{"id":"0896c752-8fd6-4add-bcdf-2efcf4b98293","arxiv_id":"2607.07169","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":5,"one_line_summary":"TACoS achieves over 96% of fully supervised segmentation performance on 2D material flakes using less than 0.6% annotated pixels via a unified framework of consistency learning, tree energy regularization, and asymmetric contrastive learning.","lead":"TACoS is a new method that segments 2D material flakes in microscope images using scribble annotations instead of full pixel-level masks, achieving over 96% of fully supervised performance with less than 0.6% labeled pixels. A smart generalist might read it to understand how weakly supervised learning can drastically cut annotation costs in materials science.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"ARCL boundary loss uses full mask annotations (§3.2.3), contradicting the scribble-supervised premise and inflating the <0.6% annotation claim","rationale":"The reader's verdict of CONDITIONAL is appropriate, but the specific load-bearing concern should be shifted from pseudo-label reliability to the ARCL full-mask dependency. The reader correctly identified the issue (point 3) but underweighted it. The pseudo-label concern is valid in principle but is partially mitigated by the fact that ARCL already uses full-mask boundaries for direct supervision — meaning the method is not purely scribble-supervised, and the pseudo-label bias problem is less severe than it would be in a truly scribble-only setting. The ARCL full-mask issue is more fundamental because it directly undermines the paper's quantitative headline: if you count the full-mask boundary pixels used in L_Boundary, the effective annotation cost is higher than 0.6%, and the 96% figure is achieved with more supervision than claimed. The ablation in Table 4 shows ARCL contributes meaningfully (without ARCL: 82.40% graphene; with ARCL: 84.09%), so removing full-mask supervision from ARCL would likely reduce performance below the 96% threshold. The per-material hyperparameter tuning (different λ values for graphene vs MoS2, §4.6) and lack of error bars are additional concerns but are secondary to this structural issue. The paper's engineering quality and the UWSD/TER modules appear sound and independently valuable; the concern is narrowly about whether the headline claim accurately reflects the annotation regime used.","tokens_in":18731,"tokens_out":1638,"duration_ms":101399,"concrete_test":"Re-run the full TACoS pipeline on the Yan graphene and MoS2 datasets with ARCL's boundary set B derived from scribble annotations only (e.g., morphological dilation of scribble strokes) instead of full masks. If mIoU drops by more than 1.5% on either dataset, the 96% claim is materially dependent on full-mask supervision and the headline annotation cost claim is misleading. Also report what fraction of pixels in E = B ∩ S overlap with true boundary pixels from full masks, to quantify how much boundary information leaks through this channel.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is achieving 96% of fully supervised performance using less than 0.6% annotated data. However, Section 3.2.3 explicitly states: 'In the implementation, a full mask is uniformly applied to extract boundaries. The morphological gradient based on the label map y_B yields the geometric boundary set B.' The boundary loss L_Boundary (Eq. 15) operates on this set E = B ∩ S, and the weight w(i) in Eq. 16 uses cross-entropy against y_B(i) — the full-mask boundary label. This means the ARCL module, which contributes +1.69% mIoU on graphene and +2.27% on MoS2 over UWSD+TER alone (Table 4), directly leverages dense pixel-level annotations during training. The paper's scribble annotations are automatically generated from existing full masks (§4.1), so full masks are available but should not be used as supervision if the method is genuinely scribble-supervised. The <0.6% annotation figure counts only scribble pixels but omits the full-mask boundary supervision in ARCL. This is not a clarification issue — it is a direct contradiction of the paper's core premise. The reader identified this as point (3) but framed it as needing clarification rather than as the load-bearing concern. The pseudo-label reliability concern (the reader's weakest_assumption) is real but secondary: if the method already uses full masks for boundaries, pseudo-label bias is partially masked by direct boundary supervision.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"The manuscript proposes TACoS, a single-stage scribble-supervised semantic segmentation framework for two-dimensional (2D) material flakes in optical microscopic images. The method integrates three components: (1) Unlabeled Weak-Strong Distribution alignment (UWSD), which enforces cosine consistency between weakly and strongly augmented views on unlabeled pixels; (2) Tree Energy Regularization (TER), which constructs minimum spanning trees on shallow and deep DINOv2 features to generate structure-aware soft pseudo labels; and (3) Asymmetric Regional Contrastive Learning (ARCL), which performs region-level prototype contrast on hard samples and applies an asymmetric boundary penalty. Experiments on graphene, MoS2, WSe2, and MoS2 datasets from two independent sources (Yan and Uslu) show that TACoS narrows the gap to fully supervised performance under sparse scribble annotations, with ablation studies isolating the contribution of each module.","tokens_in":19621,"tokens_out":1293,"duration_ms":214207,"significance":"The paper addresses a genuine practical bottleneck: pixel-level annotation of 2D material micrographs is expert-dependent and costly, limiting scalability of automated flake identification. The application of scribble supervision to this domain is novel, and the experimental design—two independent datasets, unified DINOv2 backbone for all baselines, boundary IoU as a supplementary metric, and module-level ablations—is reasonable. The framework is end-to-end trainable without iterative pseudo-label propagation, which is a practical advantage. However, the central quantitative claim of achieving over 96% of fully supervised performance using less than 0.6% annotated data is compromised by a supervision inconsistency in the ARCL module (see Major Comment 1), which must be resolved before the contribution can be properly assessed.","major_comments":[{"comment":"§3.2.3, Eqs. (15)–(16): The ARCL boundary loss L_Boundary operates on the geometric boundary set B, which is extracted from a full mask y_B. The text states: 'In the implementation, a full mask is uniformly applied to extract boundaries.' The weight w(i) in Eq. (16) is computed via cross-entropy against y_B(i), the full-mask boundary label. This means the ARCL module—which contributes +1.69% mIoU on graphene and +2.27% on MoS2 over UWSD+TER alone (Table 4)—directly uses dense pixel-level annotations during training. This contradicts the paper's core premise of scribble supervision and the claim of using less than 0.6% annotated data. The scribble annotations are generated from existing full masks (§4.1), so full masks are available, but using them for boundary supervision means the method is not purely scribble-supervised. The authors must either (a) replace the full-mask boundary labels","section":null},{"comment":"Tables 2–4: No error bars, confidence intervals, or significance tests are reported for any experiment. Given that some improvements over baselines are modest (e.g., +1.74% over URSS on graphene, +1.91% over CC4S on MoS2), it is unclear whether these gains are statistically significant or within run-to-run variance. At minimum, the authors should report results over multiple random seeds (≥3) with mean and standard deviation for the main comparison table and the ablation study.","section":null}],"minor_comments":[{"comment":"Abstract and §4.1: The phrase 'less than 0.6% annotated data' is used prominently but the actual scribble coverage varies from 0.34% to 0.56% depending on dataset (Table 1). The abstract should specify which dataset this figure corresponds to, or use a range.","section":null},{"comment":"§3.2.3, Eq. (11): The augmented label construction uses a fixed threshold τ_w=0.7. No sensitivity analysis for this threshold is provided. Given its role in pseudo-label quality, a brief study would strengthen the paper.","section":null},{"comment":"§4.2: The training uses a single-GPU batch size of 2 across two GPUs (effective batch size 4). This is small for DINOv2 fine-tuning. The authors should comment on whether this affects convergence or stability.","section":null},{"comment":"Figure 2: The overview diagram is dense and the text describing data flow between UWSD, TER, and ARCL is complex. A clearer separation of the three modules in the figure, with explicit arrow labels, would aid readability.","section":null},{"comment":"§4.5, Figure 6: The failure case analysis is qualitative only. Quantifying the correlation between optical contrast ratio and misclassification rate would make this section more rigorous.","section":null},{"comment":"§2.2: The related work discusses limitations of TEL's minimum spanning tree under low discriminative power of low-level features, but TER uses a similar MST construction. The authors should clarify how TER's dual-tree (shallow+deep) design specifically addresses the limitation they identify in TEL.","section":null},{"comment":"Table 2: The 'Multi-stage' column is marked with ✓ for some methods but the criteria for this designation are not stated. A footnote defining what constitutes multi-stage would help.","section":null},{"comment":"§3.2.4, Eq. (18): The loss weights λ_c=5, λ_t=1, λ_r=0.002, λ_b=0.00025 span several orders of magnitude. The justification given (inverse scaling between global and local gradients) is plausible but informal. A brief note on how these were tuned and whether they transfer to new materials would help.","section":null}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about ARCL using full-mask boundary labels is well-founded and is the primary reason for the major revision recommendation. This is not a presentation issue—it directly affects the validity of the central claim. If the authors can show that the boundary loss works equally well with scribble-derived or pseudo-label-derived boundaries, the paper's contribution stands. If not, the framing of the method as purely scribble-supervised needs substantial revision, and the <0.6% annotation claim must be qualified. The pseudo-label reliability concern raised by the reader is valid but secondary; it becomes moot if the boundary supervision issue is resolved by removing full-mask dependency."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"Here's the short version: TACoS applies scribble-supervised segmentation to 2D material microscopy for the first time, combining weak-strong consistency (UniMatch V2 style), tree energy regularization (from TEL), and asymmetric regional contrastive learning into a single-stage framework. The engineering is competent, the ablations are clean, and the practical motivation — replacing dense masks with scribbles for high-throughput flake screening — is real. But there's a problem the paper doesn't address that may undermine its central claim, and I think it's more serious than the reader flagged it as a clarification item. The stress-test note is right to call it out as load-bearing. In Section 3.2.3, the ARCL boundary loss L_Boundary (Eq. 15) operates on a boundary set E = B ∩ S, where B is extracted via morphological gradient on a label map y_B. The text explicitly says 'a full mask is uniformly applied to extract boundaries.' The weight w(i) in Eq. 16 computes cross-entropy against y_B(i), the full-mask boundary label. So ARCL — which contributes +1.69% mIoU on graphene and +2.27% on MoS2 over UWSD+TER alone (Table 4) — is directly supervised by dense pixel-level annotations during training. The paper's scribbles are auto-generated from existing full masks (Section 4.1), so full masks are available, but using them for boundary supervision contradicts the premise that the method is scribble-supervised with <0.6% annotated data. That figure counts only scribble pixels and omits the full-mask boundary supervision. This isn't ambiguous wording; it's a direct contradiction of the core claim. The pseudo-label reliability concern the reader raised is real but secondary — if full masks are already in use for boundaries, pseudo-label bias is partially masked by direct supervision. What the paper does well: fair backbone comparison (DINOv2 for all methods), two independent datasets, boundary IoU as a supplementary metric, and a clean ablation isolating each module's contribution. The single-stage design avoiding iterative pseudo-label propagation is a genuine practical advantage. The per-material hyperparameter tuning (different λ values for graphene vs MoS2) is a minor concern — it risks overfitting to validation sets. No error bars or multi-run statistics are reported, which matters less given the large datasets but should still be addressed. This is a legitimate new application with solid engineering, but the full-mask boundary supervision issue needs to be resolved before the central claim holds. If the authors can show ARCL works without full-mask boundaries — or reframe the claim honestly — this is a paper worth a serious referee. As it stands, I'd flag this as the key revision item.","headline":"Scribble-supervised method for 2D material segmentation — solid engineering, but ARCL boundary loss appears to use full masks, contradicting the <0.6% annotation claim","tokens_in":19551,"tokens_out":673,"would_cite":false,"duration_ms":171346,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Scribble annotations recover 96% of full-mask segmentation for 2D materials","keywords":["scribble supervision","semantic segmentation","two-dimensional materials","weakly supervised learning","contrastive learning","optical microscopy","minimum spanning tree","consistency regularization"],"falsifier":"If TACoS is trained on a 2D material system where optical contrast between flake and substrate is substantially lower than in the tested graphene and MoS2 samples, the weak-branch pseudo-labels near boundaries would be more frequently wrong, and the 96%-of-full-supervision claim would likely break down.","tokens_in":18982,"feed_emoji":"","tokens_out":1005,"duration_ms":224305,"temperature":0.7,"pith_summary":"The paper introduces TACoS, a single-stage framework for pixel-level segmentation of two-dimensional material flakes from optical micrographs using only scribble annotations covering less than 0.6% of image pixels. The approach couples three mechanisms: weak-strong consistency alignment to propagate predictions across unlabeled regions, tree-energy regularization built from minimum spanning trees on backbone features to enforce structural coherence, and asymmetric regional contrastive learning to sharpen boundary discrimination in low-contrast edges. On graphene and MoS2 datasets, the method reaches over 96% of fully supervised mIoU while maintaining superior boundary IoU compared to existing weakly supervised baselines, including multi-stage approaches. The central claim is that these three components working jointly in a single training process can recover nearly all the performance of dense pixel-level annotation at a fraction of the labeling cost, specifically for the domain of 2D material microscopy where boundaries are governed by thin-film interference contrast and are often indistinct against complex substrates.","feed_headline":"","feed_subtitle":"","key_machinery":"TACoS combines: (1) Unlabeled Weak-Strong Distribution Alignment (UWSD), which uses cosine distance between weak and strong augmentation branch predictions on unlabeled pixels with stop-gradient on the weak branch; (2) Tree Energy Regularization (TER), which constructs minimum spanning trees on both shallow (Layer 2) and deep (Layer 11) DINOv2 features, computes pixel affinity from tree path distances, and generates structure-aware soft pseudo-labels via tree filtering; (3) Asymmetric Regional Contrastive Learning (ARCL), which fuses weak-branch high-confidence predictions with scribbles into augmented labels, mines hard samples where strong-branch confidence drops below threshold, computes类","core_discovery":"Three jointly optimized constraints, consistency alignment on unlabeled pixels, minimum-spanning-tree-based soft pseudo-labels, and asymmetric contrastive learning on boundary regions, together recover over 96% of fully supervised segmentation performance from under 0.6% scribble annotation coverage on 2D material micrographs. The tree-energy module captures pixel affinity from both shallow color-gradient features and deep semantic features, generating online soft references that suppress foreground fragmentation. The contrastive module deliberately targets the hardest boundary pixels, those where augmented labels indicate foreground but strong-augmentation confidence collapses, applying rep","pith_inferences":["The reliance on weak-branch pseudo-labels with a fixed confidence threshold of 0.7 creates a potential failure mode where systematically biased predictions near low-contrast boundaries are reinforced rather than corrected, since all three modules treat these pseudo-labels as ground truth. A confidence-adaptive or uncertainty-aware threshold might be necessary for materials with even lower optical ","The single-stage design avoids multi-stage pseudo-label propagation but may sacrifice the error-correction opportunities that iterative refinement provides. Whether the tree-energy regularization can fully compensate for this in domains with more severe class imbalance remains an open question.","The method's boundary IoU, while improved over baselines, still lags substantially behind the mIoU gap to full supervision, suggesting that boundary precision remains the primary bottleneck even with the contrastive learning module."],"forward_implications":["Materials science labs could adopt automated flake screening without expert pixel-level annotation, reducing the bottleneck between sample exfoliation and characterization.","The tree-energy plus contrastive boundary design may transfer to other microscopy domains where object boundaries are governed by physics-based contrast rather than sharp edges, such as biological thin sections or crystallographic imaging.","If the 96%-of-full-supervision threshold holds across additional 2D material systems beyond graphene and MoS2, the scribble paradigm could become the default annotation standard for materials segmentation datasets.","The asymmetric boundary sampling strategy could be adapted for active learning, using the hard-sample set to suggest where additional scribble strokes would most improve the model."],"fun_headline_variants":["Scribble annotations recover 96% of full-supervision segmentation on 2D materials","Under 0.6% labels reach 96% of fully supervised 2D-material segmentation","TACoS matches full supervision on 2D materials with under 0.6% scribble labels","Scribble-level supervision hits 96% of dense-label performance on 2D-material flakes","Tree-energy and contrastive constraints recover 96% of full supervision from scribbles"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The weak-augmentation branch predictions used as pseudo-labels are assumed reliable enough to guide the strong branch, but if those predictions are systematically biased near low-contrast boundaries, the consistency, tree regularization, and contrastive modules could propagate that bias rather than correct it.","fun_headline_variants_meta":{"raw":{"variants":["Scribble annotations recover 96% of full-supervision segmentation on 2D materials","Under 0.6% labels reach 96% of fully supervised 2D-material segmentation","TACoS matches full supervision on 2D materials with under 0.6% scribble labels","Scribble-level supervision hits 96% of dense-label performance on 2D-material flakes","Tree-energy and contrastive constraints recover 96% of full supervision from scribbles","Weak scribble labels rival dense annotation for 2D-material flake segmentation"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1273,"prompt_tokens":557,"completion_tokens":716,"prompt_tokens_details":null},"tokens_in":557,"tokens_out":716,"duration_ms":48840,"temperature":1.0,"reasoning_tokens":548,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T18:22:15.899220+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If TACoS is trained on a 2D material system where optical contrast between flake and substrate is substantially lower than in the tested graphene and MoS2 samples, the weak-branch pseudo-labels near boundaries would be more frequently wrong, and the 96%-of-full-supervision claim would likely break down.","supporting_citations":[],"review_version":1}