SEAM in Research: Methods & Applications

• SEAM is a multifaceted term that denotes structured boundaries in imaging and diverse acronymic evaluation and control methods in fields like ML, robotics, and algebra.
• In imaging and graphics, seams serve as optimization objects that enhance visual continuity through techniques like seam carving, dynamic programming, and deep learning.
• Across robotics, smart contracts, and formal theory, SEAM provides structural cues to boost model robustness, control precision, and compositional consistency.

to=arxiv_search.search ￣奇米影视json {"query":"ti:SEAM OR ti:seam OR abs:SEAM", "max_results": 10, "sort_by":"submittedDate", "sort_order":"descending"}【อ่านข้อความเต็มanalysis to=arxiv_search.search 񎔊ppjson code omitted? to=arxiv_search.search 天天中彩票公众号json {"query":"(Ahmadi et al., 2019) OR (Vsevolod et al., 5 Jun 2026) OR (Lior et al., 2024) OR (Li et al., 22 Jun 2025) OR (Tang et al., 25 Aug 2025) OR (Hossain et al., 2024) OR (Zhan et al., 6 Jul 2026)", "max_results": 10, "sort_by":"relevance", "sort_order":"descending"} SEAM is a recurrent designation in contemporary research literature, used both as a literal technical term—most notably for paths, cuts, or boundaries in visual computing—and as an acronym for several unrelated frameworks, benchmarks, and defenses across machine learning, robotics, software systems, and statistics. In the arXiv literature, SEAM denotes, among other things, Shortcut-aware Evaluation, Augmentation, and Modeling for scriptedness detection, Semantically Equivalent Across Modalities for vision-language benchmarking, a Stochastic Evaluation Approach for Multi-document tasks, Smooth Execution of Action-Chunked Motion for VLA policies, Secure Automated and Maintainable Smart Contract Upgrade Framework, and Self-supervised Equivariant Attention Mechanism for weakly supervised segmentation; related seam-centric work studies image retargeting, image stitching, mesh cutting, garment manipulation, and boundary seam algebras (Vsevolod et al., 5 Jun 2026, Tang et al., 25 Aug 2025, Lior et al., 2024, Zhan et al., 6 Jul 2026, Hossain et al., 2024, Wang et al., 2020, Ahmadi et al., 2019, Langlois-Rémillard et al., 2019).

1. Terminological scope and domain-specific meanings

In visual computing, a seam is typically a structured boundary. In image retargeting, it is a connected path of pixels removed or inserted during seam carving; in image stitching, it is the dividing curve inside the overlap between aligned images; in mesh processing, it is a cut set on a surface; in garment manipulation, it is a stitched structural line that encodes assembly and part information (Ahmadi et al., 2019, Li et al., 2017, Li et al., 22 Jun 2025, Huang et al., 13 Jun 2026). These usages are literal rather than acronymic.

By contrast, many papers use SEAM as an acronym. The acronym is expanded as Shortcut-aware Evaluation, Augmentation, and Modeling in real-time scriptedness detection for interview guardrails (Vsevolod et al., 5 Jun 2026), Semantically Equivalent Across Modalities in a benchmark for vision-LLMs (Tang et al., 25 Aug 2025), and a Stochastic Evaluation Approach for Multi-document tasks in LLM benchmarking (Lior et al., 2024). Other expansions include Smooth Execution of Action-Chunked Motion for chunked robot policies (Zhan et al., 6 Jul 2026), Secure Automated and Maintainable Smart Contract Upgrade Framework for Solidity upgrades (Hossain et al., 2024), Self-supervised Equivariant Attention Mechanism for weakly supervised semantic segmentation (Wang et al., 2020), and synthetic estimated average matchup for batter-versus-pitcher spray chart estimation (Wapner et al., 2020).

A mathematical use appears in the representation theory of boundary seam algebras $\mathsf b_{n,k}(\beta=q+q^{-1})$, introduced to formulate algebraically a large class of boundary conditions for two-dimensional statistical loop models (Langlois-Rémillard et al., 2019). Here, “seam” refers to a boundary modification in Temperley–Lieb-type diagrammatics rather than to an acronym.

2. Seams as optimization objects in imaging and video

In content-aware image retargeting, the 2019 seam-carving paper augments forward-energy seam carving by controlling the positional distribution of seams so that repeated seam removals do not cluster in the same spatial corridor (Ahmadi et al., 2019). Its additional deterministic penalty is

$$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$

and is added to the forward-energy term $+LR$. The method is evaluated in MATLAB on CUHK with 57 images and RetargetMe with 80 images, using fixed parameters $a=0.0625$, $R1=4$, and $R2=5$. The reported outcome is lower Seam Coagulation Measure and higher Aspect Ratio Similarity than the improved seam-carving baseline, indicating less seam coagulation and better retargeted-image quality (Ahmadi et al., 2019).

A related algorithmic analysis compares brute-force, greedy, dynamic programming, and GPU-based parallel seam carving, formalizing the vertical-seam recurrence

$$M(i,j)=e(i,j)+\min\{M(i-1,j-1),\,M(i-1,j),\,M(i-1,j+1)\},$$

and emphasizing the large runtime gap between exact enumeration, heuristic local choice, standard DP, and GPU-parallel DP (Aijaz et al., 2024). For video retargeting, SCPL introduces parental labeling, a dynamic spatiotemporal energy buffer, and a motion-aware energy function, reporting a 90.6% average runtime reduction and a best reported 93.7% reduction relative to raw framewise seam carving, together with higher frame-wise consistency (Chuning, 2019).

In image stitching, seam-cutting is the decision boundary inside the overlap between aligned images. A perception-based energy formulation replaces raw Euclidean discrepancy with a sigmoid-metric color difference and saliency weighting so that seam optimization better tracks human visibility (Li et al., 2017). Subsequent work iteratively refines seams by patch-point evaluation and local reweighting of the overlap difference map to obtain a nearly perception-consistent seam (Liao et al., 2018), and LPAM further treats poor seam neighborhoods as local alignment problems, using SSIM-based seam quality to trigger local patch alignment under large parallax (Liao et al., 2023). Deep learning variants such as DSeam convert seam prediction into mask prediction and report about 170 FPS at input size $256\times256$, about 15× faster than GraphCut in OpenCV 2.4.9, with similar seam quality (Cheng et al., 2023). Across this literature, a seam is the locus where geometric mismatch, local appearance continuity, and perceptual visibility intersect.

3. Seams as structural primitives in geometry, garments, and robotics

In computer graphics, the seam becomes a surface cut rather than an image path. “Auto-Regressive Surface Cutting” formulates seam generation on 3D meshes as next-token prediction, with a GPT-style decoder generating sequences of seam segments represented by quantized 3D coordinates (Li et al., 22 Jun 2025). The model conditions on a point cloud sampled specifically from mesh vertices and edges—61,440 points total, split evenly into 30,720 vertex samples and 30,720 edge samples—uses 10-bit coordinates ($2^{10}=1024$ bins), and employs an hourglass-style hierarchical transformer with depth configuration $(2,(4,12,4),2)$, hidden dimension 1536, and maximum positional length 36,864 (Li et al., 22 Jun 2025). The stated objective is seam layouts that are semantically coherent and structurally aligned for UV parameterization, texture mapping, and mesh decomposition.

In robotics, seams are treated as privileged structural observations. “Seam-to-Graph Reconstruction for Garment Configuration Alignment” maps partial 3D seam observations to a topology-encoded structural skeleton graph for garment state estimation and visual servoing (Huang et al., 13 Jun 2026). The system trains on 99,000 synthetic samples, tests on 11,000 unseen synthetic samples, and runs the Seam-to-Graph network at 9.9 Hz. Its ablation shows that removing the seam branch hurts performance markedly, and the paper interprets seams as providing “critical structural information and absolute anchoring,” with point clouds supplying complementary geometric constraints (Huang et al., 13 Jun 2026). This seam-centered state representation is then used for bimanual garment alignment with human-level alignment accuracy and reduced variance in alignment error (Huang et al., 13 Jun 2026).

A closely related manipulation strategy is SIS, the Seam-Informed Strategy for T-shirt unfolding (Huang et al., 2024). SIS uses a Seam Feature Extraction Method to detect seam line segments and seam crossing segments, and a Decision Matrix Iteration Method to choose dual-arm grasp pairs from seam-derived candidates. The method is trained on real data without simulation, uses seam-type combinations as the action abstraction, and reports over 0.85 average IoU within three steps; at a threshold of 0.85 for both normalized coverage and IoU, the success rates over 20 trials are 20%, 85%, 90%, 95%, and 95% from steps 1 to 5 (Huang et al., 2024). In these robotics papers, seams are neither incidental texture nor low-level edges; they are the garment’s most informative control-oriented visual signal.

4. SEAM as benchmark and evaluation methodology

One major acronymic lineage uses SEAM for evaluation. In LLM assessment, SEAM is a Stochastic Evaluation Approach for Multi-document tasks, a conglomerate benchmark over 6 existing datasets and 3 task families: multi-document summarization, multi-hop question answering, and cross-document coreference resolution (Lior et al., 2024). Its distinctive move is repeated stochastic resampling over arbitrary factors such as instruction paraphrase, document order, few-shot example identity, few-shot order, and evaluation instances, with 10 i.i.d. runs, $$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$0 demonstrations, and up to 100 evaluated instances per dataset per run (Lior et al., 2024). It summarizes robustness with

$$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$1

where $$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$2. The benchmark’s main empirical claim is that multi-document tasks remain difficult even for 70B models, no single model dominates all datasets, and summarization is more robust than structured-output QA and coreference (Lior et al., 2024).

In multimodal evaluation, SEAM: Semantically Equivalent Across Modalities defines a benchmark pairing semantically equivalent inputs across four domains that have existing standardized textual and visual notations (Tang et al., 25 Aug 2025). The benchmark evaluates whether VLMs reason consistently across textual-symbolic and visual-spatial representations rather than merely solving different tasks with different information. Across 21 contemporary models, the reported findings are a systematic modality imbalance—vision frequently lags language in overall performance despite semantically equivalent content—and relatively low cross-modal agreement (Tang et al., 25 Aug 2025). The abstract-level error analysis attributes this to textual perception failures from tokenization in domain notation and visual perception failures that induce hallucinations, and further states that the results are largely robust to visual transformations (Tang et al., 25 Aug 2025).

A statistical usage appears in baseball analytics. SEAM, expanded as synthetic estimated average matchup, estimates batter-versus-pitcher spray chart distributions by combining the direct empirical matchup with a synthetic pitcher and a synthetic batter (Wapner et al., 2020). Its final estimator is

$$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$3

with weights

$$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$4

The paper states that the associated Shiny application can visualize any matchup in the Statcast era beginning in 2017 almost instantly and is intended for defensive alignments, lineup construction, or pitcher selection (Wapner et al., 2020).

5. SEAM as alignment, safety, and control mechanism in machine learning

In weakly supervised semantic segmentation, SEAM stands for Self-supervised Equivariant Attention Mechanism (Wang et al., 2020). The method begins from CAM-based image-level supervision, adds equivariant regularization so that predictions from transformed images remain geometrically consistent, and refines CAMs with a Pixel Correlation Module. Its overall training objective is

$$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$5

On PASCAL VOC 2012, the reported final segmentation performance is 64.5 mIoU on val and 65.7 on test with a ResNet38 backbone and no saliency supervision, compared with a same-backbone baseline at 59.7 val and 61.9 test (Wang et al., 2020). Here, SEAM denotes a consistency mechanism that narrows the gap between image-level labels and pixel-wise masks.

A safety-oriented use appears in speech and language modeling. SEAM: Shortcut-Aware Real-Time Detection of Scripted vs. Spontaneous Speech for Interview Guardrails frames scriptedness detection as a task vulnerable to corpus identity, channel conditions, and recording artifacts (Vsevolod et al., 5 Jun 2026). The framework combines uniform waveform preprocessing, seam-aware sampling, non-speech augmentation, and a compact DistilHuBERT backbone. With 8 s windows, it reports 0.971 ± 0.004 ROC-AUC on an external interview-domain evaluation set, and post-training quantization reduces the model footprint to 41.8MB with little loss in external performance (Vsevolod et al., 5 Jun 2026). Its central empirical claim is that removing the shortcut-prevention components improves internal held-out metrics but sharply reduces external performance, exposing shortcut learning (Vsevolod et al., 5 Jun 2026).

In robot action generation, SEAM: Smooth Execution of Action-Chunked Motion is a training-free inference-time method for flow-matching VLA policies (Zhan et al., 6 Jul 2026). It exploits the previous chunk’s unexecuted tail as an analytic consistency reference and applies Velocity-guided Loss Steering after each Euler step. On LIBERO-10 with $$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$6, the paper reports that SEAM reduces boundary jerk by 28%, reduces chunk transition discontinuity by 27%, preserves baseline-level task success, and keeps denoising-loop cost at $$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$7 the unguided baseline (Zhan et al., 6 Jul 2026). A related alignment-defense use appears in Self-Destructive LLM, where SEAM modifies optimization geometry so that harmful fine-tuning degrades general capability while preserving benign utility; the abstract attributes this to a novel loss that couples benign and harmful optimization trajectories, augmented with adversarial gradient ascent and an efficient Hessian-free gradient estimate with theoretical error bounds (Wang et al., 18 May 2025).

6. Infrastructure, formal theory, and cross-domain interpretation

In software systems, SEAM denotes a Secure Automated and Maintainable Smart Contract Upgrade Framework for transforming ordinary Solidity contracts into upgradeable ones using the diamond pattern (Hossain et al., 2024). The framework is explicitly designed to automate modularization into facets, manage deployment and upgrades, and target two vulnerabilities: function selector clashes and storage slot collisions (Hossain et al., 2024). The paper is architectural rather than empirical: it describes an AST Generator, Contract Assembly Unit, Security Reinforcer, Deployment Manager, Upgrade Validator, and Versioning and Changelog system, but states that large-scale implementation and real-world evaluation are future work (Hossain et al., 2024).

In algebra, seam appears in a more classical formal sense. The paper on the representation theory of seam algebras studies the boundary seam algebras

$$E_{\text{ramp}}(i,j)= \begin{cases} a (j \mod R1) & \text{if } (i \mod R2)=0\ 0 & \text{otherwise}, \end{cases}$$8

constructs their irreducible, standard, and principal modules, and describes their composition factors and non-split short exact sequences (Langlois-Rémillard et al., 2019). These algebras were introduced to formulate algebraically a large class of boundary conditions for two-dimensional statistical loop models (Langlois-Rémillard et al., 2019). This is a distinct usage from the acronymic ML literature, but it retains the core intuition of a seam as a structured boundary condition.

A plausible cross-domain implication is that “SEAM” and “seam” frequently mark interface-sensitive structure: the place where local decisions must remain coherent with global organization. In imaging, the seam is the boundary along which distortions or discontinuities are minimized; in mesh cutting and garments, it is the structural line that determines decomposition or control; in evaluation benchmarks, SEAM denotes controlled comparison across prompts or modalities; in safety and execution, it is the mechanism that suppresses spurious shortcuts or boundary discontinuities; in software and algebra, it is the formal boundary where compositional consistency must be preserved (Ahmadi et al., 2019, Li et al., 22 Jun 2025, Lior et al., 2024, Zhan et al., 6 Jul 2026, Hossain et al., 2024, Langlois-Rémillard et al., 2019). This suggests that the recurrence of the term is not purely lexical: across domains, SEAM often names the problem of preserving structure across a join.