{"id":"afe01fb7-7dc0-4731-8c5c-13da7c4270e4","arxiv_id":"2607.12292","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"SERD thresholds SAM3's internal semantic response combined with a fixed Sobel edge prior, improving cross-domain crack segmentation over native SAM3 by 3.3–3.7 IoU points without training.","lead":"This paper shows that reading SAM3's internal response map—rather than its final region proposals—recovers thin cracks that are lost during proposal formation. A frozen foundation model can then segment cracks across new road, bridge, or tunnel datasets without retraining, beating native SAM3 by about 4 IoU points.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 4's per-image min-max normalization discards absolute response calibration, so the single threshold τ is a per-image quantile; cross-domain transfer of the reported gain is not established independently of the source-selected τ.","rationale":"The reader's weakest assumption is exactly the condition I would stress: per-image min-max normalization makes τ a relative threshold and removes absolute response calibration. I agree with that identification and sharpen it with two pieces of internal evidence: the source-selected thresholds in Table 3 range from 0.30 to 0.50, and the mean target PAR of SERD (1.421) is much higher than SAM3's (0.944), which is consistent with a threshold that is too permissive on some domains after min-max rescaling. The rotated-source experiment is the paper's best defense, but it only tests six source calibrations and reports macro-averaged means; it does not establish that the fixed τ is near target-optimal within each domain. The paper does have real supporting evidence: the method is deterministic with one operating parameter, the code is released, the threshold-sweep analysis shows a stable plateau around τ=0.45, and Section 6 reports paired-bootstrap confidence intervals for the morphology metrics. Those considerations prevent the concern from being fatal. However, because the headline cross-domain IoU gain depends on a single scalar threshold selected under source supervision and a normalization that erases absolute scale, I would not move the verdict to ACCEPT. The reader's CONDITIONAL verdict is appropriate, so no adjustment is needed.","tokens_in":16597,"tokens_out":8532,"duration_ms":85876,"concrete_test":"Run the released SERD code on all six test sets. For each test image, compute the oracle threshold τ* maximizing IoU and compare the distribution of τ* across datasets with τ=0.45 and with the source-selected thresholds from Table 3. If per-image optimal thresholds vary substantially by domain, or if using per-dataset cross-validated thresholds changes the SERD-minus-SAM3 gap by more than about 1 IoU point, the transferability claim rests on an accidental calibration. As a second check, replace the min-max normalization in Eq. 4 with robust 1st–99th percentile clipping or z-score normalization and rerun the Table 1 and Table 3 protocols; if the advantage over SAM3 collapses or reverses, Eq. 4 is the load-bearing component and the method is not robust to response-scale shift.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is Eq. 4: Rq is min-max normalized independently for each image, so the absolute scale of SAM3's semantic response is removed. The fixed threshold τ=0.45 is therefore not a calibrated confidence but a relative quantile: in a low-contrast or weak-response image, the same τ can retain a much larger fraction of pixels than in a high-contrast image. This is exactly what must hold for the 'training-free transfer' claim: one threshold must remain meaningful when response distributions shift across materials, illumination, and imaging conditions. The rotated-source protocol (Table 3) only shows that the mean over five target domains is stable; it does not show that τ=0.45 is near-optimal on each target. In fact, Table 3's source-selected thresholds vary from 0.30 to 0.50, and SERD's mean target PAR is 1.421 versus 0.944 for native SAM3, indicating systematic over-segmentation in some domains. If response distributions shift, the source-selected τ can systematically over- or under-segment, and the headline 58.00% vs. 54.33% could depend on the CamCrack789 calibration choice rather than on a domain-invariant response interface. A related signal is that Table 3 shows direct Response (no Sobel) has essentially the same mean IoU as SERD (60.15 vs. 60.23), so the Sobel calibration is not what drives the headline IoU transfer; the transferable component is the uncalibrated response plus a chosen τ. Thus the central claim of a transferable training-free threshold is least secure precisely where Eq. 4 removes absolute response calibration.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SERD, a training-free decoder that replaces SAM3's discrete region proposals with the model's internal language-conditioned dense response to the prompt 'crack'. The response is sigmoid-activated, resized, min-max normalized per image (Eq. 4), modulated by a fixed Sobel edge prior (Eq. 8), and thresholded by a single scalar τ. SAM3 remains frozen; no target-domain labels or test-time optimization are used. The method is evaluated on six crack datasets under a unified fixed threshold and under six rotated source-to-target protocols, reporting average Crawl IoU gains over native SAM3 (58.00 vs. 54.33 in the CamCrack789-source protocol; 60.23 vs. 56.95 in the rotated-source protocol), alongside precision, boundary, area, robustness, and latency analyses.","tokens_in":17001,"tokens_out":5008,"duration_ms":56386,"significance":"If the results hold, the paper makes a useful empirical contribution: it shows that for thin, fragmented structures, the internal response of a frozen text-promptable foundation model can be a more informative output interface than its final proposals, and that a simple deterministic decoder can transfer across data domains. The study is unusually transparent: the single threshold is disclosed and source calibration is described, code is promised, the evaluation includes six datasets and rotating source-domain calibration, and the engineering-fidelity analysis in Section 6 provides paired-bootstrap confidence intervals. The main quantitative claim is plausible but not yet fully established, largely because the central IoU comparisons lack uncertainty quantification and the threshold-transfer property is not directly probed on target domains.","major_comments":[{"comment":"Eq. (4) normalizes each image independently, so τ is not a calibrated confidence but a per-image relative quantile. The paper's central 'one threshold transfers' claim therefore requires evidence that τ remains meaningful when response distributions shift across domains. The rotated-source protocol is helpful, but it reports only aggregate means; the selected thresholds in Table 3 vary from 0.30 to 0.50, and the mean SERD PAR is 1.421, suggesting systematic over-coverage. Fig. 7 sweeps τ only on the CamCrack789 training split. Please add target-domain threshold sweeps or an equivalent analysis, e.g., report for each target dataset the range of τ that stays within 1 IoU point of that target's best τ, and state whether the 3.67-point advantage persists when target-optimal thresholds are used. Without this, the headline gain could depend on the source-calibration choice.","section":"§3.3 (Eq. 4), §5.5, Table 3"},{"comment":"The primary IoU comparison is reported as point estimates without confidence intervals, standard errors, or paired significance tests. The gain is averaged over five datasets and varies from +8.94 to +1.19 points across datasets, so the aggregate alone does not establish that the improvement is systematic. The paper already computes paired-bootstrap intervals for the Section 6 morphology metrics; the same procedure should be applied to the main Crack IoU, F1, Precision, and Recall comparisons, with dataset-level and macro-averaged intervals.","section":"Tables 1–2, §5.1"},{"comment":"Table 3 shows that direct Response decoding (without Sobel calibration) has essentially the same cross-source mean IoU as SERD (60.15 vs. 60.23), while Table 1 omits Response from the external source-to-target comparison. The Sobel prior's benefit is presented through Precision and Boundary F1, but Tables 4–5 cover only three datasets. Since the method is named and motivated by semantic-edge calibration, include Response in Table 1 or add a cross-domain ablation on the same five target domains, and explicitly state that the headline IoU transfer is attributable to the response interface, not to the Sobel prior.","section":"Tables 1 and 3, §3.4"}],"minor_comments":[{"comment":"The Boundary F1 value is given as 67.18% in the Abstract and Section 5.1 but 67.17% in the Conclusion. Please harmonize.","section":"Abstract/Conclusion"},{"comment":"The paragraph beginning 'The default decoder contains only...' is duplicated. Remove one instance.","section":"§5.6"},{"comment":"Some rows are difficult to parse because values are run together, e.g., '0.458.4067.83' in the Response row and '0.361.76' in the SERD row. Ensure proper column separation in the final version.","section":"Table 3"},{"comment":"Eq. (12) introduces α=1, but α is never used afterward. Either use α in the expression or remove it.","section":"§5.6, Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"This is a competent and readable empirical study with a clear mechanism and a credible evaluation design. The main issue is not the idea but the strength of the evidence: significance/uncertainty reporting for the headline numbers and a more direct test of threshold transfer across target domains are needed. I would not reject the paper; the required additions are feasible within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Solid, honest empirical paper. The core idea—decode SAM3's internal language-conditioned response instead of its mask proposals for thin cracks—is simple and the evaluation is thorough. The headline gain of ~3.7 Crack IoU points over native SAM3 on five unseen datasets is plausible and consistent across source rotations. The strongest part is the rotated-source protocol, which shows the result doesn't hinge on one calibration domain, and the explicit limitations section.\n\nWhat's actually new: applying response decoding to SAM3's internal response for thin, fragmented defects. That's not a paradigm shift—MaskCLIP, NACLIP, CoCo-SAM3 already do response decoding and calibration—but the concrete rule (per-image min-max normalization, fixed Sobel prior, single threshold) is a new interface, and the multi-dataset evaluation with boundary, centerline, area-bias, and latency metrics is more careful than most papers in this space. The negative result on extra decoding modules (weak-response branch, topology filtering) is also useful.\n\nSoft spots, in proportion. The stress-test note is on target about Eq. 4: per-image normalization turns tau into a relative quantile, not a calibrated confidence. The rotated-source experiment shows the mean over target domains is stable, but not that tau is near-optimal per target. More tellingly, Table 3 shows plain Response (no Sobel) matches SERD's IoU (60.15 vs 60.23); the Sobel prior improves precision and boundary scores but isn't what drives the IoU transfer. The paper's 'semantic-edge' framing overstates the edge component. The main IoU numbers are point estimates without confidence intervals or significance tests; bootstrap CIs appear only for the morphology metrics. CoCo-SAM3, the closest related method, is cited but not compared. And the code link lacks a commit hash, so independent verification is light.\n\nNone of these are fatal. The central claim holds: response decoding recovers crack evidence lost in proposal formation, and it does so without training across six datasets. The authors already flag the threshold-adaptivity limitation in Section 7. The paper deserves a serious referee. Main asks: significance tests or CIs on the IoU differences, a direct comparison to CoCo-SAM3 if feasible, and a more precise statement about what Sobel contributes versus the response interface alone. Practitioners in infrastructure inspection and researchers interested in SAM's internal representations will both get value.","headline":"Solid, honest empirical study: decoding SAM3's internal response instead of its proposal masks gains ~3.7 Crack IoU points across unseen datasets; the main flaw is that per-image normalization makes the threshold a relative quantile, so the 'training-free transfer' claim is narrower than the abstract suggests.","tokens_in":17452,"tokens_out":3079,"would_cite":false,"duration_ms":33215,"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":"By decoding SAM3's internal text-conditioned response instead of its mask proposals, SERD raises cross-domain crack segmentation IoU from 54.33% to 58.00% with no training and no target-domain labels.","keywords":["crack segmentation","SAM3","training-free","cross-domain","semantic response decoding","Sobel edge prior","infrastructure inspection","zero-shot segmentation"],"falsifier":"Take a held-out domain with low-contrast or night-time crack images, select SERD's optimal threshold on a source split, freeze it, and measure Crack IoU on the held-out domain. If the threshold sweep on the new domain has no plateau near τ=0.45, or if per-domain optimal thresholds vary widely, then the single-operating-point transfer claim fails.","tokens_in":16524,"feed_emoji":"🔍","tokens_out":3502,"duration_ms":39744,"temperature":0.7,"pith_summary":"The paper argues that SAM3's failure to segment thin cracks across domains is largely an output-interface problem: the model's final region proposals suppress, truncate, or distort weak crack evidence that still exists in its internal, text-conditioned pixel response. SERD replaces the proposal interface with that dense response, normalizes it per image, modulates it with a fixed Sobel edge prior, and thresholds it with one global value. With SAM3 frozen and no target-domain annotation, SERD raises average Crack IoU on five unseen datasets from 54.33% to 58.00% compared with native SAM3, and it also improves boundary fidelity, precision, and foreground-area control. The central message is that a foundation model already contains usable crack evidence, and how that evidence is exposed to the downstream task can matter as much as training a better decoder.","feed_headline":"Read SAM3's hidden map, not its masks, to find cracks","feed_subtitle":"A fixed threshold on the model's internal crack response lifts cross-domain IoU from 54.3% to 58.0% with zero training.","key_machinery":"The key object is the internal language-conditioned response R_q: the dense logit field SAM3 generates for the concept 'crack' before it forms discrete proposals. SERD's decoding rule is R_e = Norm(R_q ⊙ (1 + E)), where E is a fixed normalized Sobel gradient magnitude of the grayscale input. The Sobel term acts as a modulator, not an independent detector—it only reweights existing semantic activation, so locations with negligible R_q cannot be recovered from image edges alone. This replaces proposal scoring, filtering, and merging with a single deterministic pixel-level thresholding step, keeping SAM3 fully frozen.","core_discovery":"The paper's central claim is that SAM3's internal language-conditioned response R_q—a dense activation field produced for the concept prompt 'crack' before proposal formation—preserves thin, weak, and fragmented cracks that are lost when the model converts that response into scored region proposals. SERD extracts R_q, applies a sigmoid, resizes it, normalizes it per image, modulates it with a normalized Sobel gradient magnitude, and applies a single global threshold τ=0.45. Under this deterministic decoder, average Crack IoU on five unseen test sets reaches 58.00% versus 54.33% for native SAM3; rotating all six datasets through source calibration gives 60.23% mean target-domain IoU and 67.18","pith_inferences":["The method's ceiling is set by SAM3's representation: because SERD only reweights existing evidence, domains where cracks produce almost no concept-conditioned activation cannot be recovered by any decoder choice—a limit the paper acknowledges.","The per-image min-max normalization means the transfer claim rests on one fixed threshold surviving distribution shifts; an adaptive threshold based on response statistics is a natural, testable extension and is one the paper lists as future work.","The same response-decoding interface could plausibly extend to other SAM versions or to other thin, line-like defects such as wires, scratches, or vascular structures, provided the internal response retains those patterns—this is a conjecture beyond the paper's crack-only evidence.","Because structural calibration only reweights semantic activation, its value is conditional on the semantic response being spatially noisy but locally meaningful; a stronger prior or learned calibration might be needed on highly textured surfaces."],"forward_implications":["Any SAM-family model with a dense text-conditioned response could be repurposed for thin-structure segmentation without training, auxiliary prompting networks, or test-time optimization.","Cross-domain crack segmentation can be deployed with one scalar threshold chosen on a single source dataset, with no target-domain annotation needed for calibration.","Masks decoded from the internal response are geometrically closer to ground truth: in the paper's macro-average, centerline-length error drops 33.2%, width-distribution error 22.3%, and component-length-distribution error 28.2% relative to native SAM3.","Improved boundary and precision behavior makes the decoded masks more usable for downstream image-domain inspection tasks such as length or width estimation, though physical calibration remains a separate step.","The decoder adds no meaningful end-to-end latency: SERD averages 550.6 ms per image versus 572.1 ms for native SAM3 in the reported implementation."],"fun_headline_variants":["Skip SAM3's masks: decode its inner response for cracks","SAM3's internal crack map beats its proposals, no training","Training-free trick: read SAM3's semantic edge response for cracks","Decode SAM3's hidden signal to segment cracks across domains"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that one fixed threshold applied to per-image normalized response values transfers across images, materials, and imaging conditions; if the response distribution shifts between domains, the same τ will systematically over- or under-segment.","fun_headline_variants_meta":{"raw":{"variants":["Skip SAM3's masks: decode its inner response for cracks","SAM3's internal crack map beats its proposals, no training","Training-free trick: read SAM3's semantic edge response for cracks","Decode SAM3's hidden signal to segment cracks across domains"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1304,"prompt_tokens":853,"completion_tokens":451,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":379}},"tokens_in":597,"tokens_out":451,"duration_ms":5390,"temperature":1.0,"reasoning_tokens":379,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:34:39.817982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out domain with low-contrast or night-time crack images, select SERD's optimal threshold on a source split, freeze it, and measure Crack IoU on the held-out domain. If the threshold sweep on the new domain has no plateau near τ=0.45, or if per-domain optimal thresholds vary widely, then the single-operating-point transfer claim fails.","supporting_citations":[],"review_version":2}