{"id":"10922722-1015-4284-8cd2-ee4e7cea1578","arxiv_id":"2509.03494","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"With a learned 30-pixel border prompt added to input images, a frozen mPLUG-Owl2-7B reaches 0.932 SRCC on KADID-10k using about 156K trainable parameters.","lead":"A frozen image-and-text AI model can learn to judge picture quality from a tiny trainable pixel pattern added to each image, hitting 0.93 rank correlation on one benchmark with under 0.01% of weights tuned. The result points to a near-free way to repurpose large multimodal models for new visual tasks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"KADID evaluation protocol may allow train/test content overlap and test-set prompt selection, so the headline 0.932 SRCC may not be protocol-supported.","rationale":"The reader correctly identified the missing error bars, unclear training protocol, and post-hoc prompt-type selection as reasons to be cautious. My stress-test sharpens this into a concrete, potentially fatal protocol flaw: KADID's 81-content structure means the 80/10/10 split must be content-disjoint to be comparable to baselines, and the paper does not say it is. This is more load-bearing than the mechanism-level worry about whether a global additive pattern can align logits, because the empirical claim is the central evidence; if the split is flawed, the number itself is not meaningful. I agree with the reader's conditional verdict but would add an explicit condition: demonstrate content-disjoint splitting and validation-based selection before the headline KADID number can be taken at face value. I did not choose REJECT because the source code is public and the split can be checked; this is a testable, fixable protocol issue rather than a demonstrated false result. The mechanism concern (can a single prompt align frozen features to quality) is interesting but less decisive: the reported gains show it can work empirically, provided the evaluation is sound.","tokens_in":9755,"tokens_out":4861,"duration_ms":56291,"concrete_test":"Inspect the released code/data splits and verify whether any KADID reference content appears in both train and test sets; if overlap exists, rerun the 30px padding experiment with a content-disjoint 80/10/10 split over the 81 reference images, select the prompt type on the validation split, and report mean±std SRCC/PLCC over at least 5 splits. If the KADID SRCC drops below 0.91 or the selected prompt type changes from 30px padding, the headline claim is not supported.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim—frozen mPLUG-Owl2-7B plus a 155,880-parameter padding prompt reaches 0.932 SRCC on KADID-10k—depends on the evaluation protocol being sound, but §4.1 does not state that the 80/10/10 split is content-disjoint. KADID-10k contains 10,125 distorted images derived from only 81 reference images; a random image-level split places near-duplicates (same content, different distortion levels) in both train and test. A learned additive border can then exploit content identity rather than generalize quality, artificially inflating SRCC. This matters because the comparison baselines (LIQE, Q-Align, specialized models) are typically evaluated with content-disjoint splits. Additionally, Table 2 reports all seven prompt configurations on the test set and the 30px padding is selected as best without any stated validation-based choice, so the headline number is a test-set maximum. The Table 2 vs Table 3 KonIQ PLCC inconsistency (0.874 vs 0.865) compounds the concern that the reported 'Our Proposed Method' row is not a single pre-specified protocol. If the KADID split is actually image-level rather than reference-level, the headline result is not comparable to the cited baselines, and the central claim that pixel-level prompting is competitive with full fine-tuning is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a parameter-efficient NR-IQA method that learns a pixel-space visual prompt (padding, fixed patch, or full overlay) added to the input image of a frozen mPLUG-Owl2-7B MLLM. A quality score is obtained by a softmax over positive (\"good\", \"fine\") and negative (\"poor\", \"bad\") token logits, and the prompt is trained with MSE regression against MOS. The authors report that the best configuration, a 30px padding with 155,880 trainable parameters, achieves SRCC 0.932 on KADID-10k, and they claim competitive performance with fully fine-tuned MLLMs and specialized NR-IQA models on KADID-10k, KonIQ-10k, and AGIQA-3k, while training at most ~600K parameters (<0.01% of the base model). The paper also claims to be the first to apply pixel-space visual prompting to NR-IQA.","tokens_in":10121,"tokens_out":3781,"duration_ms":40665,"significance":"If the reported results hold under a sound evaluation protocol, the contribution is significant: it demonstrates that a frozen 7B MLLM can be adapted to NR-IQA with a tiny additive pixel pattern, which would offer a highly parameter-efficient alternative to full fine-tuning and avoid model forgetting. The paper's strengths include a simple and reproducible pipeline, public code, standard datasets, explicit parameter counts, and a clear comparison with a range of baselines. The novelty of pixel-level visual prompting for NR-IQA is genuine. However, the central quantitative claim depends on the evaluation protocol, and the manuscript currently does not establish that protocol rigorously: the KADID-10k split is not shown to be content-disjoint, the prompt configuration is selected after seeing test-set results, and the reported numbers are single-run with at least one internal inconsistency.","major_comments":[{"comment":"The KADID-10k evaluation protocol is underspecified and potentially invalid. KADID-10k contains 10,125 distorted images derived from only 81 reference images. The paper states \"we select 80% of the dataset for training, 10% for validation, and the remaining 10% for testing\" but does not state that this split is reference-content-disjoint. A random image-level split will place distortions of the same content in both train and test, allowing the learned global additive prompt to exploit content identity rather than learn generalizable quality assessment. The comparison baselines (e.g., LIQE, Q-Align) are typically evaluated with content-disjoint splits, so the reported 0.932 SRCC may not be comparable. Please clarify the split and, if it is not content-disjoint, re-run the experiments accordingly.","section":"§4.1 (Experimental setup)"},{"comment":"The headline result appears to be selected post hoc from test-set performance. Table 2 reports SRCC/PLCC for seven prompt configurations (padding 10/30px, center patch 10/30px, top-left patch 10/30px, full overlay) on the same test sets. The text then says \"Based on these results, we notice that the 30px padding is the most correlated type of visual prompt with the MOS across all datasets.\" This is a test-set maximum; no validation-based model selection is described. The 0.932 SRCC is therefore the best of seven configurations chosen after seeing test labels, which inflates the expected performance of the proposed method. Please use the validation split for selection, or report the full selection rule and adjust the significance claims.","section":"§4.2 / Table 2"},{"comment":"There is an internal inconsistency in the reported results for KonIQ-10k. Table 2 reports PLCC=0.874 for the 30px padding prompt, while Table 3 reports PLCC=0.865 for \"Our Proposed Method\", which presumably is the same configuration. If they are the same, the numbers must match; if they differ, the caption or text should explain why. This discrepancy, combined with single-run reporting, makes it difficult to verify that the paper's comparison is based on a single pre-specified protocol. Please clarify and, ideally, report results over multiple seeds (mean±std).","section":"Table 2 vs Table 3"},{"comment":"The paper does not provide any analysis of what the learned visual prompt encodes or whether the underlying assumption is valid: that a single global additive pattern can steer the frozen model's softmax over the four chosen quality words into a monotonically quality-responsive score. Since neither the vision encoder nor the language model is updated, if the base model's logits for {good, fine} vs {poor, bad} are not responsive to quality for the test distribution, the prompt cannot compensate. Please include a diagnostic (e.g., prompt visualization, logit analysis, or cross-dataset transfer experiments) to support the mechanism.","section":"§4.2 (Implementation details) / §3.1"}],"minor_comments":[{"comment":"The choice of mPLUG-Owl2-7B is justified by API compatibility; please provide the exact model version/checkpoint and preprocessing details (resize and center-crop size) for reproducibility.","section":"§3.2"},{"comment":"The positive/negative token sets are fixed to \"good\", \"fine\" vs \"poor\", \"bad\". Please clarify whether these are the only tokens considered and whether results are sensitive to this choice; a brief ablation would strengthen the paper.","section":"§3.4"},{"comment":"Hyperparameters are unusual (e.g., learning rate 60 for SGD). Please specify momentum, weight decay, learning-rate schedule, and whether the learning rate is per-example or batch-scaled.","section":"§4.2"},{"comment":"The pixelated regions in Figure 2 are hard to interpret. A clearer indication of the prompt region (e.g., a highlight or zoomed inset) would improve readability.","section":"Figure 2"},{"comment":"Reference formatting is inconsistent: some entries lack venues or page numbers (e.g., [1], [18], [34]), and [34] is a bare URL. Please standardize.","section":"References"},{"comment":"Typographical issue: page 2 has the header \"APREPRINT- SEPTEMBER12, 2025\" with a missing space; please fix.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the core idea is novel and interesting. The main risk is the evaluation protocol: the KADID-10k split must be content-disjoint, and prompt selection must be validation-based rather than test-based. If those issues can be fixed and the results remain similar, the paper could be a useful contribution. I am not recommending rejection because the load-bearing concerns are addressable within the manuscript's scope, but the current reporting does not support the headline claim as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate new application of Bahng-style pixel prompts to NR-IQA, and the main effect is believable. But the headline 0.932 SRCC on KADID-10k is not protocol-supported as reported: the split is not stated to be content-disjoint, the prompt type/size was chosen on test-set numbers, and there's a PLCC inconsistency between Table 2 and Table 3 (0.874 vs 0.865 on KonIQ). Fix those and the paper is worth publishing.\n\nWhat's genuinely good: the method is simple, parameter counts are honest, the code is linked, and the comparison table includes the relevant baselines. The results show a frozen 7B MLLM with a learned border prompt clearly beats the frozen baseline and is competitive on KADID with full fine-tuning. That's a useful data point for parameter-efficient IQA.\n\nSoft spots, in order of severity:\n1. KADID protocol. The paper says 'official splits when available, otherwise 80/10/10' and never says the 80/10/10 is reference-disjoint. KADID-10k has 81 source images; a random image split puts near-duplicates across train/test. The learned 30px border then can partially memorize content identity, not just quality. The baselines (LIQE, Q-Align, MP-IQE) are typically evaluated on content-disjoint splits, so the 0.932 is not comparable as reported. This is the load-bearing issue.\n2. Prompt selection on the test set. Table 2 reports seven prompt configurations; 30px padding is chosen as best from those numbers on the same test sets. No held-out validation for prompt-type selection. The 'Our Proposed Method' row is a test-set maximum, and with no error bars you can't know how much of the gap is selection noise.\n3. Internal inconsistency. KonIQ PLCC for 30px padding: 0.874 in Table 2, 0.865 in Table 3. Minor but undermines confidence in the reported exact values.\n4. Missing analysis. No study of what the learned prompt encodes, no cross-dataset transfer, no ablation on the textual prompt or the token sets beyond what's there.\n\nThe reader's weakest_assumption is fair but secondary: the method's ceiling depends on the frozen model's logits being quality-responsive, which is plausible given Q-Bench results but untested here.\n\nBottom line: the core idea is simple, cheap, and probably real, but the central empirical claim needs a content-disjoint split, pre-registered prompt selection (or at least validation-based), and error bars before I'd trust the 0.932. The paper deserves peer review because the application is new and the experiments are standard; round one should be 'revise and re-evaluate with a clean protocol.'","headline":"A credible, cheap visual-prompting recipe for NR-IQA, but the headline KADID number is built on a test-set selection and possibly content-overlapping split.","tokens_in":10628,"tokens_out":2193,"would_cite":false,"duration_ms":21157,"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":"A frozen 7B vision-language model can be adapted to score image quality by learning a 155,880-pixel border prompt.","keywords":["no-reference image quality assessment","visual prompting","pixel-space prompts","multimodal large language models","parameter-efficient adaptation","mPLUG-Owl2","softmax quality scoring","KADID-10k"],"falsifier":"Train the 30-pixel padding prompt on KADID-10k, freeze it, and apply it to KonIQ-10k: if SRCC drops to near the unfrozen baseline of about 0.60, the prompt encodes KADID-specific shortcuts rather than a general quality axis.","tokens_in":9669,"feed_emoji":"🖼️","tokens_out":6985,"duration_ms":65343,"temperature":0.7,"pith_summary":"This paper seeks to establish that a frozen multimodal large language model can be turned into a no-reference image-quality assessor by learning only a small pattern of pixels added to the input image, leaving every model weight untouched. On the KADID-10k synthetic-distortion benchmark, that pixel prompt (a 30-pixel border) lifts Spearman correlation from 0.550 for the unadapted model to 0.932, matching or exceeding fully fine-tuned 7B models such as LIQE and Q-Align while training around 155,880 parameters, less than 0.01 percent of the base model. The same recipe works, though less strongly, on realistic (KonIQ-10k) and AI-generated (AGIQA-3k) distortions. If correct, the result means expensive fine-tuning and per-task model copies are not necessary: one frozen base model can be redirected to a new low-level vision task by swapping a tiny additive visual prompt. The authors claim this is the first pixel-level visual-prompt adaptation of an MLLM for image quality assessment.","feed_headline":"155K learned pixels turn a frozen 7B model into an image-quality scorer","feed_subtitle":"Adding a 30px border prompt to mPLUG-Owl2 matches fully fine-tuned 7B models on KADID-10k.","key_machinery":"The load-bearing object is the additive visual prompt: a trainable tensor of the same shape as the input image (a border, a fixed patch, or a full overlay), bounded by tanh and clamped to [0,1], added to the image before the MLLM's vision encoder. Because the model is frozen, the prompt is the only channel through which training signals can act. The score extraction is equally important: no regression head is learned; the quality score is a softmax over four vocabulary logits (good, fine versus poor, bad) from the model's final token, so the learned pixels must reorganize the frozen model's internal response to that vocabulary. The 30-pixel padding works best, which the paper ties to prompt","core_discovery":"The central claim is that NR-IQA can be posed as a regression whose only trainable parameters are pixels. The method feeds the frozen mPLUG-Owl2-7B a fixed textual query, \"Rate the technical quality of the image,\" alongside the input image plus a learned visual prompt; the model's final-token logits for the two positive words {good, fine} and two negative words {poor, bad} are combined into a softmax score in (0,1), trained by mean-squared error against human MOS. The best configuration is a 30-pixel padding around all four sides, with 155,880 parameters, reaching SRCC 0.932 and PLCC 0.929 on KADID-10k. Ten-pixel or small fixed patches fail, while a full overlay is competitive on KADID but w","pith_inferences":["The learned border probably acts as a global attention bias rather than encoding local distortion features; if so, its transfer across datasets should be limited, and ablating quadrants of the border should show where the signal lives.","Cross-dataset transfer is untested in the paper. A natural test is to train the padding on KADID-10k and evaluate on KonIQ-10k; because the two distributions differ sharply, the prompt may collapse toward the frozen baseline.","Combining a learned pixel prompt with a learned textual prompt is the obvious next step; the paper itself notes the gap to CLIPIQA+ on KonIQ, so co-tuning both modalities could close it.","The four-word vocabulary makes the score a coarse ordinal; expanding the positive and negative token sets or using per-level labels could change results and may be needed for a higher dynamic range."],"forward_implications":["One frozen mPLUG-Owl2 can be reused across datasets and tasks by swapping a small stored pixel prompt, instead of storing a separate 7B fine-tuned model per dataset.","Visual prompt geometry is decisive: a 30-pixel border yields 0.932 SRCC on KADID-10k while a 10-pixel patch yields 0.390, so prompt size and placement are first-class design choices.","Because neither the vision encoder nor the language model is updated, adaptation avoids catastrophic forgetting and lowers memory requirements dramatically relative to full fine-tuning.","The method still trails fully fine-tuned methods on KonIQ-10k and AGIQA-3k, indicating pixel-only adaptation is not yet a complete substitute for weight updates on realistic and AI-generated distortions.","The softmax-over-quality-words scoring trick from Q-Bench is sufficient when combined with pixel prompting to align model outputs with MOS, without architectural changes."],"supporting_citations":[{"why":"mPLUG-Owl2, the frozen base MLLM whose interface and behavior the whole method depends on.","marker":"[10]"},{"why":"Supplies the additive padding and fixed-patch visual prompting recipe, including the add-prompt-to-input operation.","marker":"[11]"},{"why":"Q-Bench provides the softmax-over-logits scoring strategy and the base textual prompt the method adapts.","marker":"[7]"},{"why":"KADID-10k, the main training and evaluation dataset behind the headline 0.932 SRCC result.","marker":"[13]"},{"why":"KonIQ-10k, the realistic-distortion dataset used to test and compare the method.","marker":"[12]"},{"why":"AGIQA-3k, the AI-generated-image dataset used to test and compare the method.","marker":"[14]"},{"why":"LIQE, the fully fine-tuned vision-language baseline that the KADID result is compared against.","marker":"[9]"},{"why":"Q-Align, the fully fine-tuned MLLM baseline for quality scoring that the method matches on KADID-10k.","marker":"[5]"},{"why":"CoOp, the textual-prompt optimization baseline that this pixel-prompt approach extends from textual to visual input space.","marker":"[23]"}],"fun_headline_variants":["Frozen 7B model scores image quality via learned pixels","30-pixel border prompt beats full fine-tuning on KADID","Train only pixels to match full MLLM tuning for IQA","155K pixels turn frozen MLLM into strong IQA scorer"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The frozen model's internal logits for the words good, fine, poor, and bad must respond monotonically to image quality, and a single fixed pixel pattern must be able to push that response into the right order on all test images.","fun_headline_variants_meta":{"raw":{"variants":["Frozen 7B model scores image quality via learned pixels","30-pixel border prompt beats full fine-tuning on KADID","Train only pixels to match full MLLM tuning for IQA","155K pixels turn frozen MLLM into strong IQA scorer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1085,"prompt_tokens":792,"completion_tokens":293,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":219}},"tokens_in":536,"tokens_out":293,"duration_ms":3705,"temperature":1.0,"reasoning_tokens":219,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:51:54.161067+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the 30-pixel padding prompt on KADID-10k, freeze it, and apply it to KonIQ-10k: if SRCC drops to near the unfrozen baseline of about 0.60, the prompt encodes KADID-specific shortcuts rather than a general quality axis.","supporting_citations":[{"cited_title":"KADID-10k: A Large-scale Artifi- cially Distorted IQA Database","cited_arxiv_id":null,"evidence_quote":"KADID-10k, the main training and evaluation dataset behind the headline 0.932 SRCC result."},{"cited_title":"KonIQ-10k: An ecologically valid database for deep learning of blind image quality assessment","cited_arxiv_id":"1910.06180","evidence_quote":"KonIQ-10k, the realistic-distortion dataset used to test and compare the method."},{"cited_title":"AGIQA-3K: An Open Database for AI-Generated Image Quality Assessment","cited_arxiv_id":"2306.04717","evidence_quote":"AGIQA-3k, the AI-generated-image dataset used to test and compare the method."},{"cited_title":"Blind Image Quality Assessment via Vision-Language Correspondence: A Multitask Learning Perspective","cited_arxiv_id":"2303.14968","evidence_quote":"LIQE, the fully fine-tuned vision-language baseline that the KADID result is compared against."}],"review_version":1}