PyTorchKR Papers

최신 논문

총 523,801편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2611.04449v1cs.LG

Near-Zero Token Embeddings Silently Stall Training in Gated Sequence Models

Among 291 surveyed open-weight language model repositories (280 distinct embedding tables), 52.6% of untied gated and hybrid checkpoints (20/38) contain reachable input embeddings with norms below $10^{-6}$, compared with 0% of weight-tied models. Such near-zero rows can arise during pretraining when unseen rare or template tokens receive zero input gradients while AdamW weight decay shrinks untied embeddings. In tested Transformer configurations, non-initial positions or a BOS token prevent this failure mode. In gated sequence models (Mamba, Gated DeltaNet, xLSTM), a near-zero token can amplify input-embedding gradients by more than 30 orders of magnitude (up to 35). At zero input, current-token readout factors (SiLU gates or linear queries) vanish, suppressing historical-state readout. Residual RMS values can then fall below the square root of the RMSNorm floor across consecutive layers, causing Pre-RMSNorm to amplify backward gradients by up to $10^3$ per layer. When a gradient entry exceeds $1.84 \times 10^{19}$, FP32 global-norm clipping in PyTorch, Hugging Face Trainer, and FSDP2 can overflow to infinity and zero the current batch's gradients. In Telugu continual pretraining of Nemotron-3-Nano-4B, this discards updates on approximately 20% of steps and leaves triggered rows frozen. We propose two remedies. For released checkpoints, GroupClip decouples input-embedding clipping using a 64-bit or rescaled norm; its rescaled implementation eliminates update loss and heals triggered rows in one step with negligible measured end-to-end overhead. For pretraining, GateFloor ($\mathrm{SiLU}(z) + \delta$, $\delta \in [0.01, 0.1]$) keeps the gate positive at zero input, reducing near-zero gradients by 28 orders of magnitude in experiments up to 370M parameters, with perplexity comparable to the baseline.

arXiv:2611.04447v1cs.AI

Formalizing PTX in Lean to Prove GPU Kernel Correctness

We present an approach to mathematically formalizing PTX, an approach that allows modeling and reasoning about patterns common in GPU kernels for Deep Learning, including acquire-release, tensor cores, proxies, async operations, multi-kernel pipelines and data-dependent control flow. We implement it in Lean and use it to prove the correctness of simple implementations of a WGMMA GEMM kernel, FlashAttention, and a 3-kernel MoE pipeline.

arXiv:2611.04442v1quant-ph

Hybrid Quantum--Classical Pauli-Correlator Architecture for Image Reconstruction

We introduce the Pauli-Correlator Autoencoder (PCAE), a hybrid quantum--classical architecture for image reconstruction. PCAE splits an image into patches and compresses each patch as a matrix product state (MPS). A small quantum circuit turns each compressed patch into a few measured numbers, and a classical neural network rebuilds the patch from these numbers and the patch position. A learnable energy term, taken from physics, is added to training and tracked as training proceeds. Two versions place the six qubits on a line (PCAE1D) or on a grid (PCAE2D). An averaging step makes the PCAE2D features follow rotations and reflections of the image exactly. Trained on one image at a time, PCAE2D rebuilds images from six public datasets with a peak signal-to-noise ratio of 25.8 to 41.6 decibels. On CIFAR-10 images not used in training, a formula-based stand-in for the PCAE2D features performs within 1 decibel of several simpler classical methods. Five trained models also run on real IBM quantum computers without retraining, and rebuild images at 25.90 to 31.52 decibels. A short argument and a synthetic test show that, with a linear final layer, measuring qubits in pairs is needed for targets built from products of inputs. These results show that PCAE works as a hybrid model that rebuilds varied images and runs on today's quantum hardware at the six-qubit scale.

arXiv:2611.04438v1cs.CV

Entropy-Gated Dynamic Rectification for Object Hallucination in Large Vision-Language Models

Large Vision-Language Models (LVLMs) routinely name objects that are absent from the image. Latent steering can reduce this failure in one model forward pass, but a fixed gate treats early and late decoding steps by the same criterion. We observe that predictive entropy varies sharply within free-running generation, while states that commit to short answers occupy a narrow, confident band. We exploit this temporal mismatch by using uncertainty as a causally available, token-local control signal for an existing steering gate. We propose Entropy-Gated Dynamic Rectification (EGDR), which lowers the intervention threshold when entropy derived from the preceding step's already available logits is high and retains the calibrated operating point when it is low. On LLaVA-1.5-7B, the underlying static gate lowers CHAIR CS/CI from 53.00/15.30 to 32.00/14.30; replacing only its constant threshold with EGDR further lowers them to 29.00/13.60. At a matched 58% intervention rate, the corresponding static control obtains 30.50/14.00. The gain over the static gate holds on LLaVA-NeXT-7B and grows with the generation budget, while POPE scores and general multimodal capability remain effectively unchanged. EGDR requires no additional model forward pass and adds roughly 2% measured latency. Our code is available at https://github.com/alvi-uiu/EGDR.

arXiv:2611.04435v1cs.CV

A Systematic Point Cloud Edge Detection Framework for Automatic Aircraft Skin Milling

The edge detection technique is an essential step for aircraft skin milling in aviation manufacturing. Most of the current detection methods focus on traditionally defined edge extraction tasks but disregard the crucial systematic requirement of edge milling. In this paper, we proposed a novel edge detection framework for automatic edge milling of aircraft skins. First, an edge probability detector is proposed by the spatial tangent continuity (STC) to provide the essential reference. Second, we propose a hierarchical branch searching (HBS) method to hierarchically strip the desired milling edges from the raw point cloud, which consists of three graded progressive steps: branch backbone generation, branch extension, and branch pruning. We demonstrate the performance of the proposed method on both synthetic models and aircraft skin workpieces. The proposed method outperforms the other baselines and shows accurate edges for the edge milling task.

arXiv:2611.04434v1cs.LG

Sketchpad: certified exact arithmetic with a small vision-language model

Language models are accurate on most long-arithmetic problems, but their accuracy is estimated on samples and carries no guarantee for an unseen input; tool use is exact but moves the computation outside the network. We show that exactness at all input lengths becomes a finite, checkable property of a network when each of its steps reads only a bounded view of an external working memory. In Sketchpad, Qwen3.5-0.8B with a low-rank adapter executes arithmetic procedures on a rendered sheet. Each forward pass reads a three-cell view and emits one step: a digit, a carry or borrow, and the next operation. A fixed runtime moves the view and writes no digit of a result. We prove in Lean that the model is then exact on all inputs of any length if and only if it is exact on the finite set of views the runtime presents, that no finite test suffices when steps also read earlier steps, and that the guarantee tolerates bounded numerical perturbation and extends to compositions of procedures. All 2,446 views pass on the deployment device, a laptop GPU running the model in a web browser. The certified model performs signed addition and subtraction, long multiplication and division, square roots and greatest common divisors at arbitrary length. It solved 200 of 200 pre-registered problems in the browser, and 220 of 220 sums of up to 1,000 digits and products of up to 20 digits, of which GPT-6.1 Sol, the strongest of ten other models, solved 217. A symbolic runtime built on these procedures solves, expands and factors polynomials with every number computed by the model, and a multiplicative gate on the adapter recovers the base model exactly when the sheet is not in use.

arXiv:2611.04423v1cs.CV

SWAP-INR: Implicit Neural Representation with Switchable Attention Paths for Time-varying Multivariate Scientific Data Compression

Visualizing multivariate scientific simulation data is essential for understanding complex physical phenomena, yet the massive scale of such data imposes a severe storage bottleneck in high-performance computing environments. Implicit neural representations (INRs) have emerged as a promising continuous alternative to traditional lossy compressors, but existing methods struggle to explicitly capture cross-variable interactions and to adapt to the diverse spectral characteristics of physical variables. To address these limitations, we propose SWAP-INR, an architecture for high-fidelity compression of multivariate, time-varying simulation data. SWAP-INR introduces two key components: (1) The switchable attention path (SWAP) module explicitly models cross-variable interactions through a cross-variable attention layer, then routes each variable to an attention or non-attention path by a deterministic, loss-driven criterion, and prunes the inactive layers for structural compression. (2) The adaptability amplification SIREN (A2-SIREN) replaces rigid frequency scaling with channel-wise learnable modulation, adapting to heterogeneous physical features with negligible overhead. Combined with post-training 16-bit quantization, SWAP-INR achieves state-of-the-art performance on small modular reactor (SMR) datasets and remains competitive on public physics simulation datasets, while maintaining a constrained memory footprint. Our code is https://github.com/hvcl/SWAP-INR.

arXiv:2611.04422v1cs.CV

GVT3D: Gesture-Vision-Text Token Enrichment for 3D Embodied Reference Understanding

Natural human-robot interaction requires robots to interpret language and gestures within a 3D environment to support tasks such as pick-and-place and object retrieval. 3D Embodied Reference Understanding (3D-ERU) supports this by identifying the 3D location of a target object from a scene (point-cloud) containing a human pointing gesture and an accompanying language description. However, current 3D-ERU models lack alignment of cross-modal representations and incur high processing costs that increase with the point-cloud's size. To address this, we propose GVT3D, which employs an Embodied Fusion module that jointly models language, gesture, and vision inputs within a shared embedding space, promoting cross-modal alignment. We additionally introduce a foveation approach that selectively processes the point cloud within a coarse region around the estimated pointing direction to reduce computations. GVT3D achieves ~8% higher accuracy and foveation enables 2.95x faster execution while maintaining comparable accuracy, relative to prior 3D-ERU models.

arXiv:2611.04421v1cs.CL

Typological Feature Induction as a Benchmark for Inductive Metalinguistic Reasoning

Existing metalinguistic reasoning benchmarks require both inference of rules and their application, making failures hard to diagnose. We introduce typological feature induction (TFI) which targets the former step. Given sentences in an unnamed language paired with translations in a known pivot language, models must predict a typological feature of the unnamed language. From Grambank and multi-parallel corpora, we automatically build evaluation questions over 195 languages. We find that LLMs identify 59--80% of these languages from only 16 sentence pairs; and misaligning the translations barely affects their accuracy, suggesting reliance on prior knowledge of the language. Enciphering text to prevent recognition increases the corpus' utility, but mainly for features where target and pivot languages agree. On languages which models cannot identify, neither open model outperforms a majority-value baseline. Since induction from raw text aligns with field linguists' use-case, this work indicates the models tested are not yet ready for this role. TFI offers a scalable path to track progress in this direction.

arXiv:2611.04420v1cs.AI

Verifier strength does not buy discovery quality: a pre-registered two-domain test of LLM-based automated research, with post-hoc evidence pointing to proposer-pool composition

Automated research pipelines pair an LLM proposer with a verifier on an implicit bet: discovery quality rises with verification strength. Under a pre-registered protocol we test this bet in two real-data domains -- job-shop scheduling and vehicular edge offloading on Beijing T-Drive trajectories -- across 100 matched LLM/random pairs per domain and three verifier tiers. The manipulation binds -- a 429-case mutation battery shows scores rising across tiers -- yet the paired LLM-over-random advantage does not scale with it (A, one-sided p=0.1115; D, p=0.9994), and in D shrinks under the strict tier (margin contrast -0.040, p=0.0013). The two lower gates promote identical candidates in 100/100 and 99/100 pairs, so the strict order is untestable when they coincide. Constructively, LLM proposals carry a real paired advantage over random controls, but a post-hoc union-pool analysis shows the fastest path to better discovery is widening the pool, not tightening the gate: mixing the two pools adds +1.52 holdout-value units in domain A (one-sided Wilcoxon, n=100, p=0.0002) -- evidence that pool composition, not verification strength, is the lever. Criteria, ledgers and scripts are archived on Zenodo (DOI: 10.5281/zenodo.23308866).

arXiv:2611.04418v1cs.CV

VidUEU-Agent: A Data-Curation Agent for Multimodal Understanding, Editing, and Unified Tasks

The advancement of multimodal learning is heavily constrained by the scarcity and high annotation cost of high-quality instruction data. While raw videos offer an abundant and dynamic source of visual knowledge, transforming them into precise instruction data remains highly challenging. To bridge this gap, we present VidUEU-Agent, an automated pipeline that synthesizes instruction data from raw videos for multimodal understanding, editing, and unified tasks. VidUEU-Agent follows a four-stage pipeline: it performs global perception to filter low-quality shots, mines keyframes and multimodal signals, constructs metadata and routes the metadata to synthesize instruction samples. More importantly, our agent framework can autonomously iterate the data distribution based on downstream training results, thereby optimizing data quality, balance, and training effectiveness. Scaling this pipeline, we construct VidUEU-100K, a large-scale dataset of 100K training samples spanning multiple domains and tasks, providing a robust foundation for developing comprehensive multimodal capabilities. Extensive experiments demonstrate that fine-tuning baseline models on VidUEU-100K yields significant and consistent performance gains across diverse representative benchmarks, validating the effectiveness of our proposed agent framework and dataset.

arXiv:2611.04416v1cs.CV

Efficient and Faithful High-Resolution Image Editing with Sparse Cross-Resolution Denoising

Instruction-based image editors still primarily operate at 1K, whereas practical editing increasingly requires faithful 2K/4K outputs. Dense high-resolution adaptation preserves the editing trajectory but incurs expensive target-source attention; editing at 1K followed by external refinement is cheaper but can separate detail restoration from the editing trajectory. We introduce FLASH, a native high-resolution editor based on sparse cross-resolution denoising. A frozen 1K branch provides a semantic and structural anchor, while a target-resolution branch reconstructs local detail within the same Transformer. An asymmetric topology retains full 1K attention, guides high-resolution queries with spatially aligned low-resolution noisy-target states, and confines high-resolution target-reference interaction to local Halo neighborhoods. Resolution-specific diffusion-state mapping and spatial YaRN adapt the pretrained editor to larger grids. An optional, training-free progressive path replaces early high-resolution denoising with a 1K prefix and completes refinement in the same editor. On VINS-4KEval, the direct model achieves the strongest structural fidelity among the compared methods at both 2K and 4K. A separate 4K forward benchmark with R = 1 records 4.2x faster total forward execution and 25.3x faster attention than dense target-reference attention. Progressive inference reduces end-to-end 4K latency by 30.5%, with a modest quality cost.

arXiv:2611.04414v1cs.CV

SphereFlow: Missing Modality Imputation via Geometric Transport on Hypersphere

Missing modality imputation aims to synthesize unobserved modalities from observed ones. Existing methods treat each modality as living on its own latent manifold, forcing the generator to learn a long, unstructured jump between them. We present SphereFlow, which takes a fundamentally different view: by projecting all modalities into a shared hyperspherical latent space via a frozen self-supervised encoder, cross-modal imputation reduces to short-range geometric transport between nearby points on the sphere. Realizing this idea with flow matching, however, exposes two geometric conflicts: the standard Gaussian noise source is catastrophically mismatched with the unit-norm data manifold, and the linear interpolation path departs the sphere surface into a region where the decoder has no training signal. We resolve both with simple, geometry-aware modifications: replacing the noise source with the observed-modality latent so that the velocity field learns only the cross-modal displacement, and training the decoder with a time-weighted loss that extends its domain to the near-surface region traversed by the flow. We prove that this data-as-source formulation reduces the worst-case off-manifold deviation from a quantity that grows with the latent dimension to a small, dimension-independent bound governed by inter-modality similarity. Experiments on BraTS multi-modal MRI imputation and CT-MRI translation demonstrate state-of-the-art generation fidelity across all settings, while downstream brain tumor segmentation with our imputed data closes roughly eighty percent of the gap to the full-modality oracle. Code is available at https://github.com/WTHH031230/SphereFlow.

arXiv:2611.04413v1cs.AI

CaliSplat: Sparse Gaussian Fitting and Calibrated Rate Allocation for Image Compression

2D Gaussian representations provide an explicit and lightweight alternative for image representation, enabling fast deterministic decoding without a heavy neural synthesis network. However, existing per-image Gaussian codecs still require costly image-specific optimization to determine how limited Gaussian capacity should be distributed across heterogeneous regions. We introduce CaliSplat, an optimization-efficient Gaussian image codec that selectively explores candidate representations and globally allocates a constrained Gaussian budget. We exploit the closed-form reconstruction error of the single-Gaussian case as an inexpensive measure of local difficulty, fitting higher-capacity candidates only in regions where additional Gaussians are likely to be useful. To reduce selection errors caused by mismatched convergence rates across candidate sizes, a small set of longer-horizon fits is used to calibrate short-horizon distortion estimates before global budget-aware selection. The selected representation is then compressed into a bitstream. Across standard and large image datasets, CaliSplat achieves competitive or superior rate--distortion performance while reducing per-image fitting time by $3.0\times$--$15.7\times$ compared with representative Gaussian codecs. It further maintains high-throughput deterministic decoding at $175.9$--$478.2$~MP/s. These results demonstrate that content-adaptive Gaussian compression can substantially reduce per-image optimization cost while preserving strong reconstruction quality and efficient explicit decoding.

arXiv:2611.04409v1cs.CV

LCM-UNet: Scale-Selective Mamba with Reparameterizable Multi-Basis Local Modeling for Lightweight Skin Lesion Segmentation

Accurate lesion segmentation underpins automated dermoscopic analysis: masks enable measurements of asymmetry, border and size. Deployment in primary care or on mobile devices requires a compact network. Existing ultra-light Mamba U-Nets offer global context, but global-modeling cost depends on each Mamba block's resolution, while their local branches rely on a single square kernel. We introduce LCM-UNet, a 33,474-parameter U-shaped network that fuses every skip at its native resolution, restricts bidirectional Mamba to stages at one-sixteenth of the input resolution or coarser, and uses local convolutions at finer scales. A multi-basis reparameterizable adapter augments each local branch during training with dilated, horizontal and vertical depthwise kernels, then folds them into the host kernel before inference. Each skip connection also learns a stage-specific gain. Nine networks were retrained on the official ISIC 2017 and ISIC 2018 partitions under a harmonized protocol, with PH$^2$ reserved for external validation and ablations run on a development split. LCM-UNet achieved Dice scores of 88.62%, 89.73% and 90.46% on ISIC 2017, ISIC 2018 and PH$^2$, respectively, ranking first on both ISIC test sets and second on PH$^2$ among the compared models. In the native-resolution host, removing Mamba from the one-eighth-resolution decoder stage eliminated 63% of Mamba operations, corresponding to 5.3% of all operations, with a Dice change of $-0.03$ percentage points. The adapter improved Dice by 0.48 percentage points over a single host kernel without adding deployed parameters or operations. With latency so far measured only on a GPU, LCM-UNet is a candidate for resource-constrained deployment.

arXiv:2611.04406v1cs.CL

The Min-Max theory of language behavior

The "Min-Max" theory of language behavior proposes that humans engage simultaneously in minimization of processing effort and maximization of information content, and that this is a primary driver in language evolution (Ring, 2025). This claim seeks to unify observations from language processing (Mahowald et al., 2013; Futrell et al., 2015; Yadav et al., 2020; Hawkins, 2014), interaction (H\"orberg and Sjons, 2023; Tucker et al., 2025; Jaeger and Tily, 2011; Smith and Culbertson, 2025), and corpora (Levshina and Moran, 2021; Levshina, 2023; Newmeyer, 2017), which suggest that considerations of processing efficiency and information content are both implicated in human communication. The interaction between these two processes provides an explanation for human linguistic behavior at various levels (Linders and Louwerse, 2023) and predicts historical developments in language, i.e. related to word order (Ring, 2025). Given that the original formulation of the Min-Max theory is based on observational data, the current paper seeks to validate the theory mathematically. Through a series of experiments that use computational models to investigate the theory's predictions, pretrained and iterated learning models are queried with synthetic data. The experiments show that manipulating the variables of token length and the weight of information content in a tabula rasa model results in word order change, as predicted by the theory. The experiments also identify important differences between computational models and observations of human language data, suggesting that current transformer-based language models may not accurately represent psycholinguistic processes. The overall findings indicate that evolutionary trajectories in language are highly susceptible to minimization of processing and maximization of information content.

arXiv:2611.04405v1cs.CV

DR-LiteNet: A Lightweight Explainable Hybrid CNN Framework for Imbalanced Diabetic Retinopathy Grading via Adaptive SMOTE Fusion

Automated diabetic retinopathy (DR) grading from colour fundus photographs is held back by two compounding challenges: severe class imbalance across five severity grades and the heavy computational cost of large scale architectures in resource constrained clinical settings. We introduce DR-LiteNet, a dual branch hybrid convolutional neural network that tackles both problems at once. The CNN branch extracts 1,280 dimensional deep features through EfficientNetB0, while the handcrafted (HC) branch produces a compact 51 dimensional lesion descriptor capturing vessel density, exudate intensity, texture entropy, and HSV colour histograms. Concatenating both branches yields a 1,331-d hybrid embedding, on which Adaptive SMOTE (ADASYN) is applied during training to synthesise minority class feature vectors, correcting distributional skew without touching the image space. A two phase training protocol first warms up the classification head on balanced features, then fine tunes the full network end to end. Grad CAM saliency maps deliver post-hoc clinical explainability. Tested on APTOS 2019 under five fold stratified cross validation, DR-LiteNet reaches 88.77% +/- 0.61 accuracy, a quadratic weighted kappa (QWK) of 0.884 +/- 0.010, and a macro F1 of 0.792 +/- 0.008 with only 4.69 M parameters beating ResNet50 (23.5 M) by DeltaQWK = 0.022 while staying practical on a single Google Colab T4 GPU. The source code of this project is available at: https://github.com/Nishi-Kanta-Paul/DR-LiteNet.

arXiv:2611.04399v1cs.AI

Silent Instrument Errors Reorder Agent Leaderboards

On a public agent leaderboard, a general-purpose agent trails a scaffold built for an airline customer-service benchmark by 13 points on the tasks that touch its database, and a reader could conclude that generality costs accuracy. The gap comes from the grader. The general agent often ends its conversations in a way that never triggers grading, so only about half of its episodes were scored and the rest were stored as failures. Grading the ungraded half under the benchmark's own rule closes the pooled gap, without rerunning any agent. We audit two carefully maintained public leaderboards, HAL and tau2-bench, from their published artifacts alone and find three more instrument errors. A cost column counts nested model calls twice, unevenly across rows. Archived trajectories record grader-written actions as agent actions. Re-grading saved trajectories under a newer evaluator is itself a new measurement: a correct and announced fix reorders a leaderboard, and a mode check that misreads four files collapses their scores by 59 to 82 points while printing no error. Each produced a plausible number, and the HAL maintainers confirmed their three within five weeks. Together the cases show that the infrastructure of agent evaluation, meaning the code that grades, logs, replays and versions an evaluation, is part of the measurement and fails silently, so scores and orderings can move with no visible sign. These failure modes follow from how agent evaluation pipelines are built, and those pipelines change as fast as the agents they measure, so we expect more of them elsewhere. We propose cheap checks that expose the instrument: report the graded fraction, aggregate cost from root spans, tag the provenance of every recorded action, and pin evaluator versions with a published re-grade diff.

arXiv:2611.04397v1cs.CV

Auditing Targetless Camera-LiDAR Extrinsic Correction before Recursive Use

Camera-LiDAR fusion repeatedly uses an extrinsic transform in projection, association, and recursive state estimation. We study whether a targetless correction should be admitted to a downstream consumer only after separate audits of correction quality, confidence ranking, and recursive risk. Our log-disjoint protocol evaluates these stages under synthetic extrinsic drift on Argoverse 2 and treats cross-dataset and odometry experiments as bounded diagnostics. On 27 test-split logs, single-frame Canny alignment is harmful in the mild tier. Multi-frame alignment improves severe-tier rotation by 2.52 degrees but worsens translation by 0.177 m; its declared joint error-magnitude improvement is also negative (-0.158, 95% CI [-0.223, -0.098]), so the rotation gain is not a six-degree-of-freedom correction success. A zero-injection check moves away from the true extrinsic, and solver-disagreement AUROC on selected splits does not yield a validated nonzero-coverage risk-coverage operating point. In a 15-log final system role, the locked conservative policy admits no candidate and equals Raw, whereas Always-correct increases trajectory error. These results support an auditable separation of correction, confidence, and recursive contamination; they do not establish a deployable calibration gate, positive odometry or SLAM improvement, map consistency, or closed-loop safety.

arXiv:2611.04395v1cs.AI

Correct Labels Are Not Enough: Reliable Metadata Can Silently Control Model Behavior

Fine-tuning data can include metadata that predicts the correct answer, such as a reward-model score or a quality label. We ask whether a model learns to rely on this metadata even when every training label is correct. We fine-tune 7-9B open-weight models from three families on multiple-choice questions in which (Reward: 5) always marks the correct option; when this annotation is moved to a wrong option at test time, models select that option on 97-100% of CommonsenseQA items. Removing it leaves most answers correct, so evaluation without it misses the dependence. Cue-following rises sharply with training-time reliability and develops even with neutral or nonsense annotation words, although test-time wording affects its strength. A Qwen3 experiment also finds cue dependence in option-letter scores after direct preference optimization. Audits can test for this dependence by moving or removing visible annotations while holding the task and correct answer fixed.