{"id":"397283cd-e4f2-4d7f-82d8-13de156bc553","arxiv_id":"2509.04502","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A CoT-annotated dataset (VaccineRAG) plus segment-level GRPO (Partial-GRPO) improves multimodal large language models' ability to ignore harmful retrieved samples in retrieval-augmented generation tasks.","lead":"This paper introduces a new dataset and training method that teach image-and-text AI models to judge whether each retrieved reference is helpful before answering a question, making them more robust when retrieval returns wrong or irrelevant content. It shows accuracy gains on a visual question answering benchmark, but code and data are not yet released and some results are inconsistent.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated by GPT-4o serving as both annotation generator and answer evaluator; independent human scoring is needed to confirm robustness.","rationale":"The reader identified GPT-4o annotation quality as the weakest assumption; my concern is closely related but distinct: GPT-4o is not only the annotation source but also the evaluator, creating a potential circularity. This is more load-bearing because even if annotations are correct, the evaluation may be biased toward the trained model's style. The ADR inconsistency on Qwen2.5-VL (GRPO 0.21 vs Partial-GRPO 10.43 in Table 1) further shows the results are not uniformly robust. I keep the verdict CONDITIONAL because a human evaluation study could settle the concern, but it is a required condition for accepting the central claim.","tokens_in":13421,"tokens_out":8494,"duration_ms":91982,"concrete_test":"Take 200 answerable questions from the WebQA validation set used in evaluation; collect final answers from Qwen2-VL trained with Partial-GRPO and with vanilla GRPO. Have two human annotators, blind to method, judge correctness against ground truth. Compute human-based MA and ADR for both methods and compare with GPT-4o-based numbers. If the ranking flips or the advantage of Partial-GRPO disappears, the reported gains are largely evaluation artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on Table 1, where accuracy is scored by GPT-4o (§6.2). The same model (GPT-4o) generated the CoT annotations in the VaccineRAG training data (§4.1). If models trained on these annotations imitate GPT-4o's answer style, the GPT-4o evaluator may systematically favor them over baselines (SURf, GRPO) that lack this style, independent of actual sample-discrimination ability. The paper asserts GPT-4o scoring 'aligns much more closely with human assessments' but provides no human study on the evaluated models. The observed MA/ADR improvements could therefore be evaluator-model affinity rather than genuine immunity to harmful retrieval samples, directly undermining the strongest claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces VaccineRAG, a Chain-of-Thought (CoT) annotated multimodal RAG dataset constructed from WebQA, together with Partial-GRPO, a variant of GRPO that applies segment-level rewards and partial importance-sampling ratios to different parts of a completion. The authors train Qwen2-VL-7B, Qwen2.5-VL-7B, and InternVL3-8B with SFT followed by Partial-GRPO, and report higher mean accuracy (MA) and lower accuracy degradation rate (ADR) than zero-shot, SURf, SURf+CoT, and vanilla GRPO under polluted retrieval settings, plus gains in Top-K retrieval. The central claim is that the method makes MLLMs robust to harmful retrieved samples while preserving accuracy.","tokens_in":13619,"tokens_out":5648,"duration_ms":63271,"significance":"If the results hold, VaccineRAG would be a useful dataset for studying retrieval robustness in multimodal RAG, and the idea of decomposing rewards by CoT segment is interesting. The three-reward design (format, helpfulness, conclusion) is intuitive, and the authors provide an ablation showing that each reward contributes. The paper also makes a methodological proposal for multi-component GRPO that could be relevant beyond RAG. However, the current evidence is weakened by an unvalidated GPT-4o evaluation, an inconsistent ADR pattern across models, and an incomplete specification of the Partial-GRPO objective. With these issues addressed, the contribution could be significant.","major_comments":[{"comment":"The 'partial importance sampling ratio' is defined as the arithmetic mean of per-token likelihood ratios. For a sequence-level policy ratio, the correct importance weight is the product over tokens of π_θ/π_θ_old. An average of ratios is not a likelihood ratio, so Eq. (4) and Eq. (10) are not unbiased surrogates for the GRPO objective as written. If the implementation uses a product, the equations should be corrected; if the average is deliberate, the authors need to supply a bias/variance analysis and show that the objective still optimizes the intended expected reward.","section":"§5.2, Eq. (3)"},{"comment":"GPT-4o is used both to generate the CoT annotations in the training data and to score the model answers in the main evaluation. The paper asserts that GPT-4o scoring 'aligns much more closely with human assessments' but reports no human study, agreement statistics, or alternative metrics. This introduces a confound: models trained on GPT-4o-generated CoT may imitate GPT-4o's answer style and receive systematically higher scores from the same model, independent of true robustness. Please provide a human evaluation on at least a random subset of all models, with inter-annotator agreement, or use an objective metric (e.g., exact match, token F1) as a second evaluation.","section":"§4.1 and §6.2"},{"comment":"No error bars, multiple seeds, or significance tests are reported. Differences such as Qwen2-VL MA5 56.27 vs. GRPO 54.50 and ADR5 15.98 vs. 23.75 could be within run-to-run noise. Please report results for at least three independent training runs (mean±std) and use a significance test (e.g., paired bootstrap or Wilcoxon) for the main comparisons. This is essential for the central claim of robustness improvement.","section":"Tables 1–3"},{"comment":"The claim in §6.4 that Partial-GRPO 'significantly improved' ADR relative to baselines is contradicted by the Qwen2.5-VL results: vanilla GRPO has ADR5=0.21 while Partial-GRPO has ADR5=10.43, and SURf+CoT has ADR5=5.83. Thus the robustness benefit is not consistent across models. If the main claim is that Partial-GRPO improves MA while preserving or improving ADR, Table 1 shows a model where ADR is much worse than vanilla GRPO. This needs to be discussed and the claim qualified.","section":"Table 1, Qwen2.5-VL row"}],"minor_comments":[{"comment":"The template lists '[reference_4]' twice; the second should presumably be '[reference_5]'.","section":"Figure 1, template"},{"comment":"The format reward is defined only as 'adheres to the format'. Please specify the exact format-checking rule (e.g., presence of the begin/end markers) and how it is applied automatically.","section":"§5.2, Eq. (5)"},{"comment":"The conclusion reward depends on detecting whether a retrieved sample is 'cited' in the conclusion. This is not defined. Please specify the citation-detection mechanism (string matching, model-based span detection, etc.).","section":"§5.2, Eq. (8)"},{"comment":"The helpfulness analysis is generated by prompting GPT-4o with the ground-truth helpfulness label. This is post-hoc rationalization rather than independent annotation. Please clarify how this could bias the CoT and whether the manual verification addresses this.","section":"§4.1"},{"comment":"The claim that GPT-4o scores are 'consistent across multiple assessments' needs a protocol description: how many repeated evaluations, what variance, and whether the same prompt was used each time.","section":"§6.2"},{"comment":"The code and dataset are promised to be 'publicly released soon'. For a reproducibility-oriented paper, please provide a link or state the intended release venue.","section":"Abstract / §7"}],"recommendation":"major_revision","confidential_remarks":"I see no evidence of citation manipulation or novelty concealment; the relation to SURf is acknowledged. The main risk is the GPT-4o annotation/evaluation confound and the mathematical specification of Partial-GRPO. The paper is within scope for a major NLP/ML venue but needs the above revisions before I would support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look. The paper builds the first CoT-annotated multimodal RAG training set (10k WebQA samples with per-reference helpfulness analysis) and introduces Partial-GRPO, a genuine modification: it applies different normalized rewards to different segments of a completion rather than one scalar. That's a plausible fix for a real issue in GRPO with long CoT, and the authors show consistent MA5 gains over SURf and vanilla GRPO across Qwen2-VL, Qwen2.5-VL, and InternVL3. The TopK experiments also show gains at K=10/15. Credit where due: the dataset is a real artifact, the method is clearly described, and the reward decomposition is sensible.\n\nSoft spots are clear too. First, the evaluation stack has a same-model issue: GPT-4o wrote the CoT annotations and GPT-4o scores the answers. The paper says GPT-4o aligns with human assessments but reports no human agreement numbers. If trained models imitate GPT-4o's style, the evaluator could favor them over baselines independent of actual discrimination ability. That's a plausible confound, not a proven fatal flaw, but it needs an independent human study or a second evaluator.\n\nSecond, the headline metric is inconsistent. On Qwen2.5-VL, vanilla GRPO gets ADR5=0.21 while Partial-GRPO gets 10.43 — that's a big regression on degradation robustness, even if MA5 is higher. The paper doesn't discuss this. Third, no error bars or multiple seeds anywhere; with three models and one run each, the 1-3 point MA differences could be noise. Fourth, the format-reward ablation is missing because training was unstable without it — that's fine, but it should be reported transparently. Fifth, code/data are promised but not yet public, which makes the dataset hard to vet.\n\nThe core claim — that CoT supervision plus segment-level GRPO reduces accuracy loss from harmful RAG samples — is plausible and worth testing, but the current evidence is one pass over three models with a single evaluator. I'd send it to review, but the reviewers should push for the human evaluation and the missing seeds.","headline":"A useful CoT/RAG training recipe with a real evaluator-bias risk; the dataset and Partial-GRPO are worth engaging, but the evidence needs error bars and independent scoring.","tokens_in":14098,"tokens_out":1570,"would_cite":true,"duration_ms":16773,"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":"VaccineRAG claims that training multimodal LLMs to analyze each retrieved sample before answering makes them resistant to harmful RAG context, with a segment-level GRPO variant doing the heavy lifting.","keywords":["Retrieval-Augmented Generation","Multimodal Large Language Models","Chain-of-Thought Reasoning","Preference Optimization","GRPO","Robustness","Visual Question Answering","VaccineRAG"],"falsifier":"Take the same training pipeline and deliberately flip the helpfulness labels on 20% of the training samples, then run the polluted-generation evaluation: if accuracy degradation remains close to the normal VaccineRAG result, the method does not depend on annotation accuracy and the claim survives; if degradation jumps toward zero-shot levels, annotation quality is load-bearing and the immunity claim needs qualification.","tokens_in":1753,"feed_emoji":"💉","tokens_out":2017,"duration_ms":89162,"temperature":0.7,"pith_summary":"VaccineRAG attacks an underappreciated weak point in retrieval-augmented generation: even when the retriever works well enough to find useful references, the generator often absorbs irrelevant or misleading ones. The paper introduces a 10k-example multimodal dataset in which each retrieved image or passage is paired with a structured chain of thought—a summary, a helpfulness judgment, a conclusion that cites only helpful references, and then a final answer—so a model can be explicitly taught which evidence to trust. To make that long chain of thought learnable, the authors modify GRPO so that each segment of the output gets its own reward and gradient signal, calling the result Partial-GRPO. Experiments across Qwen2-VL, Qwen2.5-VL, and InternVL3 show higher mean accuracy and lower accuracy degradation than a strong robustness baseline and than vanilla GRPO when harmful retrieved samples are added. If correct, RAG reliability can be improved by training the generator to discriminate evidence rather than by endlessly improving the retriever.","feed_headline":"VaccineRAG cuts accuracy loss from bad retrievals by a third","feed_subtitle":"Chain-of-thought labels plus Partial-GRPO raise mean accuracy and soften the harm from harmful RAG samples.","key_machinery":"The two load-bearing components are the VaccineRAG dataset and Partial-GRPO. VaccineRAG provides dense token-level supervision: each of the roughly five retrieved references per question is paired with a GPT-4o-generated summary, a helpfulness analysis verified against the original WebQA label, a citation-aware conclusion, and the final answer. Partial-GRPO splits a sampled completion into helpfulness-analysis, conclusion, and final-answer segments; each segment has its own reward (helpfulness reward, conclusion reward, format reward), and each reward is normalized across the sampled group and applied through a partial importance-sampling ratio so that gradients affect only the segment the r","core_discovery":"The paper's central claim is that a multimodal RAG system's vulnerability to harmful retrieved samples is not only a retriever problem; it is a generator problem that can be trained away with dense reasoning supervision. VaccineRAG converts sparse retrieval labels into a chain of thought: for each retrieved image or passage the model must produce a summary, decide whether the reference is helpful, write a conclusion that cites only helpful references, and only then give the final answer. Training on this dataset with a short SFT warm-up followed by Partial-GRPO yields higher mean accuracy and lower accuracy degradation than SURf and vanilla GRPO on polluted generation, and in TopK generation","pith_inferences":["Beyond the paper: evaluating the trained models on a different multimodal RAG benchmark or on adversarially constructed distractors would test whether the discrimination skill transfers beyond WebQA's distribution.","Beyond the paper: if the GPT-4o-generated chain-of-thought is replaced with a cheaper or weaker annotator, the drop in performance would reveal how much of the gain is annotation-driven rather than method-driven.","Beyond the paper: Partial-GRPO's segment-level credit assignment should generalize to other long structured generation tasks such as multi-step tool use or long-form QA, though the paper only tests it on RAG.","Beyond the paper: a deployed system could tune the retriever for speed and recall rather than precision, letting the trained generator serve as the final filter against noise."],"forward_implications":["If correct, multimodal RAG systems can tolerate imperfect retrievers: a trained model keeps most of its accuracy even when harmful samples are injected into the context.","Robustness no longer depends on the retriever staying accurate after deployment, since the generator itself learns to discriminate helpful from unhelpful evidence.","The structured chain-of-thought output gives an interpretable audit trail showing which retrieved samples the model accepted, which it rejected, and why.","The ablation results imply all three reward functions—format, helpfulness, and conclusion—are necessary; removing any one measurably hurts accuracy or output stability.","The method transfers across three different 7B-8B vision-language models, suggesting the recipe is not tied to one architecture."],"supporting_citations":[{"why":"Supplies the WebQA dataset—questions, images, retrieved references, and ground-truth helpfulness labels—that VaccineRAG re-annotates with chain-of-thought.","marker":"Chang et al. 2022"},{"why":"SURf is the main robustness baseline; VaccineRAG extends it by adding CoT supervision and compares against it in all main tables.","marker":"Sun et al. 2024"},{"why":"Introduces GRPO, the preference-optimization algorithm that Partial-GRPO builds on and modifies for segment-level credit assignment.","marker":"Shao et al. 2024"},{"why":"Provides the chain-of-thought prompting paradigm that structures VaccineRAG's reasoning annotations.","marker":"Wei et al. 2022"},{"why":"Qwen2.5-VL is one of the three base vision-language models trained and evaluated with the proposed method.","marker":"Bai et al. 2025"},{"why":"Qwen2-VL is another base model used in the experiments and ablations.","marker":"Wang et al. 2024"},{"why":"InternVL3 is the third base model used to test whether the method transfers across architectures.","marker":"Chen et al. 2024c"},{"why":"Provides BGE-VL-base, the retriever used in the TopK generation experiments to construct realistic retrieval sets.","marker":"Zhou et al. 2024"}],"fun_headline_variants":["RAG models get vaccinated against harmful retrieval samples","Chain-of-thought training makes RAG robust to noisy retrievals","VaccineRAG: teach multimodal LLMs to reject bad RAG inputs","Partial-GRPO improves RAG accuracy under polluted retrieval","CoT labels plus RL method make RAG resilient to harmful samples"],"cache_read_input_tokens":16000,"weakest_assumption_plain":"The central claim assumes the GPT-4o-generated summaries, helpfulness judgments, and conclusions—manually checked only when they disagree with the original labels—are correct enough to serve as a teaching signal; if those annotations are noisy or systematically biased, the model may imitate the annotator rather than learn genuine sample discrimination.","fun_headline_variants_meta":{"raw":{"variants":["RAG models get vaccinated against harmful retrieval samples","Chain-of-thought training makes RAG robust to noisy retrievals","VaccineRAG: teach multimodal LLMs to reject bad RAG inputs","Partial-GRPO improves RAG accuracy under polluted retrieval","CoT labels plus RL method make RAG resilient to harmful samples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000576,"raw_usage":{"total_tokens":2569,"prompt_tokens":771,"completion_tokens":1798,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":1724}},"tokens_in":515,"tokens_out":1798,"duration_ms":14283,"temperature":1.0,"reasoning_tokens":1724,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:00:50.965023+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same training pipeline and deliberately flip the helpfulness labels on 20% of the training samples, then run the polluted-generation evaluation: if accuracy degradation remains close to the normal VaccineRAG result, the method does not depend on annotation accuracy and the claim survives; if degradation jumps toward zero-shot levels, annotation quality is load-bearing and the immunity claim needs qualification.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the WebQA dataset—questions, images, retrieved references, and ground-truth helpfulness labels—that VaccineRAG re-annotates with chain-of-thought."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SURf is the main robustness baseline; VaccineRAG extends it by adding CoT supervision and compares against it in all main tables."},{"cited_title":"V.; and Zhou, D","cited_arxiv_id":null,"evidence_quote":"Provides the chain-of-thought prompting paradigm that structures VaccineRAG's reasoning annotations."}],"review_version":1}