{"id":"110395e2-2c83-412c-99dc-52af9acce5e3","arxiv_id":"2411.13697","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Decomposing MLLM responses into atomic verification tasks and checking them with an ensemble of open-source expert models yields preference data that reduces hallucination in LLaVA and Qwen-VL-Chat.","lead":"This paper introduces DecompGen, a system that judges an AI image-description response by breaking it into small facts and checking each one with a specialized open-source tool, then uses those judgments to build a training set that makes image chatbots hallucinate less. A reader might care because it offers a cheaper, open alternative to using GPT-4 as a judge for making multimodal AI more trustworthy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DGPref preference labels inherit unvalidated noise from decomposition and expert errors, making the feedback-quality premise central but unverified.","rationale":"The reader's weakest_assumption identifies the same point, and I agree it is the most load-bearing. The main text claims DecompGen provides 'precise assessments' (Sec. 3.4), but the evidence for precision is a 50-case manual check with one expert at 50% accuracy. The ablations show expert quality matters (PrefObj-0.1 vs PrefObj, Table 3), which supports the mechanism but does not quantify the noise level or its effect on pairwise labels. The missing test is a direct human evaluation of DGPref's preference labels, not just of expert accuracies in isolation or of final model outputs. Until that is done, the central claim should remain conditional. I do not think the paper is fatally flawed: the cross-benchmark gains and the expert-quality ablation provide independent support. But the load-bearing premise—accurate fine-grained feedback—is not directly validated, so the reader's CONDITIONAL verdict is appropriate and unchanged by this stress-test.","tokens_in":28276,"tokens_out":9169,"duration_ms":86513,"concrete_test":"Sample ~150 preference pairs from DGPref across all five aspects, with two independent annotators per pair, and measure (a) agreement between DGPref's chosen preferred response and human majority (chance = 50%), and (b) per-claim precision/recall of DecompGen's 0/-1 scores against human judgments on a subset of extracted check-worthy parts, especially OCR, general attribute, and relation. Additionally, rerun DPO after removing all pairs whose score difference is driven entirely by OCR or decomposition-flagged 'invalid' parts; if the CHAIR/MMHal gains persist, the noisy components are not load-bearing, and if they vanish, the premise is confirmed as critical and currently unsubstantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DecompGen's fine-grained scores are the sole basis for assigning preferred/rejected labels in DGPref (Sec. 2.3). The paper's own quality checks leave this premise in doubt: Appendix E Table 6 reports OCR expert accuracy of 50.0%, general attribute 76.2%, and general relation 79.4% on only 50 manual cases, and Sec. 3.4 finds invalid check-worthy part extraction to be the most common decomposition error (77% of errors). An invalid part such as 'other' as an object is usually scored -1 because no bounding box exists, penalizing faithful responses; a missed part can let a hallucination go unscored. Yet no direct validation is reported for the resulting pairwise preference labels—e.g., agreement with human judgments on which response in a DGPref pair is more accurate. If per-part scores are this noisy, many DPO training pairs are mislabeled, and the observed hallucination drops could stem from length/conservatism artifacts or regularization rather than from accurate feedback. This is the load-bearing premise behind the central claim that DecompGen feedback is what improves trustworthiness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DecompGen, a decomposable framework that evaluates MLLM image-description responses by first decomposing each response into visually check-worthy parts (objects, relations, attributes, counts, image text) and then verifying each part through a set of open-source expert models (OWL-ViT, BLIP-2, EasyOCR, heuristic rules). The resulting fine-grained feedback scores are aggregated into an overall score, which is used to construct a preference dataset, DGPref, by pairing higher- and lower-scoring responses. MLLMs (LLaVA-v1.5 and Qwen-VL-Chat at 7B) are then aligned with DGPref via DPO with LoRA. The main empirical claim is that this alignment sharply reduces hallucination rates on ObjHal, MMHal, and AMBER while preserving informativeness, and that DGPref-aligned models outperform or match prior preference-data methods such as Silkie, HSA-DPO, and RLAIF-V.","tokens_in":28586,"tokens_out":5060,"duration_ms":85972,"significance":"If the central claim holds, the paper makes a useful contribution: it replaces closed-source evaluator feedback (e.g., GPT-4) with an ensemble of open-source experts, at a total parameter count of roughly 4B, and it shows large hallucination reductions on two different base MLLMs. The decomposition into atomic verification tasks is interpretable, and the framework allows swapping individual expert models. The paper also releases code, reports a human evaluation, and provides an ablation study with an interesting negative result (PrefObj-GT fails to generalize to MMHal/AMBER). The significance is conditional, however, on the reliability of the automatically generated preference labels and on the independence of the evaluation benchmarks from the training image distribution; both of these points need to be established before the reported gains can be fully credited.","major_comments":[{"comment":"The load-bearing premise—that DecompGen's per-part scores correctly separate faithful from hallucinated response content—is not directly validated. Appendix E Table 6 reports expert accuracies on only 50 manually checked cases, with OCR at 50.0%, general attribute at 76.2%, and general relation at 79.4%. Section 3.4 adds that invalid check-worthy part extraction is the most common decomposition error, accounting for 77% of errors. Because each DGPref preferred/rejected label is a deterministic function of these scores, label noise is inherited from both stages. The paper, however, reports no evaluation of the resulting pairwise preference labels—e.g., no human agreement on which response in a DGPref pair is more accurate, and no analysis of how many pairs would flip if the noisy parts are removed or corrected. I request (a) a human validation of a random sample of DGPref pairs, (b) a per-error-propagation analysis showing the fraction of pairs whose labels change when OCR/general-attribute/general-relation errors are corrected, and (c) an experiment training on a variant of DGPref that excludes or down-weights the least reliable expert dimensions (particularly OCR).","section":"Sec. 2.3, Sec. 3.4, Appendix E Table 6"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any evaluation metric. The key comparisons—DGPrefQwen versus HSA-DPO and RLAIF-V, and DGPref versus its ablations—involve differences of a few points or less on several metrics (e.g., MMHal HalRate 0.34 vs. 0.29 for RLAIF-V, AMBER s.CHAIRi 1.5 vs. 3.1 for RLAIF-V). Given the known variance of hallucination benchmarks and the fact that evaluation of MMHal uses GPT-4 with stochastic sampling, these differences may not be robust. Please report results across at least three seeds or bootstrap intervals for the main models and ablations, and state the number of evaluation samples underlying each metric.","section":"Tables 2, 3, 5"},{"comment":"The training and evaluation image sources may overlap, which could inflate the reported gains. DGPref is constructed from Visual Genome images, while ObjHal is a COCO-based benchmark and AMBER may also draw on common image sources; Visual Genome contains many images that also appear in COCO. The paper does not quantify this overlap. If a substantial fraction of evaluation images appear in the DGPref training set, the CHAIRs/CHAIRi reductions on ObjHal (e.g., from 36.0 to 8.0 for Qwen) would partly reflect in-distribution leakage rather than improved trustworthiness. Please report the exact image overlap between DGPref and each evaluation benchmark, and provide results on a strictly disjoint subset of each benchmark.","section":"Sec. 3.1, Table 2"},{"comment":"The aggregation step that converts per-part scores into an overall response score is not specified. The text says 'Weighted averaging is applied over the fine-grained feedback scores' but does not give the formula, the weights, how skipped verifications (where existence fails) are treated, or how the number of check-worthy parts per response is normalized. This is not a cosmetic omission: the preferred/rejected ordering of every DGPref pair depends on this aggregation, and without it the dataset construction is not reproducible or auditable. Please provide the exact aggregation function, the weight values, and the treatment of skipped parts.","section":"Sec. 2.3"},{"comment":"The human evaluation that supports the claim that trustworthiness improvements do not sacrifice informativeness is based on very low inter-annotator reliability: Cohen's kappa is 0.13 for 'fewer errors', 0.15 for 'more informative', and 0.24 for 'preferable'. With these levels of agreement, the reported 81.7–88.3% preference rates are difficult to interpret. The human evaluation should either use more reliable annotation procedures (e.g., adjudication, more annotators, clearer criteria) or be reported with appropriate caveats, and it should not serve as the primary evidence against the 'conservatism/length artifact' concern without a length-controlled analysis.","section":"Appendix M, Table 4"}],"minor_comments":[{"comment":"There is a duplicated word: 'image content content' in the Introduction.","section":"Abstract / Sec. 1"},{"comment":"Typo: 'Fianally' should be 'Finally'.","section":"Sec. 3.2"},{"comment":"Typo: 'HAS-DPO' should be 'HSA-DPO' in the sentence comparing with the strongest baselines.","section":"Sec. 3.2"},{"comment":"Typo: 'comparbale' should be 'comparable'.","section":"Sec. 3.2 / Table 2 caption"},{"comment":"The figure caption contains 'Hallicination' which should be 'Hallucination'.","section":"Fig. 3"},{"comment":"The dataset name 'DGPref' is said to stand for 'DecompGen Feedback Generated Preferance dataset'—'Preferance' is misspelled; it should be 'Preference'.","section":"Sec. 2.3"},{"comment":"The reference to BLIP-2 in the description of [RELA]general cites [46] (Woodpecker) in one place and [20] (BLIP-2) elsewhere; the reference numbers appear inconsistent. Please verify all expert-model citations.","section":"Appendix A"},{"comment":"Please clarify whether the 52k DGPref samples are all distinct pairs per image and how the '8 diverse instructions' are paired with images; the paper says 'pairwise combinations' but the numbers (16k images, 8 responses) imply 448k possible pairs, so the sampling procedure should be stated explicitly.","section":"Sec. 3.1 / Appendix I"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claim is plausible and the paper is well structured, but the missing validation of the automatically generated preference labels is a substantive gap that affects the interpretation of the main results. The possible image overlap between Visual Genome (training) and the COCO-based evaluation benchmarks is a second concern that the authors should address with explicit overlap statistics. The low inter-annotator agreement in Appendix M and the missing error bars should also be fixed before the paper can be considered for acceptance."},"author_rebuttal":null,"desk_editor":null,"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper builds a preference dataset from decomposed, expert-verified claims and shows it cuts MLLM hallucination rates by about fivefold without losing informativeness.","keywords":["multimodal large language models","hallucination","preference optimization","direct preference optimization","response decomposition","expert models","trustworthiness","open-vocabulary detection"],"falsifier":"Take a random sample of, say, 200 MLLM responses from a different image distribution, have annotators label every extracted claim and the preferred response, and then measure DecompGen's per-claim agreement and pairwise ranking accuracy against those labels; if per-claim agreement falls far below the 76-92% the paper reports on its small check, or if the pairwise preference accuracy is near chance, the method's hallucination reductions should fail to replicate on that distribution.","tokens_in":28038,"feed_emoji":"🧩","tokens_out":10308,"duration_ms":91015,"temperature":0.7,"pith_summary":"Multimodal large language models (MLLMs) often produce detailed image descriptions that contain confident factual errors. The paper argues that the best way to fix this is to stop asking a single evaluation model to judge a whole response, and instead break every response into small, checkable claims, verify each claim with a specialist open-source model, and use those fine-grained verdicts to build a preference dataset (DGPref) for direct preference optimization. Across two 7B MLLMs, alignment with DGPref reduces object-hallucination rates by roughly an order of magnitude on the response-level metric (LLaVA from 54.7 to 10.3 CHAIRs; Qwen-VL-Chat from 36.0 to 8.0) while keeping or improving informativeness scores. If these results hold, preference-data construction for trustworthy multimodal models no longer requires expensive closed-source judges.","feed_headline":"Decomposing AI responses cuts hallucination rates by about fivefold","feed_subtitle":"Aligning LLaVA and Qwen-VL with expert-verified preferences cuts errors without losing informativeness.","key_machinery":"The load-bearing mechanism is DecompGen's two-stage decomposition. In stage one, an instruction-tuned LLM (Llama-3.1-8B-Instruct) uses eight in-context examples per aspect to split a response into check-worthy parts. In stage two, each part is reduced to atomic tasks via a fixed mapping (e.g., relation verification becomes existence checks for both entities plus a relationship check), and each task is routed to a dedicated expert: OWL-ViT for [DET], BLIP-2 for [RELA]general and [ATTR]general, EasyOCR for [OCR], and hand-coded geometric rules for spatial relations and size-aware attributes. Each expert emits a score of 0 (faithful) or -1 (problematic), the scores are weighted and averaged into an overall response score, and the resulting preferences are used for Direct Preference Optimization.","core_discovery":"The central claim is that a decomposable, interpretable feedback generator can outperform a monolithic judge for preference learning. DecompGen first prompts an LLM to extract 'check-worthy parts' from a response across five aspects—object existence, relations, attributes, counts, and image text—and then maps each part to atomic verification tasks ([DET], [RELA], [ATTR], [COUNT], [OCR]). Each atomic task is executed by a specialist: an open-vocabulary detector (OWL-ViT), a VQA model (BLIP-2) for general relations and attributes, coordinate/size heuristics for spatial relations and counts, and an OCR tool for scene text. The binary pass/fail scores for every part are aggregated into a single response score, and pairwise comparisons form DGPref, consisting of 52k preference pairs from 16k Visual Genome images. After DPO with LoRA, LLaVA-v1.5 and Qwen-VL-Chat show large drops in hallucination on ObjHal (CHAIRs 10.3 and 8.0) and lower or comparable error rates than preference-data baselines built with GPT-4 or a 34B evaluator, while preserving coverage and informativeness on MMHal and AMBER.","pith_inferences":["The same decompose-and-verify recipe could be applied to other long-form generation tasks (e.g., text-only summarization or statement verification): any output that can be split into atomic claims and checked by a portfolio of tools could yield preference data without training a monolithic judge. This is an extension the paper does not make.","The paper's own error analysis (OCR expert at 50% accuracy; invalid extraction as 77% of decomposition errors) suggests that the main headroom for further gains lies in better extraction and text verification; if those components improve, DGPref's quality and the resulting hallucination reductions should improve correspondingly—a testable prediction.","The fact that aligning with ground-truth annotations (PrefObj-GT) improved ObjHal but hurt generalization to MMHal/AMBER cautions that expert-based labels, while noisy, can be more diverse and robust than sparse human annotations; this is a nuance the paper surfaces but does not fully resolve.","The reported gains are measured on three specific hallucination benchmarks; treating 'trustworthiness' as a broader property would require testing on additional dimensions such as safety, bias, and out-of-distribution images."],"forward_implications":["Preference data for MLLM alignment can be generated entirely from open-source components, eliminating the need for closed-source judges like GPT-4 in this pipeline.","Each error in a model response becomes traceable to a specific extracted claim and the expert that verified it, making feedback interpretable and easy to audit.","The modular expert design means stronger or additional specialists (e.g., for commonsense or fine-grained OCR) can be swapped in to improve the preference data without redesigning the framework.","Because automatic generation scales, DGPref (52k pairs) can exceed the size of human-annotated preference sets (~1k-10k pairs), which the paper shows correlates with better alignment outcomes."],"supporting_citations":[{"why":"Supplies the Direct Preference Optimization algorithm used to align MLLMs with DGPref preferences.","marker":"[33]"},{"why":"OWL-ViT, the open-vocabulary detector that implements the [DET] atomic tasks for object existence, relations, attributes, and counts.","marker":"[28]"},{"why":"BLIP-2, the VQA model used for [RELA]general and [ATTR]general verification in the expert ensemble.","marker":"[20]"},{"why":"Visual Genome, the image source used to construct the 16k-image DGPref dataset.","marker":"[18]"},{"why":"Silkie, a preference-distillation baseline built with a GPT-4 evaluator that DGPref must match or beat.","marker":"[21]"},{"why":"HSA-DPO, a fine-grained AI-feedback (GPT-4 based) baseline for hallucination mitigation, compared against DGPref-aligned models.","marker":"[44]"},{"why":"RLAIF-V, an open-source AI-feedback baseline using a 34B evaluator model, compared against the 4B-parameter expert ensemble.","marker":"[48]"},{"why":"ObjHal, the benchmark used to measure CHAIRs and CHAIRi hallucination rates, the primary quantitative evidence.","marker":"[34]"},{"why":"MMHal, the benchmark used to evaluate informativeness (Score) and hallucination rate (HalRate).","marker":"[40]"},{"why":"AMBER, the multi-dimensional hallucination benchmark measuring existence, attribute, and relation errors plus coverage.","marker":"[42]"}],"fun_headline_variants":["Decomposing AI responses cuts hallucination rates fivefold","Expert ensemble for MLLMs reduces hallucinations by 5x","Open-source decomposable judge beats GPT-4 for MLLM alignment","Break down AI replies to fix hallucinations without losing info","Atomic verification of AI answers cuts errors fivefold"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's value depends on the LLM's extraction and the expert models' verdicts being reliable enough that the binary scores separate faithful from hallucinated content, and the paper's own manual checks show this foundation is imperfect (OCR 50%, invalid extraction 77% of decomposition errors).","fun_headline_variants_meta":{"raw":{"variants":["Decomposing AI responses cuts hallucination rates fivefold","Expert ensemble for MLLMs reduces hallucinations by 5x","Open-source decomposable judge beats GPT-4 for MLLM alignment","Break down AI replies to fix hallucinations without losing info","Atomic verification of AI answers cuts errors fivefold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000985,"raw_usage":{"total_tokens":4193,"prompt_tokens":975,"completion_tokens":3218,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":3136}},"tokens_in":591,"tokens_out":3218,"duration_ms":23060,"temperature":1.0,"reasoning_tokens":3136,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:58:46.183044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of, say, 200 MLLM responses from a different image distribution, have annotators label every extracted claim and the preferred response, and then measure DecompGen's per-claim agreement and pairwise ranking accuracy against those labels; if per-claim agreement falls far below the 76-92% the paper reports on its small check, or if the pairwise preference accuracy is near chance, the method's hallucination reductions should fail to replicate on that distribution.","supporting_citations":[{"cited_title":"Manning, Stefano Ermon, and Chelsea Finn","cited_arxiv_id":null,"evidence_quote":"Supplies the Direct Preference Optimization algorithm used to align MLLMs with DGPref preferences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"BLIP-2, the VQA model used for [RELA]general and [ATTR]general verification in the expert ensemble."},{"cited_title":"Shamma, Michael S","cited_arxiv_id":null,"evidence_quote":"Visual Genome, the image source used to construct the 16k-image DGPref dataset."},{"cited_title":"Aligning large multimodal models with factually aug- mented RLHF","cited_arxiv_id":null,"evidence_quote":"MMHal, the benchmark used to evaluate informativeness (Score) and hallucination rate (HalRate)."}],"review_version":1}