{"id":"de64397d-9df0-43a2-8eda-7c8b7f514cbb","arxiv_id":"2505.01958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The paper finds the LLM is mostly faithful given good captions, the CLIP vision encoder contributes perception errors, and the projector preserves visual information but aligns it poorly with text.","lead":"This paper analyzes where visual object hallucination comes from in LLaVA-style vision-language models, testing the language model, vision encoder, and projector separately. It also introduces two new hallucination benchmarks and several training tweaks that produce small, inconsistent gains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'LLM is faithful' finding is not established: the text-substitution probe in §2.2 uses incomplete VG scene graphs, and Table 1's own relation split contradicts it.","rationale":"The reader's weakest assumption is exactly the point on which the whole diagnosis depends. The three findings are presented as a chain: if the LLM is faithful given complete textual scene content, then any remaining hallucination in the full LVLM must come from perception or alignment. Testing that requires a complete and unbiased textual surrogate. The QA-VisualGenome relation split in Table 1 is direct evidence that the surrogate is not equivalent at 7B, because text-only Vicuna-7B_LLaVA is 4.4 points worse than image-based LLaVA-7B. This is not a small discrepancy, and the reversal at 13B makes the claim scale-dependent. The same concern applies to POPE, where the text input lists exactly the objects that the question asks about, making the high accuracy partly tautological; it does not demonstrate what the LLM would do when the textual description is noisy or missing an object. Thus the strongest claim is conditional at best. The projector-preservation evidence also has a reporting gap: Table 3 shows only LLaVA-13B while the text claims both models, and the mitigation tables show only marginal, sometimes negative, improvements, so the causal story is weak. The paper is a useful systematic analysis and the benchmarks are valuable, but the central attribution is not proven. A synthetic oracle-scene-graph test would settle whether the LLM is truly faithful under complete text input, and should be run before the diagnosis is accepted as stated.","tokens_in":16241,"tokens_out":5019,"duration_ms":50104,"concrete_test":"Build or render a synthetic scene set (CLEVR-style, or another domain with an oracle scene graph) with fully known object, attribute, and relation ground truth. For QA-VisualGenome-style yes/no questions, run the §2.2 probe by feeding the complete oracle scene graph as text to Vicuna-7B_LLaVA, and compare to image-based LLaVA-7B on the same questions. If text-only relation accuracy is not at least comparable to image-based accuracy, the LLM is not exonerated for relation hallucination; if it is, the original low relation split is attributable to incomplete VG annotations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central diagnosis rests on Conjecture 1 (§2.2), which replaces the image with a text list of objects, attributes, and relations from Visual Genome and concludes that the LLM is faithful, leaving the vision encoder and projector as the hallucination sources. This is load-bearing because the other two findings inherit their interpretation from it. The inference requires the textual scene graph to be an unbiased, complete proxy for the image, but the paper provides no completeness check and Visual Genome annotations are known to be sparse and noisy. The paper's own data undercut the inference: on QA-VisualGenome relation, Vicuna-7B_LLaVA with text-only input scores 63.2 accuracy versus 67.57 for image-based LLaVA-7B (Table 1), so the LLM is less faithful with supposedly complete text than with the image at the 7B scale. At 13B the ordering reverses, so the conclusion is not stable across scales. Either the text surrogate is incomplete or biased, invalidating the isolation, or the LLM genuinely contributes to relation hallucination; both possibilities contradict the strong 'LLM is faithful' claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a component-level analysis of hallucination in LLaVA-like large vision-language models. It claims three findings: (1) the LLM generates faithful content when image information is supplied as text, (2) the CLIP vision backbone produces perception-level hallucinations, and (3) the projector preserves visual information but fails to align visual and textual spaces. Based on these diagnoses, it proposes targeted mitigations: fine-grained contrastive tuning of CLIP, fine-grained perception-based visual instruction tuning, and three contrastive projector-alignment objectives. It also introduces two benchmarks, QA-VisualGenome for attribute/relation hallucinations and QA-FB15k for cognition-based hallucinations, and evaluates the proposed methods on POPE, POPE-NoCaps, QA-VisualGenome, QA-FB15k, Amber, and LLaVA-Bench.","tokens_in":16504,"tokens_out":5155,"duration_ms":49718,"significance":"If the component-level diagnosis is correct, the paper makes a useful contribution by shifting attention from decoding-side or LLM-side hallucination fixes to perception and alignment problems, and it provides concrete training objectives for those components. The two new benchmarks also fill real gaps in attribute/relation and cognition-based hallucination evaluation. The paper's strengths are its breadth: it probes each component, proposes several mitigation variants, and evaluates on multiple benchmark suites. However, the central 'LLM is faithful' result is not established because the text-substitution probe relies on incomplete Visual Genome scene graphs, and the paper's own Table 1 relation split contradicts the claim at one model scale. The projector-alignment claim also lacks direct causal evidence, and the reported mitigation gains are small and unreplicated. The contribution is promising but needs reanalysis and additional experiments before the central claims can be accepted.","major_comments":[{"comment":"The claim that the LLM is faithful when visual information is replaced by text is not established by the reported experiments. For QA-VisualGenome, the text input is a Visual Genome scene-graph listing, but no completeness check against the image is provided, and Visual Genome annotations are known to be sparse. Table 1 shows that text-only Vicuna-7B_LLaVA scores 63.2 on the relation split while image-based LLaVA-7B scores 67.57, whereas at the 13B scale the ordering reverses (84.30 versus 56.40). Either the text surrogate is an incomplete or biased proxy, which invalidates the isolation of the LLM, or the LLM itself contributes to relational hallucination at the 7B scale; both possibilities contradict the strong 'LLM is faithful' conclusion, which is load-bearing for the paper's other findings.","section":"§2.2, Conjecture 1; Table 1"},{"comment":"The projector information-preservation claim is not fully supported by Table 3. The table reports only LLaVA-13B, yet the text says that 'the performance percentage drop of post-projection features is less than 2%, indicating that the visual features are well preserved by the projectors in both models.' No 7B probing results are shown even though LLaVA-7B is the model used in all later mitigation experiments, and no variance or repeated-seed statistics are reported. The claim that the projector is not a source of information loss is therefore currently supported for only one model scale.","section":"§2.4, Conjecture 3; Table 3"},{"comment":"The conclusion that low cosine similarity between projected image features and caption embeddings causes hallucination is a functional claim that the paper does not test directly. Cosine similarity in an anisotropic embedding space can be low even when features are perfectly usable for next-token prediction, so the appropriate control is to compare against random caption embeddings or to show that increasing this similarity reduces hallucination. The mitigation experiments in Table 6 do not provide that evidence: contrastive alignment variants improve POPE accuracy by at most about 1.1 points, and on QA-VisualGenome they often decrease attribute accuracy (e.g., from 64.67 to 60.95 in the first Int. Align. row). Section 4.2 itself concedes that object hallucination may not be directly related to alignment, which undercuts the causal interpretation of Table 4.","section":"§2.4, Table 4; §4.2, Table 6"},{"comment":"The claims that w-ECLIP and w-FineIns reduce hallucination are based on very small accuracy and F1 differences with no error bars or significance testing. For example, on POPE Random, LLaVA-7B has Acc 87.42/F1 86.36 and w-ECLIP has Acc 87.80/F1 86.87; on QA-VisualGenome Relation, w-ECLIP declines from 67.57 to 67.00. Without multiple seeds or confidence intervals, these differences are within plausible run-to-run variation, so the mitigation claim is not yet demonstrated.","section":"§4.1, Table 5"}],"minor_comments":[{"comment":"Several entries concatenate two numbers without a space (e.g., '79.5080.79', '100.00100.00', '99.6799.67', '87.9089.15'); please fix the formatting for readability.","section":"Table 1"},{"comment":"Equation (7) writes L_itg = -p(R|I,T); as written this is not a loss. It should presumably be -log p(R|I,T).","section":"§3.2.1, Eq. (7)"},{"comment":"The phrase 'the LLM after the pertaining and instruction tuning' should read 'after the pre-training and instruction tuning.'","section":"§2.2"},{"comment":"The notation H_V(Y) is used in the V-information definitions but never defined; please define it as the minimal V-entropy of Y without conditioning, or clarify the notation.","section":"§2.4"},{"comment":"The caption says caption embeddings are processed by the Vicuna tokenizer, but the paper does not specify how the caption embedding is computed (e.g., mean-pooled token embeddings, last hidden state) or what 'Image Res. (575, 488)' denotes; please clarify.","section":"Table 4"},{"comment":"The two 'Int. Align.' rows are distinguished only by an unlabeled symbol; please add a readable key such as 'trainable λ' and 'fixed λ' that is consistent with Section 3.2.1.","section":"Table 6"},{"comment":"The text says w-ECLIP is 'comparable or even better' than w-FineIns, but Table 5 shows mixed ordering across columns; please state the comparison more precisely.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and the component-level framing is potentially valuable. The main issue is internal: the load-bearing 'LLM is faithful' finding is contradicted by the relation split in Table 1 at the 7B scale, and the text-substitution probe lacks a completeness check. I believe this is fixable by replacing the QA-VisualGenome text surrogate with complete captions or full dense scene graphs, adding 7B projector probing, and reporting repeated-seed statistics for the small mitigation gains. Please also ask the authors to document the construction of QA-FB15k in more detail, as it is a claimed contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take on arXiv:2505.01958. The paper does a systematic component-level analysis of LLaVA-like models to localize object hallucination, and it contributes two new benchmarks and a set of measurement numbers that aren't elsewhere. That's real value. But the headline conclusion—that the LLM is faithful and hallucination lives in the vision encoder and projector—is not actually supported by their own data.\n\nWhat's good: QA-VisualGenome extends evaluation to attributes and relations, QA-FB15k covers cognition-based hallucination. The direct CLIP text-image matching test (Table 2) and the cosine similarity measurements (Table 4) are concrete, and the finding that CLIP itself fails on such matches is consistent with prior work. The probing experiment on the projector is a reasonable way to check information preservation, and the numbers suggest the projector retains most of the classification-relevant information. The mitigation methods are sensible even if gains are modest.\n\nNow the soft spots. The load-bearing step is Conjecture 1: they replace the image with a text list of objects, attributes, and relations from Visual Genome, then watch the LLM answer accurately, and conclude the LLM is not the source of hallucination. That inference depends on the text being a complete, unbiased proxy for the image. Visual Genome annotations are notoriously sparse and noisy, and the paper gives no completeness check. More damning, Table 1 shows that on QA-VisualGenome relation, the text-only Vicuna-7B_LLaVA scores 63.2 versus 67.57 for image-based LLaVA-7B. If the text is complete, the LLM is less faithful with it than with the image; if the text is incomplete, the isolation fails. Either way, the strong 'LLM is faithful' claim doesn't hold. At 13B the ordering reverses, so it's not stable across scale.\n\nThere are also presentation issues. Table 3 says 'both models' but only reports LLaVA-13B. There are no error bars anywhere, so marginal differences in Tables 5 and 6 are hard to interpret. The QA-VisualGenome benchmark is built from Visual Genome and the fine-grained instruction data is also from Visual Genome, so gains there may partly reflect train/test overlap.\n\nOverall, the paper is worth engaging with because the benchmarks and measurements are useful and the component framing is clean. But the central diagnosis needs to be reworded: the evidence supports 'CLIP perception is imperfect and the projector doesn't align visual and textual spaces well', not 'the LLM is faithful'. Send it to peer review—a serious referee can push for the necessary tempering and for the missing error bars.","headline":"Systematic component analysis and two new benchmarks are real contributions, but the headline 'LLM is faithful' claim is contradicted by the paper's own relation results.","tokens_in":17061,"tokens_out":2800,"would_cite":false,"duration_ms":25183,"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":"Visual hallucination in LLaVA-like models comes mainly from the vision encoder and the projector, not the language model.","keywords":["visual object hallucination","large vision-language models","LLaVA","vision encoder","projector alignment","hallucination benchmarks","contrastive alignment","cognition-based hallucination"],"falsifier":"Run the text-substitution experiment with the full ground-truth Visual Genome annotation for each image rather than a filtered subset: if LLaVA's accuracy on QA-VisualGenome remains far below the reported 63–68% attribute and relation accuracy even when exhaustive visual facts are supplied as text, the claim that the language model is faithful would be falsified.","tokens_in":1639,"feed_emoji":"🖼️","tokens_out":4463,"duration_ms":80248,"temperature":0.7,"pith_summary":"Large vision-language models such as LLaVA hallucinate objects, attributes, and relations when answering questions about images. This paper tries to locate where those hallucinations originate by taking the model apart into its three components — the language model, the CLIP vision encoder, and the projector that connects them — and testing each one in isolation. It concludes that the language model is largely faithful when it receives the image content as text, that the vision encoder itself misperceives objects, and that the projector preserves visual information but aligns it poorly with the language model's embedding space. If this diagnosis is right, hallucination is primarily a perception and cross-modal alignment problem, and the field's emphasis on decoding strategies and language-model corrections addresses a downstream symptom rather than the root cause.","feed_headline":"Hallucination traced to vision encoder and projector, not the LLM","feed_subtitle":"When captions replace images, LLaVA stays faithful; fixes should target perception and alignment.","key_machinery":"The load-bearing mechanism is component-wise decomposition of a LLaVA-like model, with each component probed by a dedicated diagnostic. The language decoder is isolated by text substitution: replace the image with ground-truth objects, attributes, and relations, then measure question-answering accuracy. The vision encoder is isolated by CLIP text-image matching on yes/no templates. The projector is probed twice: V-information-based linear probing (comparing pre- and post-projection features on classification) to test information preservation, and cosine similarity between projected image features and caption embeddings to test visual-textual alignment. The proposed mitigations are CLIP post-training with synthetic hard negatives, fine-grained perception-based instruction tuning, and a contrastive alignment loss for the projector with three instantiations.","core_discovery":"The paper's central claim is a three-part localization of visual object hallucination in LLaVA-like models. First, when a model is given a complete textual description of an image — objects for POPE, objects plus attributes and relations for QA-VisualGenome — the underlying LLM answers almost perfectly (up to 99.67% on POPE random), so the language model is not the hallucination source. Second, CLIP, acting as the vision encoder, makes substantial errors on the simple text-image matching task \"There is a {object} in the image,\" scoring only 83.33% on POPE random, which indicates perception-level hallucination. Third, linear probing shows the projector retains essentially all visual information (post-projection accuracy drops less than 2% on CIFAR and ImageNet), but the cosine similarity between projected image features and the embeddings of their own captions is only 0.03–0.06, indicating a failure to align the visual and textual spaces. The paper then proposes component-specific fixes and shows they reduce hallucination on perception-based benchmarks and, for the contrastive alignment variants, on a new cognition-based benchmark.","pith_inferences":["If the vision-encoder locus holds, hallucination rates should scale with CLIP's per-category matching error; a testable prediction is that models with improved CLIP fine-grained discrimination will hallucinate less on exactly the object categories where CLIP currently confuses present and absent objects.","The low cosine similarity between projected visual features and caption embeddings may mean that projection is not performing alignment in a simple cosine-similarity geometry; the authors do not explore whether the autoregressive loss itself is doing the alignment implicitly, in which case cosine similarity is the wrong diagnostic.","The text-substitution experiment could be reused as a general audit tool: for any new vision-language model, the gap between image-input accuracy and caption-input accuracy localizes hallucination to perception versus generation without needing component-level retraining.","The QA-FB15k results hint at a connection between modality alignment and factual reasoning that the paper does not develop: aligning the projector better may unlock knowledge already present in the language model, which could be probed directly by measuring knowledge-query accuracy under varying alignment strengths."],"forward_implications":["Hallucination-mitigation efforts in LLaVA-like models should prioritize the vision encoder and the projector; decoding-time or LLM-only fixes address a secondary symptom.","Fine-grained CLIP post-training or fine-grained instruction tuning yields measurable gains on POPE, POPE-NoCaps, and QA-VisualGenome without adding parameters at inference time.","Adding a contrastive alignment stage for the projector (about 12 minutes of training in the paper's setup) improves a cognition-based hallucination benchmark, suggesting that better alignment helps the model draw on world knowledge stored in the LLM.","Attribute and relation hallucinations are harder to cure by projector alignment alone, implying that better visual representations are the bottleneck for fine-grained perception.","The new QA-VisualGenome and QA-FB15k benchmarks extend hallucination evaluation beyond object existence to attributes, relations, and knowledge-dependent cognition."],"supporting_citations":[{"why":"Defines the POPE benchmark and its random, popular, and adversarial sampling settings, which are used throughout the paper to evaluate object-existence hallucination in the LLM, CLIP, and mitigation experiments.","marker":"(Li et al., 2023e)"},{"why":"Supplies the dense object, attribute, and relation annotations from which QA-VisualGenome is built, enabling the paper's fine-grained hallucination evaluation.","marker":"(Krishna et al., 2017)"},{"why":"Provides the FB15k multimodal knowledge graph used to construct the cognition-based QA-FB15k benchmark.","marker":"(Bordes et al., 2013)"},{"why":"Defines V-information, the formal basis for testing whether the projector preserves visual information by comparing pre- and post-projection features.","marker":"(Hewitt et al., 2021)"},{"why":"Introduces the LLaVA architecture and visual instruction tuning pipeline whose components the paper analyzes and modifies.","marker":"(Liu et al., 2023e)"},{"why":"Documents object hallucination in CLIP models and motivates the fine-grained negative-sample generation used for CLIP post-training.","marker":"(Liu et al., 2024b)"}],"fun_headline_variants":["LLM exonerated: image hallucination stems from vision encoder","Vision encoder, not LLM, causes object hallucination","Text captions make LLaVA faithful — vision alignment is the fix","Hallucination source: vision encoder and projector, not the language model","Perception, not reasoning: where LVLMs see things that aren't there"],"cache_read_input_tokens":19200,"weakest_assumption_plain":"The argument that the LLM is faithful rests on the assumption that a textual list of objects, attributes, and relations is a complete substitute for the image; if that list omits or misrepresents visual content, the low error with text input does not prove that the LLM is not a hallucination source.","fun_headline_variants_meta":{"raw":{"variants":["LLM exonerated: image hallucination stems from vision encoder","Vision encoder, not LLM, causes object hallucination","Text captions make LLaVA faithful — vision alignment is the fix","Hallucination source: vision encoder and projector, not the language model","Perception, not reasoning: where LVLMs see things that aren't there"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1251,"prompt_tokens":923,"completion_tokens":328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":234}},"tokens_in":539,"tokens_out":328,"duration_ms":3720,"temperature":1.0,"reasoning_tokens":234,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:04:46.034744+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the text-substitution experiment with the full ground-truth Visual Genome annotation for each image rather than a filtered subset: if LLaVA's accuracy on QA-VisualGenome remains far below the reported 63–68% attribute and relation accuracy even when exhaustive visual facts are supplied as text, the claim that the language model is faithful would be falsified.","supporting_citations":[],"review_version":1}