{"id":"9456737d-6801-41be-a276-85906f02f0d9","arxiv_id":"2507.16753","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CMP combines LLM-generated semantic references, automatically composed SAM prompts, and frequency-domain alignment to set new state-of-the-art results on four cross-domain few-shot segmentation benchmarks.","lead":"This paper proposes CMP, a method that automatically creates prompts for SAM using LLM-expanded semantics and frequency alignment, and reports state-of-the-art mIoU on four cross-domain few-shot segmentation datasets. Reading it matters if you want to know whether foundation-model prompting can replace manual annotation in specialized imaging domains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not protocol-controlled: CMP's target-domain fine-tuning stage (Sec. 4.2) may explain part of the 10.5-point gain, and the baselines appear to receive no equivalent adaptation.","rationale":"The paper describes a plausible system: LLM-expanded semantic prototypes, composable prompt generation for SAM, and frequency-domain interaction are all reasonable design choices, and the ablation study in Table 2 shows each module contributes on DeepGlobe. I found no clear internal inconsistency in Equations 1-11 or in the module descriptions. However, the headline empirical claim is only as strong as the comparison protocol, and the protocol is not held fixed. The reader's weakest assumption—that all compared methods used the same evaluation protocol—is exactly the load-bearing concern. CMP's extra target-domain fine-tuning is explicitly stated, while no baseline is described as receiving it; without a controlled comparison, the 10.5-point margin cannot be attributed to the proposed architecture. A source-only CMP run would settle this directly. I would keep the verdict at CONDITIONAL: the scientific claim may be true, and the modules may well be effective, but the preprint as written does not establish SOTA under a fair protocol. The paper does deserve credit for a clear ablations and a relatively lightweight method, and the issue is empirical controllability, not fraud or sloppiness in the derivations.","tokens_in":7952,"tokens_out":3373,"duration_ms":41267,"concrete_test":"Re-run CMP with the target-domain fine-tuning stage removed, i.e., evaluate immediately after source meta-training on the same support/query episodes used for Table 1. If the average mIoU drops below the best baseline (APSeg/DR-Adapter) or by more than a few points, the reported SOTA is not attributable to the proposed modules alone; if it remains above 65 mIoU in 1-shot, the fine-tuning concern is weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that CMP outperforms all compared methods by 10.5/8.5 mIoU—rests on Table 1, but the comparison is not controlled for training protocol. Section 4.2 states that CMP uses a two-stage process: source meta-training followed by target-domain fine-tuning, where limited support samples are augmented into pseudo queries for each episode. This is a gradient-based adaptation stage on target-domain data. The compared baselines (APSeg, HQ-SAM, PerSAM, Matcher, and the ResNet-based methods) are not reported as receiving any equivalent target fine-tuning; most are designed to use the support set only as conditioning at inference. If CMP is the only method that updates network weights on target support data, part of the reported gain—especially the +15.1 mIoU on ISIC 1-shot—could come from this extra adaptation rather than from the RCT, CMPG, or FAI modules. This is not an internal inconsistency, but it makes the SOTA claim unverifiable from the preprint. The concern is compounded by the incomplete Matcher row in Table 1 (Chest X-ray and FSS-1000 missing) and by the absence of the target fine-tuning budget (number of epochs, sample counts), code, and error bars.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CMP, a framework that adapts the Segment Anything Model (SAM) for cross-domain few-shot segmentation (CD-FSS). Three modules are introduced: the Reference Complement and Transformation (RCT) module, which uses an LLM and CLIP text/visual encoders to expand semantic prototypes with co-occurring negative classes; the Composable Meta-Prompt Generation (CMPG) module, which combines semantic and geometric information into sparse and dense prompts compatible with SAM's prompt encoder; and the Frequency-Aware Interaction (FAI) module, which aligns cross-domain frequency statistics through a memory bank and performs bidirectional support-query amplitude enhancement. The model is meta-trained on PASCAL VOC and then fine-tuned on target support sets, and is evaluated on DeepGlobe, ISIC2018, Chest X-ray, and FSS-1000. The paper reports state-of-the-art mIoU of 71.8% (1-shot) and 74.5% (5-shot), with ablations on DeepGlobe showing that each module contributes to performance.","tokens_in":8262,"tokens_out":8004,"duration_ms":81773,"significance":"If the reported gains hold under a matched evaluation protocol, the paper would make a strong empirical contribution to SAM-based CD-FSS, demonstrating that composable semantic/geometric prompts together with frequency alignment can improve cross-domain few-shot segmentation by a substantial margin. The method is clearly motivated, the module design is well integrated with SAM's architecture, the internal arithmetic of the reported gains is consistent with Table 1, and the ablation study isolates the contribution of each module in a sensible way. However, the central SOTA claim is currently not protocol-controlled: CMP performs target-domain fine-tuning while the baselines are not reported to receive an equivalent adaptation stage, so the magnitude of the claimed improvement is not yet established. The paper also provides no code, no error bars, and omits several hyperparameters of the fine-tuning and frequency-memory components, which limits verification.","major_comments":[{"comment":"The central SOTA claim is not supported by a protocol-controlled comparison. Section 4.2 states that CMP uses a two-stage process: meta-training on the source domain and fine-tuning on the target domain, in which limited support samples are augmented into pseudo queries for each episode. Table 1 compares CMP with APSeg, HQ-SAM, PerSAM, Matcher, and ResNet-based methods without stating whether any of these baselines received an equivalent target-domain fine-tuning stage. Most of these methods are designed to use the support set only as conditioning at inference. Since CMP updates network weights on target support data, part of the reported +10.5/+8.5 mIoU margins, and especially the +15.1 point gain on ISIC 1-shot, could come from this additional adaptation step rather than from the proposed RCT, CMPG, or FAI modules. The authors should state explicitly whether all compared methods were fine-tuned under the same protocol, and if not, provide a matched comparison, e.g., evaluating CMP without target fine-tuning or fine-tuning all baselines on the same support data. This is necessary to substantiate the claim that CMP surpasses previous best results by 10.5% and 8.5%.","section":"Sec. 4.2 / Table 1"},{"comment":"The target fine-tuning stage is not specified precisely enough to reproduce or interpret the reported results. The paper does not report the number of fine-tuning epochs, the number of pseudo-query augmentations per support sample, the optimizer settings for the fine-tuning stage (beyond the shared learning rate), or any regularization, early stopping, or validation selection criterion. Because this stage is part of the reported 71.8/74.5 mIoU numbers, the missing details block verification of the main claim. The authors should provide a complete description of the fine-tuning budget and the data-augmentation strategy.","section":"Sec. 4.2"}],"minor_comments":[{"comment":"The Matcher row has missing entries for Chest X-ray and FSS-1000; since Table 1 is the basis for the 'previous best' comparison, these entries should be filled in or a footnote should explain their absence.","section":"Table 1"},{"comment":"The notation φv(Is, l) suggests pixel-level CLIP visual features, but CLIP's visual encoder operates on image patches; please clarify how per-pixel or per-region prototypes are obtained.","section":"Sec. 3.2.1 / Eq. (1)"},{"comment":"The hyperparameters τ (threshold in Eq. (6)), T (memory bank size), and α (momentum in Eq. (7)) are not given anywhere in the paper; please report their values or state how they were selected.","section":"Sec. 3.2.3 / Eqs. (6)-(7)"},{"comment":"The normalization operator Norm(·) is not defined; please specify which norm is used and along which axes it is applied.","section":"Sec. 3.2.3 / Eq. (9)"},{"comment":"No code or trained models are provided, and the evaluation protocol (number of random episodes, number of runs, random seeds) is not described; releasing these would substantially aid reproducibility.","section":"Sec. 4.2"},{"comment":"All results are reported as single numbers without standard deviations or the number of runs; given that the ablation difference for CDFA is only 1.0 mIoU, error bars or multi-seed results would help establish the reliability of the smaller ablation effects.","section":"Tables 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The protocol mismatch is the decisive issue. If the authors can supply a matched-protocol comparison (or clearly demonstrate that the baselines already include equivalent target fine-tuning) and provide the missing fine-tuning details, the paper would be suitable for publication. I do not see grounds for rejection, as the proposed method is coherent and the reported internal numbers are consistent, but the SOTA claim is not verifiable from the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read CMP. The paper is a competent system paper for SAM-based cross-domain few-shot segmentation. What's actually new is the specific composition: LLM-expanded semantic prototypes, a composable prompt generator, and frequency-domain alignment combined into one pipeline. Each ingredient has prior art, but the combination and the reported benchmark numbers are new. The ablations are clean and show each module earns its keep, with CMPG the biggest contributor. That part is solid.\n\nThe soft spot is the one the stress-test note flags. Section 4.2 describes a two-stage process: meta-training on source, then fine-tuning on target, where support samples are augmented into pseudo queries. The baselines are not reported as receiving any equivalent target-domain fine-tuning. If CMP is the only method that updates weights on target support data, part of the 10.5-point average gain may come from that extra adaptation, not from the proposed modules. That makes the SOTA claim unverifiable from the preprint. There are also no error bars, no code, missing hyperparameters (memory bank size, momentum, tau, fine-tuning budget), and the Matcher row is incomplete in Table 1. These are fixable in revision, but they matter for a claim that rests entirely on Table 1.\n\nI want to be fair: this is not a broken paper. The internal numbers are consistent, the ablations are sensible, and the limitation section acknowledges extreme domain shifts and coarse masks. The central mechanism is plausible. But the headline 'surpasses previous best by 10.5%' is conditional on a protocol match that the paper does not establish.\n\nWho this is for: people working on CD-FSS, SAM prompting, and lightweight domain adaptation. They would get a practical recipe and a clear baseline to beat. The paper deserves a serious referee, but the referee should push for a protocol-controlled comparison, full hyperparameters, and code release. I would not desk-reject it.","headline":"A plausible SAM-based CD-FSS system whose SOTA claim is undercut by an uncontrolled target fine-tuning stage; worth a serious referee but needs protocol fixes.","tokens_in":8767,"tokens_out":2054,"would_cite":false,"duration_ms":19791,"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":"CMP composes reference-driven semantic and geometric prompts for SAM and aligns source-target frequency statistics, claiming 71.8% and 74.5% mean IoU for 1-shot and 5-shot cross-domain few-shot segmentation on four benchmarks.","keywords":["cross-domain few-shot segmentation","Segment Anything Model","composable meta-prompt","frequency-aware interaction","domain adaptation","few-shot learning","semantic segmentation","prompt engineering"],"falsifier":"Run the full benchmark with every baseline given CMP's exact two-stage protocol, including the same target-domain fine-tuning, the same support-to-pseudo-query augmentation, and the same number of epochs; if the mIoU advantage over APSeg and DR-Adapter on DeepGlobe, ISIC2018, Chest X-ray, and FSS-1000 collapses from 10.5 points to roughly the ablation margins, then the headline gain is mostly protocol rather than the modules.","tokens_in":7749,"feed_emoji":"🧩","tokens_out":7336,"duration_ms":69966,"temperature":0.7,"pith_summary":"This paper tries to establish that the Segment Anything Model can be made to work for cross-domain few-shot segmentation without manual prompts, by composing reference information into an automated meta-prompt and by aligning features in the frequency domain. It introduces CMP, a framework with three modules: RCT expands semantic context using LLM-generated co-occurring classes, CMPG composes semantic and geometric prompts compatible with SAM's prompt encoder, and FAI aligns source-target frequency statistics and enhances support-query consistency. On four target domains (DeepGlobe, ISIC2018, Chest X-ray, and FSS-1000) with PASCAL VOC as the source, CMP reports 71.8% and 74.5% mean Intersection over Union for 1-shot and 5-shot settings, surpassing the previous best by 10.5 and 8.5 percentage points. If correct, this would make foundation-model prompting a practical route for segmentation in medical and satellite domains with only one or a few annotated examples.","feed_headline":"Automated meta-prompts lift SAM to 71.8% mIoU on four domains","feed_subtitle":"Prompt synthesis plus frequency alignment beats the prior best by 10.5 points in 1-shot cross-domain segmentation.","key_machinery":"The central object is the Composable Meta-Prompt (CMP), a learnable composition of a semantic embedding (enhanced foreground and background prototypes, including LLM-generated negative classes) and a geometric mask embedding (SAM's mask encoder plus a learned adapter), fused through a prompt alignment block into SAM's sparse and dense prompts. The Frequency-Aware Interaction (FAI) module carries the domain-adaptation load through Fast Fourier Transform decomposition, a memory bank of source frequency statistics for cross-domain amplitude alignment, and bidirectional affine transforms between support and query amplitude spectra. These modules are the only trainable parts of the framework, and the paper attributes its performance gains to them through ablations.","core_discovery":"The central claim is that SAM's two obstacles in cross-domain few-shot segmentation—reliance on manual prompts and limited cross-domain ability—can be overcome by a composable meta-prompt mechanism plus frequency-aware interaction. CMP trains only lightweight prompt-generation and alignment modules while keeping SAM's image encoder and mask decoder frozen. The RCT module asks a large language model what objects co-exist with the target class, encodes those negative classes with CLIP, and combines them with visual prototypes from the support mask; the CMPG module turns these semantic and mask priors into sparse and dense prompts compatible with SAM's prompt encoder; and the FAI module decomposes support and query features with FFT, aligns amplitude statistics through a memory bank, and performs bidirectional amplitude exchange between support and query. The paper's evidence is its benchmark table: average mIoU of 71.8% in 1-shot and 74.5% in 5-shot settings across DeepGlobe, ISIC2018, Chest X-ray, and FSS-1000, with a DeepGlobe ablation showing CMPG is the largest contributor (-6.7%), followed by RCT semantic expansion (-4.0%) and the full FAI module (-3.2%).","pith_inferences":["Beyond the paper: the comparison is only fair if every SAM-based baseline also received target-domain fine-tuning; the paper does not state that APSeg, HQ-SAM, PerSAM, or Matcher were given the same two-stage protocol, so a re-benchmark with matched fine-tuning could shift the reported margins.","Beyond the paper: the RCT module depends on LLM-generated co-occurring classes, so a failure mode worth probing is whether a hallucinated or contextually inappropriate negative class degrades the prompt; this could be tested by perturbing the LLM output or using class names from different domains.","Beyond the paper: the same composable-prompt-plus-frequency-alignment recipe could transfer to other promptable segmentation models and to related low-data tasks such as referring segmentation or medical image segmentation, where support sets are naturally small.","Beyond the paper: because target fine-tuning augments support samples into pseudo queries, the framework implicitly assumes the support mask is reliable; noisy or coarse support masks at test time could weaken the gains, consistent with the paper's stated limitation."],"forward_implications":["A single annotated support image can produce a domain-adaptive prompt, so the pipeline becomes usable in settings where expert annotation is scarce, such as medical imaging and remote sensing.","Because SAM's parameters remain frozen, the method adds only lightweight trainable modules, keeping adaptation cost far below full fine-tuning of the foundation model.","Frequency-domain alignment offers a complementary intervention to spatial-only feature transformation in cross-domain few-shot segmentation, with the ablation attributing 3.2 mIoU points to it on DeepGlobe.","If the reported numbers hold under a matched protocol, CMP becomes the leading method on this four-dataset cross-domain few-shot segmentation benchmark, ahead of the previous best by 10.5 points in 1-shot and 8.5 points in 5-shot settings."],"supporting_citations":[{"why":"Supplies the Chest X-ray target domain used for pulmonary screening evaluation.","marker":"[1]"},{"why":"Supplies the ISIC2018 skin-lesion target domain used for cross-domain evaluation.","marker":"[2]"},{"why":"Supplies the DeepGlobe satellite-imagery target domain where the main ablations are run.","marker":"[3]"},{"why":"Provides the frozen Segment Anything Model backbone whose prompt encoder and mask decoder CMP adapts.","marker":"[9]"},{"why":"PATNet is a cross-domain few-shot segmentation baseline whose reported results CMP compares against.","marker":"[16]"},{"why":"DMTNet is a recent ResNet-based baseline whose results anchor the comparison table.","marker":"[17]"},{"why":"DR-Adapter is a domain-rectifying adapter baseline that defines the prior 1-shot standard on several target domains.","marker":"[18]"},{"why":"APSeg is the best prior SAM-based auto-prompt baseline that CMP must beat.","marker":"[19]"},{"why":"ABCDFSS is an adapt-before-compare baseline included in the state-of-the-art comparison.","marker":"[20]"},{"why":"Supplies the FSS-1000 target domain covering 1000 classes of natural-object few-shot segmentation.","marker":"[24]"}],"fun_headline_variants":["Meta-prompts automate SAM for cross-domain few-shot segmentation","CMP: Composable meta-prompts lift SAM to 71.8% mIoU","Frequency-aware prompts beat prior best in few-shot segmentation","SAM with auto-prompts hits 71.8% mIoU on four domains","Composable meta-prompt framework improves SAM across domains"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported superiority assumes every compared method was judged under the same protocol, yet CMP alone receives an extra target-domain fine-tuning stage that turns support samples into pseudo queries; if the baselines were not allowed that same extra adaptation stage, part of the measured gain could come from that advantage rather than from the proposed modules.","fun_headline_variants_meta":{"raw":{"variants":["Meta-prompts automate SAM for cross-domain few-shot segmentation","CMP: Composable meta-prompts lift SAM to 71.8% mIoU","Frequency-aware prompts beat prior best in few-shot segmentation","SAM with auto-prompts hits 71.8% mIoU on four domains","Composable meta-prompt framework improves SAM across domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000447,"raw_usage":{"total_tokens":2273,"prompt_tokens":979,"completion_tokens":1294,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":595,"completion_tokens_details":{"reasoning_tokens":1201}},"tokens_in":595,"tokens_out":1294,"duration_ms":10036,"temperature":1.0,"reasoning_tokens":1201,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:02:30.274110+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full benchmark with every baseline given CMP's exact two-stage protocol, including the same target-domain fine-tuning, the same support-to-pseudo-query augmentation, and the same number of epochs; if the mIoU advantage over APSeg and DR-Adapter on DeepGlobe, ISIC2018, Chest X-ray, and FSS-1000 collapses from 10.5 points to roughly the ablation margins, then the headline gain is mostly protocol rather than the modules.","supporting_citations":[{"cited_title":"CMP: A Composable Meta Prompt for SAM-Based Cross-Domain Few-Shot Segmentation","cited_arxiv_id":"2507.16753","evidence_quote":"Supplies the Chest X-ray target domain used for pulmonary screening evaluation."},{"cited_title":"Few-Shot Segmentation Few-shot semantic segmentation (FSS) tackles novel class segmentation with limited labeled examples through two main pipelines","cited_arxiv_id":null,"evidence_quote":"Supplies the ISIC2018 skin-lesion target domain used for cross-domain evaluation."},{"cited_title":"For an image containing [class name Ts], what other objects might co-exist?","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepGlobe satellite-imagery target domain where the main ablations are run."},{"cited_title":"One-shot learning for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the frozen Segment Anything Model backbone whose prompt encoder and mask decoder CMP adapts."},{"cited_title":"Relevant intrinsic feature enhancement network for few-shot semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"PATNet is a cross-domain few-shot segmentation baseline whose reported results CMP compares against."},{"cited_title":"Hierarchi- cal dense correlation distillation for few-shot segmenta- tion,","cited_arxiv_id":null,"evidence_quote":"DMTNet is a recent ResNet-based baseline whose results anchor the comparison table."},{"cited_title":"Extending segment anything model into auditory and temporal dimensions for audio-visual segmentation,","cited_arxiv_id":null,"evidence_quote":"DR-Adapter is a domain-rectifying adapter baseline that defines the prior 1-shot standard on several target domains."},{"cited_title":"Bridge the points: Graph-based few- shot segment anything semantically,","cited_arxiv_id":null,"evidence_quote":"APSeg is the best prior SAM-based auto-prompt baseline that CMP must beat."},{"cited_title":"Asam: Boosting segment anything model with adversarial tuning,","cited_arxiv_id":null,"evidence_quote":"ABCDFSS is an adapt-before-compare baseline included in the state-of-the-art comparison."},{"cited_title":"Apseg: Auto- prompt network for cross-domain few-shot semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the FSS-1000 target domain covering 1000 classes of natural-object few-shot segmentation."}],"review_version":1}