g., an image of the Empire State Building or an OK gesture), so in the conflict setting the model can produce the 'original' answer by following visual evidence rather than by retrieving parametric knowledge. The clean subset, entity knowledge (e.g., birth year), does not have this confound because the image does not contain the queried fact. Yet Tables 2 and 3 show that on EK, CAR is comparable to or higher than OAR: in Table 3 open-ended context-memory, Qwen2.5-VL-7B OAR=0.26 vs CAR=0.62, InternVL3-8B OAR=0.27 vs CAR=0.53, GPT-4o mini OAR=0.36 vs CAR=0.47; in inter-context, Qwen2.5-VL-7B OAR=0.14 vs CAR=0.76. Thus the central claim that LMMs favor parametric knowledge is reversed on the subset that isolates parametric knowledge from image evidence. The paper must redefine the parametric-knowledge baseline (e.g., answer with the image only, or with no image) and report conflict-type-specific results, not averages across confounded types.","section":"5.1, 5.2, Table 3"},{"comment":"Observation 1 selectively reports context-memory results. The inter-context open-ended rows in Table 3 show the opposite of the claim for two open-source models: Qwen2.5-VL-7B has OAR=0.38 vs CAR=0.50, and InternVL3-8B has OAR=0.30 vs CAR=0.54. Since the abstract and conclusion state generally that LMMs favor internal parametric knowledge over external evidence, the paper should either qualify the claim to the context-memory setting only or present all settings transparently and explain the discrepancy.","section":"5.2, Table 3"},{"comment":"The statement 'entity recognition conflicts yield an OAR as low as 0.26 on Qwen2.5-VL-7B' is inconsistent with Table 3, where Qwen2.5-VL-7B ER OAR is 0.66 (context-memory) and 0.65 (inter-context); the 0.26 value corresponds to entity knowledge OAR. This appears to be a factual error in the text and undermines the reliability of the qualitative observations.","section":"5.2, Observation 2"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any of the OAR/CAR comparisons. Several differences are small (e.g., Table 3 Qwen2.5-VL-7B context-memory avg OAR=0.44 vs CAR=0.40), and some categories contain as few as 147 instances (visual semantic). The paper should report variance across repeated runs (if any) or statistical significance to support the claimed trends.","section":"5.2, overall"}],"minor_comments":[{"comment":"In Section 4.1, 'from the MMKC-Bench dataset' should read 'from the MMKE-Bench dataset' (reference [37]); the sentence also says 'Since these types are already included in MMKC-Bench,' which is confusing because MMKC-Bench is the current benchmark being built.","section":"4.1"},{"comment":"The row label 'Visual Entity Conflict' is inconsistent with the term 'Entity Recognition Conflict' used throughout the text and Figure 1; please align the terminology.","section":"Table 1"},{"comment":"The metric is defined as 'Irrelevant Answer Ratio' with acronym IAO, but Tables 2 and 3 and the surrounding text use IAR; please use one acronym consistently.","section":"5.1"},{"comment":"In the case study, 'curvewaringsign' appears to be a typo; it should likely be 'curve warning sign'.","section":"5.4"},{"comment":"The text mentions sizes '13B' and '70B' for Qwen2.5-VL, but the paper's model list says Qwen2.5-VL (3B, 7B, 32B, 72B); please correct the model sizes.","section":"5.2, Observation 4"},{"comment":"In the conclusion, 'real-world multimodal knowledge conflict benchmarks, such as [10]' cites WikiContradict, which is a text-only benchmark; please either use a multimodal real-world benchmark or clarify that [10] is text-only.","section":"6"}],"recommendation":"major_revision","confidential_remarks":"The benchmark construction is the paper's main strength, but the headline finding is load-bearing and currently unsupported. The confound in OAR is not a stylistic issue; it changes the conclusion on the cleanest subset of the data. I would suggest the editor require the authors to redo the behavior analysis with a proper parametric baseline and report per-conflict-type results, and to soften the abstract and conclusion if the new results do not support the original claim. The paper also needs basic statistical rigor (error bars or significance tests). With those revisions, the manuscript could be publishable as a benchmark contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this paper builds a genuinely useful multimodal conflict benchmark, but its central empirical claim is undermined by a measurement confound. On the clean entity-knowledge subset, the numbers actually point the other way.\n\nWhat's new and good: MMKC-Bench covers context-memory and inter-context settings across three conflict types (entity recognition, entity knowledge, visual semantic), includes both behavior and conflict-detection tasks, and is human-verified. The construction pipeline is reasonable and the detection task fills a real gap. The authors clearly put effort into instance filtering and quality control, and the dataset is diverse in scope.\n\nThe soft spots are real. The OAR baseline is defined as consistency with the model's answer in the non-conflict setting, where the correct image and correct text are both present. For entity recognition and visual semantic conflicts, the image itself determines the answer—an image of the Empire State Building or an OK gesture carries the information visually—and the same image appears in the conflict setting. So high OAR for those types just means the model is following visual evidence, not that it prefers parametric memory. That conflation contaminates the averaged results.\n\nThe clean test is the entity-knowledge subset, where the image does not contain the queried fact (e.g., birth year). There, Table 3 (open-ended, context-memory) shows CAR exceeds OAR for all three models: Qwen2.5-VL-7B 0.62 vs 0.26, InternVL3-8B 0.53 vs 0.27, GPT-4o mini 0.47 vs 0.36. That is the opposite of the paper's observation 1. Table 2 (MCQ) shows no clear OAR > CAR on entity knowledge either. So the load-bearing claim is unsupported, and the paper's own data reverse it on the subset that actually isolates parametric knowledge from image evidence.\n\nOther issues are minor by comparison: Section 4.1 says the visual semantic knowledge comes from MMKC-Bench when it must mean MMKE-Bench; there are no error bars or significance tests; and the dataset is not released with the paper, which limits immediate reuse.\n\nThis is a paper for the multimodal RAG community. As a dataset contribution it deserves a serious referee, but as evidence about parametric-memory behavior it needs substantial rework. I would engage with it as a benchmark proposal, not as a finding about LMM internals.\n\nRecommendation: send it to peer review, but with the expectation that the authors re-analyze using the entity-knowledge subset as the primary test, and either separate the perception-based conflicts or change the baseline so that following visual evidence is not conflated with parametric memory.","headline":"A useful benchmark resource, but the headline claim that LMMs favor parametric knowledge is contradicted by the paper's own clean entity-knowledge numbers.","tokens_in":14897,"tokens_out":2707,"would_cite":false,"duration_ms":23023,"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 detect knowledge conflicts but usually answer from memory, a new benchmark shows.","keywords":["multimodal knowledge conflict","large multimodal models","retrieval-augmented generation","context-memory conflict","inter-context conflict","conflict detection","counterfactual editing","benchmark construction"],"falsifier":"Run the context-memory conflict task again with the image removed from the conflicting evidence, or replaced by a different image, and compare the Original Answer Ratio; if it drops substantially, the apparent preference for internal knowledge is actually a preference for the unchanged visual input.","tokens_in":13788,"feed_emoji":"🧠","tokens_out":4311,"duration_ms":38883,"temperature":0.7,"pith_summary":"This paper introduces MMKC-Bench, a human-verified benchmark of 1,573 multimodal knowledge conflicts, and uses it to ask how large multimodal models behave when external evidence contradicts what they already know. The paper claims that current models can identify when a conflict exists, but when answering they tend to reproduce their internal parametric knowledge rather than follow the external evidence, a pattern that holds across open-source and closed-source models. The authors care because retrieval-augmented generation depends on models using retrieved evidence; if models quietly override that evidence with memory, RAG systems become unreliable. The benchmark covers three conflict types (entity recognition, entity knowledge, and visual semantics) under both context-memory and inter-context settings.","feed_headline":"Multimodal models pick memory over evidence in knowledge conflicts","feed_subtitle":"New 1,573-instance benchmark shows vision-language models detect conflicts yet mostly answer from internal knowledge.","key_machinery":"The load-bearing object is MMKC-Bench itself: a counterfactually edited dataset in which each original knowledge item (image plus text) is paired with one or two conflicting versions, produced by replacing entity names, attribute values, or semantic labels. Each conflict is cast as a multiple-choice or open-ended question whose answer options include the original, the conflicting, and a distractor answer. Behavior is summarized by three ratios (Original Answer Ratio, Counter Answer Ratio, and Irrelevant Answer Ratio) computed against a non-conflict baseline, and conflict detection is scored as binary accuracy in coarse-grained and fine-grained settings.","core_discovery":"On the paper's own terms, the central discovery is an asymmetry: multimodal knowledge conflicts are recognized but not followed. Across nine models in the Qwen2.5-VL, InternVL3, and GPT-4o mini families, conflict-detection accuracy averages roughly 75 to 79 percent, while in behavior analysis the Original Answer Ratio (answers matching the model's non-conflict response) exceeds the Counter Answer Ratio (answers matching the conflicting external evidence) in nearly every configuration. The paper interprets this as models favoring internal parametric knowledge over external evidence, and notes that this contrasts with text-only LLMs, which generally yield to external knowledge. It also finds that sensitivity is higher for knowledge-level conflicts (such as birth-year contradictions) than for recognition-level conflicts (such as which building is shown), and that larger models show an even stronger preference for internal knowledge.","pith_inferences":["A testable extension would remove or change the image in the conflict setting; because the baseline uses the correct image and the conflict setting keeps it, the current Original Answer Ratio may partly measure image-following rather than parametric memory.","The synthetic counterfactual design, acknowledged by the authors, means real-world conflicts such as naturally contradictory web pages could behave differently, so building a naturalistic multimodal conflict set would test whether the memory preference persists.","If the memory-over-evidence pattern holds, multimodal RAG training should include interleaved multi-source contexts rather than isolated image-text pairs, directly addressing the training-data gap the paper proposes as an explanation.","The finding that models detect conflicts better in non-conflict scenarios than in conflict scenarios may reflect a bias to answer 'yes' only when evidence is emphatic; per-type error analysis could confirm this."],"forward_implications":["In multimodal RAG pipelines, retrieved evidence may not shift model answers as intended, so systems should verify whether the model actually used the retrieved context.","Model scaling within a family increases agreement with internal knowledge, so bigger is not automatically better for evidence-following.","Knowledge-level conflicts such as birth years are more likely to make models switch answers than recognition-level conflicts such as which entity is in the image, suggesting that perception-style conflicts need different mitigation.","The benchmark provides a reusable protocol: counterfactual editing plus human verification, with behavior analysis and conflict detection as two separable evaluation tasks.","Average detection accuracy around 75 to 79 percent means current LMMs can flag conflicts but still need better mechanisms for resolving them."],"supporting_citations":[{"why":"Supplies the three-way taxonomy of intra-memory, context-memory, and inter-context conflicts that organizes the benchmark's design.","marker":"[8]"},{"why":"ConflictBank is the textual benchmark whose finding that LLMs favor external knowledge is the baseline the paper contrasts against.","marker":"[11]"},{"why":"WikiContradict provides real-world text conflicts that the paper cites as the kind of future naturalistic benchmark needed for multimodal settings.","marker":"[10]"},{"why":"Contributes the fine-grained conflict-detection protocol (single-sentence evidence) and the behavior-analysis style adapted for this benchmark.","marker":"[9]"},{"why":"Prior multimodal context-memory conflict work using commonsense conflicts that MMKC-Bench extends to factual knowledge.","marker":"[12]"},{"why":"Prior work distinguishing perception versus cognition conflicts in document understanding, used to frame the recognition-versus-knowledge sensitivity results.","marker":"[13]"},{"why":"Prior cross-modality conflict work in intra-memory settings that the paper positions itself against when claiming coverage of context-memory and inter-context conflicts.","marker":"[14]"},{"why":"MMKE-Bench is the source of visual semantic knowledge instances reused for gesture, action, emotion, and symbol conflicts.","marker":"[37]"}],"fun_headline_variants":["LMMs spot conflicts but side with internal knowledge","New benchmark: multimodal models favor memory over evidence","Detect yes, but LMMs answer from parametric memory","Multimodal models see conflict, still trust internal facts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's main behavioral conclusion assumes that a model's answer when given the correct image and correct text is a pure measure of what it remembers; because the same correct image is also present when the conflicting text arrives, a model that simply follows the image would be miscounted as following memory.","fun_headline_variants_meta":{"raw":{"variants":["LMMs spot conflicts but side with internal knowledge","New benchmark: multimodal models favor memory over evidence","Detect yes, but LMMs answer from parametric memory","Multimodal models see conflict, still trust internal facts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000173,"raw_usage":{"total_tokens":1281,"prompt_tokens":953,"completion_tokens":328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":569,"tokens_out":328,"duration_ms":4347,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:12:49.817457+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the context-memory conflict task again with the image removed from the conflicting evidence, or replaced by a different image, and compare the Original Answer Ratio; if it drops substantially, the apparent preference for internal knowledge is actually a preference for the unchanged visual input.","supporting_citations":[{"cited_title":"Knowledge conflicts for llms: A survey.EMNLP, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the three-way taxonomy of intra-memory, context-memory, and inter-context conflicts that organizes the benchmark's design."},{"cited_title":"Conflictbank: A benchmark for evaluating the influence of knowledge conflicts in llms.Advances in Neural Information Processing Systems, 37:103242–103268, 2024","cited_arxiv_id":null,"evidence_quote":"ConflictBank is the textual benchmark whose finding that LLMs favor external knowledge is the baseline the paper contrasts against."},{"cited_title":"Wikicontradict: A benchmark for evaluating llms on real-world knowledge conflicts from wikipedia.Advances in Neural Information Processing Systems, 37:109701–109747, 2024","cited_arxiv_id":null,"evidence_quote":"WikiContradict provides real-world text conflicts that the paper cites as the kind of future naturalistic benchmark needed for multimodal settings."},{"cited_title":"Resolving knowledge conflicts in large language models.COLM, 2024","cited_arxiv_id":null,"evidence_quote":"Contributes the fine-grained conflict-detection protocol (single-sentence evidence) and the behavior-analysis style adapted for this benchmark."},{"cited_title":"Is cognition consistent with perception? assessing and mitigating multimodal knowledge conflicts in document understanding.arXiv preprint arXiv:2411.07722, 2024","cited_arxiv_id":null,"evidence_quote":"Prior work distinguishing perception versus cognition conflicts in document understanding, used to frame the recognition-versus-knowledge sensitivity results."},{"cited_title":"Mmke- bench: A multimodal editing benchmark for diverse visual knowledge.ICLR, 2025","cited_arxiv_id":null,"evidence_quote":"MMKE-Bench is the source of visual semantic knowledge instances reused for gesture, action, emotion, and symbol conflicts."}],"review_version":1}