PyTorchKR Papers

최신 논문

총 523,801편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.04181v1cs.AI

Clocks all the way down: does sequence-processing AI need endogenous temporal structure?

Mainstream sequence-processing AI is temporally reactive: each update is driven by input and learned state transitions, rather than by an independently evolving internal temporal reference. Biological brains, by contrast, continuously generate rhythmic activity that structures perception, memory, and cognition, and this rhythmicity reaches all the way to behavior. We ask whether this difference is merely implementational or genuinely computational. Four lines of evidence suggest that endogenous temporal structure supports putative functions that input-driven architectures may not readily replicate: 1) phase-dependent routing of information, 2) delay-structured combinatorial representation, 3) cross-frequency multiplexing of items, and 4) oscillation-gated offline consolidation. A proof-of-concept delay-ablation test dissociates representational capacity from memory retrieval. In recurrent and spatial spiking networks with fixed synaptic weights and connectivity, shuffling delays preserves part of the capacity but substantially impairs the sequence retrieval. The relative contrast is more pronounced for spatial networks and longer propagations. Instantaneous transmission collapses capacity itself. We then outline a cross-frequency multiplexing scheme that combines item identity and serial order with alternative mechanisms for selective memory retrieval. Finally, we state one falsifiable hypothesis, conditions for matched-resource tests, and three research directions: the components already exist; the disciplined comparison does not.

arXiv:2611.04180v1cs.AI

Validation-Gated ML Pipeline Synthesis with Expert-Grounded Self-Evolving Skills

Recent advances in large language models have enabled agentic systems that automate machine learning pipeline construction from minimal user input. However, existing agentic AutoML methods either depend on rigid templates that limit adaptability, target narrow task domains, or rely on free-form code generation that lacks reliability guarantees, leaving intermediate errors to propagate silently and degrade the final model. We present EvoML, an agentic framework for end-to-end ML pipeline synthesis built on expert-grounded skills that evolve across runs, executed through a validation-gated harness. The framework decomposes pipeline construction into four phases coordinated by three agents: an Orchestrator that enforces validation gates, advancing each phase only when its outputs pass structural, execution, and semantic checks, which precludes error propagation and keeps every artifact auditable; an Executor that runs expert-grounded skills, grounded in the nnU-Net workflow and pairing reusable templates with online adaptation; and a Curator that performs post-run evolution, consolidating the scripts and guidance recovered during a run into the skill library so it improves across datasets. We evaluate EvoML on eight datasets spanning six sophisticated, real-world ML task types, against agentic AutoML systems and command-line coding tools. EvoML achieves full automation and the best final model quality on all eight, spanning segmentation (up to 0.94 Dice), classification (0.99 accuracy), reconstruction (0.95 SSIM), and denoising (0.79 normalized PSNR), while transferring to broader ML tasks.

arXiv:2611.04173v1cs.CV

ReconDreamer#: Self-Evolving Closed-Loop Synergy of Reconstruction and Generation for Dynamic Driving Scenes

Dynamic driving scene reconstruction has made remarkable strides in recent years by leveraging generative priors, particularly in the quality of novel trajectory rendering. Nevertheless, a significant gap remains between the generated data and real-world sensor observations, severely hindering further improvements in reconstruction fidelity. In this paper, we present ReconDreamer#, a self-evolving framework that synergizes reconstruction and generation for dynamic driving scenes, which substantially bridges the gap between synthesized novel trajectory data and recorded real-world observations. Within this self-evolving closed-loop framework, the generative and reconstruction models mutually reinforce each other through iterative refinement, ultimately yielding a significant enhancement in the 3D representation capability of the reconstruction model. Additionally, we introduce a geometry-prior-based adaptive ground model to further enhance the rendering quality of the reconstruction model. To validate the effectiveness of the proposed method, we conduct comprehensive evaluations on Waymo, nuScenes, PandaSet, and EUVS benchmarks. Specifically on Waymo, at a 6-m lateral shift, ReconDreamer# improves NTA-IoU and NTL-IoU by 8.8% and 3.0%, respectively, and reduces FID by 15.1% relative to ReconDreamer++.

arXiv:2611.04169v1cs.CL

Length Is Cheap, Competition Is Not: Attention Predicts Context Rot in Retrieval

Why do language models fail on long inputs? The common answer is length. We show that, for retrieval, length is cheap and competition is not. On five open models (0.5B to 3B parameters, four families), neutral filler costs nothing: accuracy stays at 0.97 to 1.00, up to 32K tokens on Qwen2.5-1.5B. Statements in the format of the evidence, placed at the same positions, drop it to 0.16 to 0.33. A single statement that gives the queried key a rival value multiplies the loss by 9 to 110, and every additional rival adds to it. Inside the model, one quantity tracks the failure: the attention mass left on the evidence. Across 126,473 trials it explains up to 58% of the variance in loss; log length explains at most 8%. The masked-filler protocol behind the claim that length alone hurts reproduces on MMLU (11 to 18 points lost at 30K tokens), but fails a visible control in retrieval: masking costs 45 accuracy points at 32K tokens, while the same filler left visible costs none. In retrieval, context rot is a competition for attention. We release the probe and all 126K trial-level measurements.

arXiv:2611.04166v1cs.LG

Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs

Reasoning vision-language-action (VLA) models pair physical control with natural-language rationales that supervisors can read to monitor embodied safety. Prior attacks on VLA reasoning corrupt the rationale itself and leave visible evidence in the text. We show that visual perturbations can instead alter the physical action while keeping the generated rationale word-for-word identical. At the shared text/action interface of Alpamayo for driving and InstructVLA for manipulation, our text-neutral action-subspace (TNAS) diagnostic shows that the locally most action-sensitive direction is nearly orthogonal to the active text-loss gradients, with 1.87% and 5.28% norm overlap. At equal hidden-state perturbation norms across both models, interventions along this text-neutral residual shift actions 4.64x to 47.8x more than text-gradient perturbations, whereas text-gradient perturbations induce 21.2x to 277x more text drift. Steering visual perturbations along this direction reaches the same decoupling from the image. Across 154 obstacle-annotated driving scenes in Alpamayo, TNAS-guided attacks induce a new collision under a word-for-word identical rationale in 52.6% of scenes under full-image perturbations and 35.1% under a localized forward-camera patch, compared with 4.55% for unguided action attacks. Relative to a text-preserving baseline, TNAS guidance raises exact rationale preservation from 83.0% to 89.5% and cuts teacher-forced text drift by 68.4% at a comparable collision rate. In InstructVLA, TNAS guidance cuts text drift by 40.5% and preserves the exact response on 85.0% of examples alongside a 1.73-unit action-chunk shift. An unchanged rationale therefore cannot certify a safe action, so monitors for reasoning VLAs must verify the predicted action channel directly.

arXiv:2611.04164v1cs.CV

What Remains in Sight? Autoregressive Video Decoding as Representation-Guided Context Rewriting

Long-horizon video generation is increasingly moving toward autoregressive rollout, where each newly generated segment becomes part of the evidence used for future prediction. Recent causal decoders, long-context training, and KV-cache mechanisms have extended feasible duration, but most inference pipelines still decide the visible past implicitly through recent windows, compressed caches, or external memory. This raises a direct question: \textbf{can the choice of which past frames or segments to show to the decoder be treated as an explicit test-time decision?} We answer this question with \textsc{ReCR}, a training-free method that formulates autoregressive video decoding as representation-guided context rewriting. At each step, \textsc{ReCR} builds a selected visible context from the generated history by using internal decoder representations to favor context units that match the global rollout state, avoid redundancy with the recent continuation, preserve temporal boundary coverage, and widen the range of spatial layouts in view. The selected units are then mapped onto a compact causal axis before decoding the next block. Across long-horizon text-to-video benchmarks, \textsc{ReCR} consistently improves multiple autoregressive backbones under matched visible-token budgets. Empirical analyses such as representation-space diagnostics and transition-count prompt switching further show that improving what remains visible is an effective path toward more stable long-video decoding.

arXiv:2611.04158v1cs.LG

On the Feature Evolution of Temporal Difference Learning

Temporal-difference (TD) learning trains neural representations using bootstrapped targets generated by evolving value estimates, yet how predictive features emerge and propagate under incomplete neural updates remains poorly understood. We investigate this problem using finite-width two-layer ReLU critics on a constructed family of smooth stochastic horizon-$H$ MDPs with unknown quadratic structure and $d$-dimensional observations. We first show that an informative TD update acquires the quadratic Bellman signal and yields a representation admitting a bounded predictive readout. The simultaneous hidden-weight and head update achieves a strict finite-sample prediction-risk improvement over matched head-only SGD, with a $2(d+1)$-fold larger population quadratic response. We then establish that an imperfect learned continuation can already provide informative supervision to an earlier zero-reward stage. By tracking coupled neural updates with one SGD step per target copy, without requiring intermediate Bellman regressions to be solved, we obtain a uniform prediction-error transient $\exp[-\Omega(K/(Hd))]$, up to action-independent shifts and explicit sampling and finite-width errors. On the constructed signed-action MDP family whose total reward is bounded by 1, we show that an exploration-then-commit algorithm can achieve $\widetilde{O}(d^3H^{1/3}T^{2/3})$ regret over $T$ episodes, using $K=\widetilde{O}(Hd)$ global updates. Our analysis connects bootstrapped feature acquisition, concurrent neural refinement, and finite-budget policy performance.

arXiv:2611.04157v1stat.ML

Nonparametric GARCH: An Iterated Neural Estimator

We study nonparametric estimation and forecasting for GARCH-type processes and propose an iterated neural estimator that alternates between estimating the recursion map and updating the unobserved conditional variances. Under mild conditions, we establish finite-sample bounds for the time-averaged in-sample variance-path error and show that the estimator is minimax rate-optimal up to logarithmic factors. Under stationarity and mixing conditions, we establish the same rate for recursion-map estimation and derive finite-sample bounds for out-of-sample variance forecasts and expected excess Gaussian quasi-likelihood (QLIKE) loss. The convergence rates depend on the true lag orders and remain unchanged when the fitted orders are overspecified. Simulation studies illustrate the effectiveness of our estimators under nonlinear and time-varying volatility dynamics. Real-data applications suggest that our neural methods perform competitively and can incorporate option-price and high-frequency information.

arXiv:2611.04155v1cs.CV

Improving PI-RADS Prediction and Prostate Cancer Detection from Multiparametric MRI with Anatomical Guidance

Accurate prediction of Prostate Imaging Reporting and Data System (PI-RADS) scores from multiparametric MRI (mpMRI) remains challenging, more so than clinically significant prostate cancer (csPCa) detection. Given the importance of lesion location in PI-RADS scoring, we evaluated whether incorporating prostate zonal masks improves PI-RADS prediction and csPCa detection across multiple architectures (3D ResNet-18, ResNet-50, and ViT) using a large-scale internal dataset. Anatomical guidance generally improved performance across architectures, with the best results achieved by ResNet-18 with 2-zone guidance for PI-RADS prediction (QWK = 0.511) and ResNet-50 with 5-zone guidance for csPCa detection (AUC = 0.809). Fine-tuned results on external datasets further demonstrated the benefits of anatomical guidance. Furthermore, models with anatomical guidance improved the utility of model-predicted PI-RADS scores for PI-RADS-based csPCa detection (positive defined as PI-RADS>=3). Combining model-predicted and expert-assigned PI-RADS scores significantly increased the AUC from 0.711 to 0.756 (p < 0.05), suggesting that AI-derived PI-RADS scores may provide complementary information to expert assessment for csPCa detection. Overall, our findings highlight the value of incorporating prostate zonal anatomy into deep learning-based prostate MRI assessment.

arXiv:2611.04153v1cs.AI

Palimpsestic Accountability in the US Federal AI Inventories

An AI inventory is a structured, often public record in which a government or organization lists and describes the AI systems it uses or plans to use. Studies of single AI inventory releases describe what systems do but say little about how they are governed over time. We linked entries across three releases of the US Federal AI Use Case Inventory (2023--2025), compared answers only where both releases required them, and analyzed governance accounts through a dual deductive-inductive approach, guided by the Algorithmic Decision-Making Adapted for the Public Sector (ADMAPS) framework. As the inventory grew from 710 to 3,611 entries from 2023 to 2025, fewer required answers were left blank, but accountability was rewritten through three mechanisms. Each release \emph{redrew} the scope of the questions; agency-assigned classifications \emph{gated} what was reported; and answers were also \emph{overwritten}, with gains and losses concentrated in single agencies and little account of human oversight and judgment. These findings show that each release redraws the boundary of accountability and leaves the public to \emph{re-stitch} the records, a condition we term \textbf{palimpsestic accountability}. We recommend versioned records, reasons for every reclassification, public files that distinguish kinds of missing answers, and oversight documented as conditions to improve public sector AI accountability.

arXiv:2611.04152v1cs.CL

PACT: Certifying What Is Written in Parallel Decoding of Diffusion Language Models

Parallel decoding speeds up masked diffusion language models by writing several tokens per forward pass. Published commit rules choose these tokens by testing the predicted distributions. Under sampling the written tokens are drawn, and these tests do not see them. We find that answer quality follows the reliability of the tokens actually written. An answer's log probability splits into this reliability and the dependence within each step. We propose PACT, a training-free commit rule that draws the tokens first, certifies the set it will write and spaces the positions of each step apart. This certificate holds at every temperature, and we show that no rule with the same guarantee commits more positions. Experiments use four models of the LLaDA and Dream families and HumanEval, MBPP, GSM8K, MATH500 and IFEval benchmarks. PACT reaches the quality of published commit rules with fewer forward passes. Spacing removes positions from each step, yet it usually lowers the total cost because later steps then write more tokens. At the same speed, published decoders also write a word twice in a row far more often than PACT. These results suggest that commit rules should certify the tokens they write. Code is available at https://github.com/LiZaiyuan0619/PACT.

arXiv:2611.04147v1cs.CL

AdSteer: Native Advertising in Large Language Model Generation

Large language model (LLM) services face rapidly rising serving costs. To sustain free or affordable access for a broad user base, alternative monetization mechanisms are needed. Native advertising offers a promising approach, as it creates economic value by aligning user intent with merchant offerings. Yet how to integrate sponsored content naturally into LLM responses remains largely underexplored. Existing approaches primarily rely on prompt-level injection, multi-pass rewriting, or fine-tuning, which can compromise response quality or introduce additional computational costs. We propose AdSteer, a lightweight decoding-time advertising framework that adaptively adjusts token probabilities to integrate contextually appropriate advertisement into responses. AdSteer calibrates the strength and duration of advertising influence, making sponsored content a natural part of the response. Experiments on GEM-Bench across three application scenarios show that AdSteer achieves an average improvement of 14.1% in overall score over the strongest baselines and 22.5% fewer total tokens on average than rewrite-based methods. These results demonstrate that AdSteer provides an effective and practical solution for native advertising in LLMs.

arXiv:2611.04146v1cs.LG

RGDM: Robust Graph Diffusion Model

Diffusion models represent a powerful class of generative models known for their solid theoretical foundations and remarkable performance across diverse tasks and domains. Existing node level diffusion methods synthesize nodes and edges within a graph, but are not designed to learn a robust generator from adversarially corrupted graphs. In this study, we introduce the Robust Graph Diffusion Model (RGDM), designed to generate labeled synthetic graph structures consisting of nodes and edges that integrate seamlessly into a given graph. The RGDM consists of a Robust Graph Autoencoder (RGAE) and a Latent Diffusion Model (LDM). Leveraging an edge selection mechanism and an innovative low-rank regularization on the latent feature, the RGDM produces clean and high-quality synthetic graph structures, even when trained on graphs subject to adversarial attacks. Comprehensive experimental evaluations reveal that Graph Neural Networks (GNNs) trained on the augmented graph, which is formed by merging the original attacked graph with the synthetic graph structures, exhibit significantly improved robustness against various graph adversarial attacks in the context of semi-supervised node classification. The code of the RGDM is available at https://github.com/Statistical-Deep-Learning/RGDM.

arXiv:2611.04142v1cs.AI

Interpretable Early Termination of Web Navigation Agents via Closed Sequential Pattern Mining

Web navigation agents built on open-weight language models fail often and fail late, with most of the tokens spent after the trajectory has become unrecoverable. Recent work has shown that early termination is a viable cost-reduction strategy, but existing approaches produce opaque termination decisions that do not support diagnosis or human oversight. In this work we apply closed sequential pattern mining (BIDE+) to K-step prefixes of agent execution traces to produce a small, human-reviewable failure-signature library that organizes into a four-category failure taxonomy with associated repair recommendations. We use a multi-variant evaluation protocol that explicitly removes contamination artifacts (zero-step crashes and trivially separable single-step successes) so that reported detection accuracy reflects behavioral discrimination rather than trace-length distributions. We evaluate our methodology on 1,544 MiniWoB++ traces collected from three open-weight language models, achieving the best F1 at K=3 on the Exclude-errors configuration across eight methods (0.717), recovering 88% of test-set failures within three steps, and supporting an early-termination policy that saves 26.3% of tokens at 92% precision.

arXiv:2611.04141v1cs.CL

Where Do Distilled Students Store Thinking?

Small students degrade on large reasoning models' (LRMs') long solutions, so recent work shortens them by interpolating an LRM with a short-answer model at ratio $\lambda$. We ask what student weights retain when this teacher stops thinking, comparing LoRA adapters of three students from three teacher pairs at each $\lambda$. In each pair, the adapters' update subspace shifts most where the teacher stops thinking and solutions shorten most, locating the change without attributing it. We identify a shared adapter subspace tied to thinking-like behavior (reflective markers, monologue format and long solutions). Shared by adapters before that point, it keeps about half (post hoc) while one pair's teacher halves its solutions but keeps its think tag, and largely empties after thinking stops. Removing it largely turns this behavior off for every pair, student and benchmark. Only in that pair and one other student does it leave clearly fewer markers than removing the update's top components; elsewhere, same-energy random directions often nearly match it. Grafted into a short-answer student at twice its weight, it turns early markers (a post hoc count) on, like the whole update, but the student rarely finishes. Our work identifies a weight subspace tied to distilled thinking-like behavior.

arXiv:2611.04139v1astro-ph.IM

COSMIO: A Benchmark for Cross-Survey Modality Imputation

Large-scale astronomical surveys define a heterogeneous, partially observed multi-survey learning setting: each object is observed by only a subset of surveys, while survey observations differ in resolutions, wavelength coverage, noise characteristics, and footprints. This setting gives rise to a natural machine learning problem: given observations of an object in some surveys, how to infer its appearance in other surveys. We formalize this as the problem of cross-survey modality imputation and introduce COSMIO, a multi-survey benchmark of aligned observations obtained by forward-rendered shared astrophysical scenes across LSST, Euclid, and Roman under survey-specific instrument characteristics. The benchmark enables systematic evaluation of imputation across source-target survey combinations and missing-survey patterns. We evaluate a broad range of machine learning methods for imputation, which reveal a clear domain gap: methods developed for existing imputation settings transfer poorly to the survey setting. We identify rare-event imputation as a central challenge. This regime arises in the early stages of new surveys, when rare but scientifically valuable phenomena may be available only from prior surveys. We use strong gravitational lensing, where light from a background galaxy is gravitationally deflected by a foreground galaxy, as a representative rare-event setting. We evaluate model performance on zero-shot imputation of gravitational lenses and on downstream scientific tasks using the imputed lens observations. On our constructed evaluation set of galaxy-galaxy strong lenses, performance degrades under rare-event distribution shift, and this degradation further propagates to the downstream lens-detection task. Code: https://github.com/fw-ic/COSMIO.

arXiv:2611.04137v1cs.LG

LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers

Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's.

arXiv:2611.04126v1cs.AI

GlucoAgent: Personalized Glucose Forecasting from Multimodal Personal Context

Continuous glucose monitors, smartwatches, meal logs, and health profiles form a rich multimodal personal context that can significantly enhance glucose forecasting. However, existing methods struggle to effectively utilize these signals to improve glucose prediction because the relevance of different modalities varies drastically across individuals and time. We present GlucoAgent, an LLM-based agentic system that addresses this challenge through dynamic, case-by-case analysis at inference time. For each forecast, GlucoAgent writes and runs code over an individual's multimodal context and returns predictions with a reason and supporting evidence. The generated programs are highly diverse, ranging from retrieving similar past meals to fitting and back-testing physiological and statistical models. On CGMacros, a multimodal glucose dataset, GlucoAgent outperforms time series foundation models on every forecasting horizon, achieving an average RMSE of 22.5 mg/dL compared to 23.5 mg/dL of Chronos-2, the leading multimodal time series foundation model. Furthermore, GlucoAgent is highly effective at leveraging multimodal data. The addition of multimodal data reduces its average RMSE by 1.4 mg/dL (24.0 mg/dL to 22.5 mg/dL), compared to a marginal 0.3 mg/dL (23.8 mg/dL to 23.5 mg/dL) improvement for Chronos-2. This advantage is most pronounced in challenging post-meal windows, where GlucoAgent uses multimodal context to reduce 60-minute forecast RMSE by 4.6 mg/dL (31.0 mg/dL to 26.5 mg/dL), far exceeding the 0.8 mg/dL (31.9 mg/dL to 31.1 mg/dL) reduction achieved by Chronos-2.

arXiv:2611.04123v1astro-ph.GA

An automated photometric pipeline for measuring the corotation radius of spiral galaxies

The corotation radius R_CR of a spiral galaxy marks the radius where the angular speed of the spiral pattern matches the orbital speed of the disk material. Despite its importance for understanding spiral structure and galaxy evolution, very few galaxies have corotation measurements, often derived using heterogeneous techniques. We present an automated photometric pipeline that identifies corotation radii by tracking the azimuthal phase offset between young and old stellar populations. Building on the color-gradient framework of \citet{Puerari1997}, we extract azimuthal brightness profiles from deprojected dual-band images, fit periodic models to the profiles, and identify corotation candidates at radii where the phase difference between the two stellar tracers changes sign. Each candidate is assigned an adaptive uncertainty interval and reported with a goodness-of-fit metric. We benchmark the method against a cross-method literature compilation, scoring every technique -- including our own -- against a reference built from the published measurements alone and evaluated leave-one-out. Over the $90$ galaxies for which such a reference exists, the pipeline places a correct corotation radius in $80.0\%$ of them, recovers $63.6\%$ of their known corotation radii, and reaches a harmonic mean of recall and purity of $0.58$, the highest values of any method in the comparison. Unlike the published methods, each of which was applied to a small, curated set of galaxies, our pipeline is run on the entire reference sample without any pre-selection, a sample two to six times larger than existing methods. The pipeline is scalable to large surveys, flexible with respect to spiral mode, and naturally recovers multiple corotation radii in galaxies with complex spiral structure, providing a uniform framework for studying disk resonances across diverse galaxy populations.

arXiv:2611.04122v1cs.CR

ScrambleMark: Zero-Query Watermark-Agnostic Black-Box Evasion of Modern Speech Watermarking

Speech watermarking is becoming a key defense for digital copyright and content authentication. Existing attacks treat evasion as a removal problem: they erase the watermark, relying on detector access or on speech data that is hard to collect, such as paired clean-watermarked or clean target speaker speech. Commercial platforms are moving to close off these assumptions. TrustedAudio, for instance, restricts detection to authenticated users and denies meaningful detection feedback to unauthorized parties. Adversaries thus cannot tell whether a sample is watermarked or which watermark it carries, nor reliably remove it without degrading speech quality. In this work, we show that evading a speech watermark does not require erasing it: detection relies on the structural consistency of the embedded pattern, and disrupting this consistency is enough to defeat it. Building on this observation, we identify harmonic structure as an effective attack surface, as perturbing it breaks watermark consistency while preserving perceptual quality. We therefore propose ScrambleMark, a zero-query, watermark-agnostic, black-box attack that evades detection through harmonic scrambling. ScrambleMark combines a deterministic engine for discovering transferable harmonic perturbations with a generative model that learns randomized scrambling patterns, avoiding reliance on fixed attack rules. Across three commercial platforms (Google SynthID, Resemble AI, and TrustedAudio), ScrambleMark is the only attack to both evade detection and keep speech natural, achieving nearly 97% evasion and a 4.2/5 human rating, far above the 1.1 to 1.7 of prior attacks. Beyond evasion, ScrambleMark withstands adaptive recovery designed to reverse it, evidence that structural evasion is hard to mitigate in practice.