{"id":"1a2985d1-0dea-4aa0-9e37-b5c4d3810c75","arxiv_id":"2505.07172","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Rationale-augmented instruction tuning plus self-critic DPO reduces visual hallucination and improves multimodal reasoning in LVLMs across several benchmarks.","lead":"This paper presents Re-Critic, a training recipe that adds teacher-written explanations to image-question pairs before the model answers, then uses the model's own judgments to pick better answers for preference tuning. It reports that models trained this way hallucinate less and score higher on several vision-language benchmarks, even when only a small fraction of the training data is augmented.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control with generic rationale; VCIT bundles content and insertion, so the reported gains may not be caused by image-grounded rationales.","rationale":"The reader already issued a CONDITIONAL verdict, and this stress-test does not require moving away from that. The proposed control directly tests whether the image-grounded rationale content is the active ingredient, which is the assumption the reader flagged. The concern is not that the authors are dishonest or that the method is impossible; it is that the experimental design, as reported, cannot distinguish rationale content from generic insertion. This is a load-bearing gap because the paper's headline contribution is specifically the rationale-augmented instruction mechanism. The condition for acceptance should be explicit: run the generic-insertion control and release the augmented dataset and exact training/evaluation protocol. Until then, the empirical claim remains plausible but not fully established.","tokens_in":13645,"tokens_out":7518,"duration_ms":84336,"concrete_test":"Retrain LLaVA-v1.5 with the same 10k augmented subset in three arms: (A) Re-Critic GPT-4o rationales as described; (B) generic rationales that do not reference image content, matched for length and formatting (e.g., templated 'To answer this, inspect the image carefully and reason step by step'); (C) original QA pairs without insertion. Keep all other training details fixed and evaluate on POPE, MMHalBench, HallusionBench, MathVista, and LLaVA-Bench using identical inference prompts. If B is statistically indistinguishable from A, the rationale-specific mechanism is not supported; if A significantly exceeds B, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that rationale-augmented instruction tuning is what drives the reported gains on hallucination and broader reasoning benchmarks. For this to be supported, the content of the inserted rationale must be shown to matter, not just the presence of extra instructive text or a longer prompt. VCIT bundles two changes at once: the training prompt is augmented with GPT-4o-generated rationale text, and the prompt is lengthened with pedagogical framing. There is no control arm in which the same 10k-30k samples are augmented with a generic, image-independent rationale (e.g., 'Read the image carefully and reason step by step before answering') or with rationales generated without access to the image. Table 4 compares VCIT on/off, but that contrast cannot separate the effect of rationale content from the effect of the insertion itself. Moreover, because the teacher is prompted with the gold answer and told to 'consider the image contents,' and because no human audit or filtering of the generated rationales is reported, it is possible that the augmented data teaches a shallow text pattern rather than visual reasoning. If a generic-insertion control matched Re-Critic, the central claim of 'methodological learning before reasoning' would be unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Re-Critic, a two-stage framework for reducing hallucination in large vision-language models (LVLMs). The first stage, VCIT, augments raw instruction-tuning samples with GPT-4o-generated rationales that are inserted into the question text; the second stage builds preference pairs by having the fine-tuned model critique its own two responses and applies DPO. Experiments on LLaVA-v1.5-7B, InternVL2-2B, and MiniGPT-4 report gains on hallucination benchmarks (POPE, MMHalBench, HallusionBench, Object HalBench) and on several general multimodal benchmarks (MME, MathVista, LLaVA-Bench, MMBench, RWQA, GQA, OCRBench). Ablations in Table 4 attribute improvements to both VCIT and the self-critic DPO.","tokens_in":13860,"tokens_out":5061,"duration_ms":46607,"significance":"If the claims hold, Re-Critic is a simple and scalable data-augmentation recipe that improves both hallucination metrics and general reasoning without external reward models or human preference annotation. The paper includes ablations, multiple backbone models, and a small-data regime analysis, and it provides a concrete recipe that other groups could reproduce. However, the significance is currently limited by the absence of a control that isolates the content of the inserted rationale from the mere fact of insertion, by the lack of any statistical uncertainty quantification, and by the self-referential nature of the preference-pair construction.","major_comments":[{"comment":"The central claim that rationale-augmented instruction tuning drives the gains is not supported because the VCIT condition confounds the content of the inserted rationale with the presence of extra instructive text. Table 4 compares VCIT on/off, but an arm that inserts a generic, image-independent instruction (e.g., \"Read the image carefully and reason step by step before answering\") or rationales generated without access to the image is missing. Without such a control, the improvements could be caused by longer prompts or pedagogical framing alone, not by the image-grounded rationale content that the paper emphasizes.","section":"§3.1 / Table 4"},{"comment":"The reported \"average improvement\" is not defined and mixes higher-better and lower-better metrics. For example, in Table 2, HallusionBench drops from 59.4 to 56.0 for LLaVA-v1.5 and from 56.0 to 55.0 for InternVL2, and in Table 3 MME drops by 3.4 points and InternVL2 MMBench drops by 0.1 points, yet the text states consistent average improvements of 6.2%, 2.7%, 2.8%, and 3.7%. The paper must state the direction of each metric, how percentage changes are computed (e.g., relative to the base, with lower-better metrics inverted), and which benchmarks are included in each average; otherwise the headline claim is misleading.","section":"§4.2, Tables 2 and 3"},{"comment":"No error bars, confidence intervals, or multiple seeds are provided. Many reported deltas are small (e.g., POPE +0.6, MMHal-B +0.14 for LLaVA-v1.5, MMBench -0.1 for InternVL2). Without estimates of variance, the paper cannot establish that these differences are statistically meaningful, especially on binary or score-based benchmarks where a single-seed run can differ by this amount.","section":"§4.2 / Tables 2 and 3"},{"comment":"The self-critic mechanism is self-referential: the same model that is being optimized generates the two response candidates and also evaluates them using its own judgment. The paper argues that this avoids distribution shift, but it also introduces a risk that the model simply prefers outputs that match its own style, independent of factual correctness. No validation of the preference pairs is provided (e.g., human agreement, comparison with an external critic, or a random-pair baseline). Table 4 shows that self-critic alone improves LLaVA-Bench by 6.7 points, which is larger than the VCIT-only gain and suggests that generic DPO effects may be at play.","section":"§3.2 / Table 4"},{"comment":"The comparison with prior methods (LLaVA-RLHF, HA-DPO, POVID) relies on numbers taken from previous papers without re-evaluation under a single protocol. Since Object HalBench evaluation typically depends on the prompt set and decoding settings, and MMHal-Bench relies on a GPT-4 judge, the baseline values may not be comparable to the runs of Re-Critic. The paper should either re-run all baselines under the same protocol or clearly state the source and the exact evaluation settings for each reported number.","section":"§4.1 / Baselines"}],"minor_comments":[{"comment":"The section title contains a typo: \"Prefernece\" should be \"Preference\".","section":"§3.2"},{"comment":"The caption reads \"on on four hallucination benchmarks\"; remove the duplicated \"on\".","section":"Table 2"},{"comment":"The label \"T extQA\" should be \"TextQA\".","section":"Figure 4"},{"comment":"The \"Base Model\" row in Table 5 reports OCRBench 802, whereas Table 3 reports 784 for the same InternVL2 base model. This inconsistency should be resolved.","section":"§4.4 / Table 5"},{"comment":"The text claims an average improvement of 2.8% on general benchmarks for LLaVA-v1.5, but Table 3 shows a decrease on MME (-3.4). The definition of the average and the direction of each metric must be clarified before the claim can be assessed.","section":"§4.2 / Table 3"},{"comment":"The hard-sample sampling strategy (BERT-based clustering, top-k selection, difficulty sorting) is described only briefly. The paper should provide details (number of clusters, k, how difficulty accuracy was measured, and whether the selection was done on the training set or a held-out set) so that the experiment is reproducible.","section":"§4.4 / Table 5"},{"comment":"The small-data experiment uses LLaVA-Instruct-80K, but it is unclear whether the base model is LLaVA-v1 or LLaVA-v1.5. State the exact base model and training recipe for this ablation.","section":"§4.3 / Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting and timely problem, and the proposed framework is simple enough to be adopted widely if the claims hold. The main blockage is the absence of a generic-insertion control, which directly bears on the paper's central thesis. This is fixable with a modest additional experiment, so I recommend major revision rather than rejection. The average-improvement reporting also needs to be made transparent before the paper can be considered conclusive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Re-Critic is a cheap, scalable recipe: take a small slice of your instruction data, have GPT-4o write a short 'judgment basis' rationale grounded in the image and gold answer, insert that rationale into the question rather than the answer, and fine-tune. Then use the fine-tuned model's own self-critic to pick preferred vs dispreferred responses for DPO. The result is consistent, modest gains on hallucination benchmarks and some general reasoning benchmarks over LLaVA-v1.5-7B and InternVL2-2B.\n\nThe genuinely new things are the insertion into the question (VCIT) rather than appending to the answer, and the self-critic pair selection with no external reward model. That combination is worth taking seriously. The ablations in Table 4 show both components contribute, and the MiniGPT-4 result gives some generality. The gains on MathVista and LLaVA-Bench suggest the effect is not just on hallucination metrics.\n\nThe main soft spot is the missing control for rationale content. VCIT bundles two changes at once: adding extra instructive text to the prompt, and grounding that text in the image. Without a control that inserts a generic, image-independent rationale ('read the image carefully, reason step by step') into the same subset, you cannot attribute the gains to the image-grounded content. The paper's central claim of 'methodological learning before reasoning' is under-supported without that control. This is a real weakness, though not a fatal one; the evidence is at least consistent with the rationale content mattering.\n\nOther issues: no error bars or multiple seeds, and several baseline numbers are taken from prior papers without re-running, so the deltas are not apples-to-apples. The 'average improvement' mixes higher-better and lower-better metrics, which is sloppy. The self-critic DPO is self-referential—the same model labels its own pairs—which could reinforce biases, but it's a known bootstrap risk, not a demonstrated failure. And there's no audit of the GPT-4o rationales; if they contain errors, training on them could reinforce flawed reasoning.\n\nWho is this for? Practitioners working on hallucination mitigation and instruction data augmentation for LVLMs. It's a legitimate empirical contribution, but it needs revision before it can be trusted as a strong claim. I'd send it to peer review rather than desk reject, but I'd make the revision conditional on adding a generic-insertion control, reporting seeds/error bars, and releasing the data and code. If those hold, the method is a useful addition to the toolbox.","headline":"Re-Critic's rationale-in-the-question twist and self-critic DPO show consistent gains, but the missing generic-insertion control and lack of seeds keep this at promising-preliminary rather than established.","tokens_in":14380,"tokens_out":4150,"would_cite":true,"duration_ms":37642,"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":"Re-Critic inserts reasoning rationales into visual questions before the answer, reducing hallucination while improving general multimodal reasoning.","keywords":["hallucination","large vision-language models","instruction tuning","rationale augmentation","chain-of-thought","self-critic","direct preference optimization","multimodal reasoning"],"falsifier":"Retrain Re-Critic with rationale texts that are deliberately scrambled or describe a different image while preserving the same format and training pipeline, then compare hallucination and reasoning benchmarks: if performance stays about the same, the gains are not caused by the rationale content itself.","tokens_in":13451,"feed_emoji":"🧠","tokens_out":5759,"duration_ms":51823,"temperature":0.7,"pith_summary":"This paper argues that a major cause of hallucination in large vision-language models is missing reasoning context: the model sees a question and an answer but never learns the judgment basis that connects them to the image. To fix that, Re-Critic inserts a machine-generated rationale—a short statement of the relevant principles plus a chain of ideas—into the question before answering, then fine-tunes the model on these augmented instructions. A second stage uses the model's own critique of two of its candidate responses to build preference pairs for DPO, avoiding any external reward model. The experiments report that this recipe improves hallucination benchmarks by 6.2% on average for LLaVA-v1.5 and 2.7% for InternVL2, while also lifting scores on general multimodal reasoning benchmarks. If correct, this makes rationale augmentation a cheap, scalable alternative to curating larger datasets.","feed_headline":"Inserting reasoning steps before the answer cuts hallucinations","feed_subtitle":"Re-Critic's rationale-augmented fine-tuning also lifts general visual reasoning on LLaVA, InternVL2, and MiniGPT-4.","key_machinery":"The central mechanism is the visual chain insertion technique (VCIT), which rewrites each training question by inserting a generated rationale—a statement of the underlying judgment basis and a brief chain of ideas—before the answer is expected. This turns the rationale into part of the input context rather than part of the output, so the model is trained to reason from the rationale first. The second mechanism is the self-critic preference-learning loop, where the currently optimized LVLM evaluates two of its own responses under explicit criteria (image-content understanding, contextual reasoning) and the preferred pair is used for Direct Preference Optimization (DPO). Together they let the model learn methodology before answering.","core_discovery":"Re-Critic claims that hallucinations in LVLMs stem from insufficient context reasoning, and that inserting rationale guidance between question and answer during instruction tuning teaches the model to reason before responding. Concretely, the visual chain insertion technique (VCIT) prompts a proprietary LVLM to generate a \"basic judgment basis\" and a brief chain of ideas from each raw question-answer pair, and places that rationale inside the question input. Then an in-context self-critic mechanism makes the tuned model rank two of its own responses using criteria of image-content understanding and comprehensive contextual reasoning, turning the winner-loser pair into DPO training data. The paper reports consistent gains: average hallucination-benchmark improvements of 6.2% for LLaVA-v1.5 and 2.7% for InternVL2, plus gains on MME, MathVista, LLaVA-Bench, and OCRBench, and larger relative gains when training on the smaller LLaVA-Instruct-80K set.","pith_inferences":["We infer that the method's ceiling is set by rationale quality: because nothing checks the teacher's reasoning, any systematic factual error in the generated rationales would be imprinted into the fine-tuned model.","We infer that the self-critic step inherits the base model's blind spots; a model that cannot see its own mistakes may rank a plausible hallucination over a correct answer, so gains should depend on base capability.","We infer the recipe is portable: inserting rationales into prompts is not tied to vision data and could be applied to any instruction-tuning corpus where a teacher can generate a judgment basis.","A direct test of causality is to corrupt a random fraction of rationales and measure the drop on POPE and MMHalBench; a monotone drop would confirm the rationale content drives the effect."],"forward_implications":["Mixing rationale-augmented instructions into standard SFT data improves hallucination benchmarks without replacing the original dataset.","Self-critic preference learning produces DPO pairs from the model's own judgments, avoiding distribution shift from third-party feedback.","Data efficiency: with only 10k augmented instructions in LLaVA-Instruct-80K, Re-Critic yields large gains on MME and POPE.","Gains transfer to general reasoning: improvements on MathVista, MME, LLaVA-Bench, and OCRBench across two backbones.","The recipe is architecture-agnostic, showing gains on LLaVA-v1.5, InternVL2, and MiniGPT-4."],"supporting_citations":[{"why":"Supplies the chain-of-thought reasoning pattern that Re-Critic inserts as rationale context.","marker":"[Wei et al., 2022]"},{"why":"Provides the DPO objective used for preference tuning after self-critic pair selection.","marker":"[Rafailov et al., 2024]"},{"why":"Defines the LLaVA-v1.5 backbone and training setup that Re-Critic augments and evaluates.","marker":"[Liu et al., 2024b]"},{"why":"Defines the InternVL2 backbone used to test generalizability of the method.","marker":"[Chen et al., 2024a]"},{"why":"Baseline instruction-tuning method for hallucination mitigation that Re-Critic compares against.","marker":"[Liu et al., 2023a]"},{"why":"Baseline RLHF method and the MMHalBench evaluation that Re-Critic uses.","marker":"[Sun et al., 2023]"},{"why":"Baseline HA-DPO method for hallucination-aware preference optimization.","marker":"[Zhao et al., 2023]"},{"why":"Baseline POVID method for modality alignment via preference fine-tuning.","marker":"[Zhou et al., 2024]"},{"why":"Defines POPE, a central benchmark for measuring object hallucination in the experiments.","marker":"[Li et al., 2023]"},{"why":"Defines HallusionBench, used to evaluate visual illusion and knowledge hallucination.","marker":"[Guan et al., 2024]"}],"fun_headline_variants":["Pre-answer reasoning curbs LVLM hallucinations","Re-Critic: reason first, then answer to cut hallucinations","Rationale-before-answer tuning reduces visual hallucinations","Inserting rationale steps before answers cuts LVLM hallucination"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes the GPT-4o-generated rationales are factually accurate and grounded in the image, yet the paper does not audit, filter, or verify them.","fun_headline_variants_meta":{"raw":{"variants":["Pre-answer reasoning curbs LVLM hallucinations","Re-Critic: reason first, then answer to cut hallucinations","Rationale-before-answer tuning reduces visual hallucinations","Inserting rationale steps before answers cuts LVLM hallucination"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3185,"prompt_tokens":931,"completion_tokens":2254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":547,"completion_tokens_details":{"reasoning_tokens":2189}},"tokens_in":547,"tokens_out":2254,"duration_ms":17284,"temperature":1.0,"reasoning_tokens":2189,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:23:22.268080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain Re-Critic with rationale texts that are deliberately scrambled or describe a different image while preserving the same format and training pipeline, then compare hallucination and reasoning benchmarks: if performance stays about the same, the gains are not caused by the rationale content itself.","supporting_citations":[{"cited_title":"Chain-of-thought prompting elicits rea- soning in large language models","cited_arxiv_id":null,"evidence_quote":"Supplies the chain-of-thought reasoning pattern that Re-Critic inserts as rationale context."},{"cited_title":"Direct preference optimization: Your lan- guage model is secretly a reward model.Advances in Neu- ral Information Processing Systems,","cited_arxiv_id":null,"evidence_quote":"Provides the DPO objective used for preference tuning after self-critic pair selection."},{"cited_title":"Hallusion- bench: an advanced diagnostic suite for entangled lan- guage hallucination and visual illusion in large vision- language models","cited_arxiv_id":null,"evidence_quote":"Defines HallusionBench, used to evaluate visual illusion and knowledge hallucination."}],"review_version":1}