Model Pretraining

To train handwritten text recognition systems we need word images and their corresponding transcriptions, and these transcriptions are produced manually. For a script that can be read by only a small number of specialists, this manual transcription is a limitation, because the trained models are supposed to save the time of those same specialists. A relevant question therefore arises: how many transcriptions are needed before a recogniser becomes useful, and how much of that cost can pretraining remove? In this study the answer is measured directly for handwritten Devanagari. We keep the recogniser, optimiser and evaluation protocol the same and change only the number of real transcribed words used for fine-tuning across nine budgets from 10 to 4,000 and four initialisation regimes, with six seeds at every point. The resulting curves are then converted into annotation-equivalent terms. A CER of 0.50 is reached by supervised synthetic pretraining using only 81 transcribed words, whereas random initialisation requires 355, which gives a label multiplier of 4.40 [3.56, 4.99]. There is a zero-shot reference point as well: with no real transcribed words at all, this pretraining is worth about 136 of them. This advantage gets smaller as the target accuracy improves, and at the most demanding target we measure, it cannot be distinguished from no saving at all. A fourth arm in which only the encoder is transferred separates the effect of the pretraining method from that of transfer scope, and masked image modelling is observed to transfer negatively over a bounded range of budgets. We emphasise that the scarcity in this study is constructed by subsampling a large corpus.

The growing use of electroencephalography (EEG) motivates automated analysis that is accurate, transferable, and deployable on constrained hardware. Recent EEG foundation models learn general representations from large-scale pretraining, but their size and computational cost limit edge and wearable deployment. We introduce BRIDGE-EEG, an efficient multi-task EEG classification pipeline that preserves the benefits of pretraining while reducing model size. A unified preprocessing scheme maps heterogeneous recordings with different channel counts, montages, and sampling rates to a device-agnostic 62-channel time--frequency representation. We pretrain an SE-ResNet18 teacher (11.84 M parameters) with SimCLR on unlabeled EEG from five heterogeneous datasets, then compress it into SE-ResNet8 (1.56 M) and SE-ResNet4 (0.48 M) students using task-agnostic and task-specific distillation. We evaluate six benchmarks spanning abnormality detection, motor imagery, and emotion recognition. For abnormality detection and emotion recognition, the students achieve accuracy comparable to or better than several recent EEG foundation models with 10--1,000$\times$ more parameters. Motor imagery shows a remaining representation gap, highlighting the importance of pretraining diversity. Inference profiling on a server GPU, desktop CPU, and NVIDIA Jetson Orin Nano shows up to 3.0$\times$ lower edge energy per inference (15.64 mJ vs. 46.67 mJ). The compact models further support future deployment on MCU-class wearables.

fMRI foundation models increasingly aggregate heterogeneous data across brain states, cohorts, and acquisition settings, yet pretraining domains are commonly treated as a flat mixture and downstream tasks are adapted independently. We study whether measured learning relations can organize both stages without modifying the backbone. During pretraining, a lightweight Brain-DiT proxy estimates difficulty and directed facilitation across ten fMRI domains, yielding a priority-guided cumulative domain curriculum combined with high-to-low-noise timestep scheduling and joint consolidation. During adaptation, controlled first- and higher-order transfer across fifteen tasks constructs a directed taskonomy, from which budgeted integer programming (BIP) selects directly supervised source tasks and target-specific routes. The joint priority-domain and high-to-low-timestep curriculum reduces v-NMSE, PSD-NMSE, and FC-MSE by 6.5%, 16.3%, and 10.5%, respectively, relative to uniform sampling over both dimensions, and shows strong downstream performance across six in- and out-of-domain tasks. The taskonomy reveals asymmetric, target-dependent transfer, while exploratory sealed-test evaluation shows larger descriptive gains for BIP policies when higher-order route spaces are available than for matched random controls. Together, these findings support organizing fMRI pretraining and adaptation by measured learning relations rather than treating domains and tasks as independent flat sets.

Time-series foundation models are evaluated almost exclusively on public archives that predate them, so a strong score cannot be separated from having seen the test set during pretraining. The obvious remedy is a hold-out that postdates the models. We build one: thirteen forecasters -- four classical, three trained per dataset, six pretrained -- on seven groups drawn from five domains, every observation published after the last model was released, and every dataset rebuildable without an API key. Under this protocol pretrained models win 5 of 7 groups, lose one to a Theta baseline, and on daily exchange rates are indistinguishable from a seasonal naive forecast, along with every other method tested. We then ask what separates the wins from the losses, and report a negative result: the two intrinsic properties one would reach for -- seasonal strength and spectral entropy, measured on the input window -- do not account for the pattern, and seasonal strength is if anything negatively associated with the advantage. What does track it is corpus familiarity. Our largest gain (28% lower MASE than the best classical method, on weekly Wikipedia pageviews) falls on Wikipedia pageviews, the domain TimesFM's authors describe as the bulk of its pretraining corpus, at the same granularities and differing only in time window. Within the pretrained family, where every model forecasts identical series so that series difficulty cancels, the TimesFM family outranks the Chronos family by -0.53 ranks on Wikipedia against -0.09 everywhere else (1,500 vs. 754 series, Mann-Whitney p < 1e-5). We conclude that a temporal hold-out removes memorisation of a window but not familiarity with a domain, that benchmarks therefore need domain hold-outs stated relative to disclosed corpora, and that the practitioner's question is less which model is better than whether their domain is one the model was raised on.

Data augmentation is fundamental to training modern deep vision and multimodal models. While individual methods, such as RandAug, CutMix, Mixup, RandErase, and DropPath, offer strong regularization effects, their combined use has saturated in performance due to overlapping functionalities, and aggressive pixel-level manipulations may disrupt delicate cross-modal alignment. This saturation motivates the search for a new augmentation axis within the embedding space rather than the input space. We introduce Aether, a simple plug-in method that applies diffusion-style random perturbations in the embedding space via controlled alpha-mixing, specifically designed to provide isotropic regularization that remains semantically consistent. Inspired by feature-space perturbations in language models and image degradation in generative pretraining, Aether induces mild yet effective perturbations that smooth the representations without compromising the fine-grained structural information required for strong vision-language encoders. Across diverse architectures and across multiple recognition tasks, Aether delivers consistent gains over the advanced recipe combining CutMix, Mixup, DropPath, and RandAug---a level of improvement rarely observed with modern augmentation alternatives. Notably, Aether demonstrates superior effectiveness in multi-modal alignment, succeeding where traditional pixel-space augmentations fail by providing a stable, isotropic regularization signal that respects the integrity of the high-dimensional feature space.

Vision--language--action (VLA) models acquire broad generalization through large-scale pretraining, yet adapting them to a new task and robot embodiment still requires post-training on newly collected data. Unlike pretraining, post-training targets task- and embodiment-specific adaptation, making it particularly sensitive to data quality. In practice, collected robot datasets often contain heterogeneous errors, including execution mistakes, sensor drift, and timestamp misalignment, which can impair post-training and policy performance. Manual inspection is costly, while existing data-cleaning methods are typically tailored to particular corruption types. To address these challenges, we introduce \textsc{RoboDrop}, a data-curation framework that audits supervision using local gradient compatibility measured along the training trajectory as a proxy for its effect on post-training performance. During a one-epoch warm-up run, RoboDrop scores each candidate sample online by comparing its gradient with those of task-semantic and visually matched validation samples. The resulting sample scores are aggregated at the episode level, and a simple automatic post-processing rule converts them into filtering decisions. We evaluate RoboDrop on controlled observation--action corruptions, naturally suboptimal demonstrations in simulation, and real-robot datasets containing non-expert collection errors. Across these settings, RoboDrop more accurately distinguishes unreliable demonstrations than prior methods, while post-training on the curated data consistently yields stronger downstream policies, with average real-robot rollout success rising from $35.0\%$ to $67.5\%$. These results establish training-trajectory-aware, context-conditioned supervision auditing as an effective approach to robust VLA post-training.

Choosing a deep learning architecture for label-free single-cell classification remains an open question, with microscopy benchmarks reporting conflicting conclusions about CNNs versus transformers. We present a controlled benchmark on LIVECell phase-contrast microscopy data using source-image-disjoint train/validation/test splits to prevent parent-image leakage and matched optimisation, augmentation, and evaluation protocols across EfficientNet, Vision Transformer (ViT), and EVA-02 models. This allows the effects of architecture, pretraining, fine-tuning, tokenisation, and distillation to be disentangled. We find that the previously reported CNN advantage is largely explained by pretraining rather than architecture: the smallest pretrained model outperforms the strongest model trained from scratch despite far fewer parameters. Pretraining improves macro-F1 by 3-4 points, while the gap between the best pretrained CNN and transformer is below 0.5 points. Architectural choices nevertheless matter: ViT-S/8 outperforms ViT-S/16 and matches the four-times-larger ViT-B/16 at a quarter of the parameters, showing that finer tokenisation benefits small cell crops. Conversely, layer-wise learning-rate decay, central to the EVA-02 fine-tuning recipe, degrades performance, highlighting that transfer heuristics from natural-image recognition may not generalise to microscopy. Finally, knowledge distillation substantially improves the deployment frontier: compact EfficientNet-B0 students distilled from teacher councils outperform every individually trained backbone, including the EfficientNet-B5 and EVA-02 teachers. Overall, our results show that rigorous control of pretraining and evaluation is essential for interpreting biomedical architecture benchmarks, while distillation may be a more effective route to practical single-cell classification than architecture choice alone.

Speculative decoding is critical for accelerating LLM inference. However, the speedup is fragile: drafters are typically trained against a narrow distribution for a single target model, and their acceptance rate collapses under workload shifts. This is a striking inversion of modern LLM development, where target models are valued precisely for the broad generalization they acquire through large-scale pretraining. We argue that the natural remedy, pretraining, has been hard to apply to drafters because existing recipes are target-specific: the drafter consumes the target's hidden states and is distilled on the target's logits, so pretraining must be repeated for each target. We introduce Osprey, which instead bootstraps drafters from off-the-shelf pretrained small language models, treating broad pretraining as a reusable, target-agnostic asset and reducing per-target work to a lightweight adaptation step. Realizing this requires overcoming two challenges: small LMs are far deeper than a latency-bound drafter can afford, and their pretrained computation must remain intact while the drafter learns to ingest target hidden states and emit tokens in the target's vocabulary. Osprey addresses both by pruning to a shallow backbone, restoring its language-modeling capability with target-agnostic next-token pretraining, and adapting it to each target through vocabulary alignment, zero-initialized QKV expansion, and distillation from the target model's output distribution. Empirically, a single pretrained Osprey backbone transfers across targets and improves mean acceptance length by 16.1% for Qwen3-8B, 21.2% for Llama-3.3-70B-Instruct, and 22.7% for the 229B MiniMax-M2.5 (with 17.5% higher tokens per second), with the largest gains on out-of-domain and multilingual data. Our code is available at https://github.com/LeanModels/Osprey.

Language-model checkpoints are commonly selected by pretraining loss or benchmark scores, assuming that the highest-scoring checkpoint will remain the best starting point for subsequent training. We show that this assumption can fail in a full 30B mixture-of-experts training pipeline. The checkpoints that perform better after the full downstream training stack also have higher solution density, i.e., retain downstream performance under local weight perturbations.

Adversarial robustness in computer vision is still largely achieved through adversarial training or test-time adversarial purification, both of which introduce significant computational overhead by generating adversarial examples during training or performing iterative denoising at test time. We study whether empirical robustness can instead emerge from architectural and representation-learning inductive biases. We introduce Oscillatory Predictive Learning (OPL), a two-stage framework that combines Artificial Kuramoto Oscillatory Neurons (AKOrN) with predictive self-supervised pretraining using X-PhiNet. Because our default checkpoint uses randomized initial oscillator states, we compare it with other randomized adversarial defense methods that provide precise, reproducible, and strong attack protocols. Experiments on CIFAR-10 and CIFAR-100, with additional corruption evaluation on CIFAR-10-C, demonstrate that our method achieves competitive results under the AutoAttack-rand evaluation protocol. On CIFAR-10 and CIFAR-100, OPL attains 76.63$\pm$0.76$\%$ and 50.44$\%$ robust accuracy, respectively, under $\ell_\infty$, $ε=8/255$, AutoAttack-rand with EoT $K=20$.

Natural Language Processing (NLP) in the climate domain requires models to process heterogeneous text sources, including scientific literature, policy disclosures, and synthetic reports. However, how to effectively combine diverse domain corpora during continued pretraining (CPT) remains underexplored. We introduce Climate-ModernBERT, a family of climate-adapted encoder models obtained through continued pretraining of ModernBERT-Base on three climate corpora: academic climate text, climate-filtered web data, and synthetic climate documents. We systematically compare joint continued pretraining on corpus mixtures with parameter-space merging of independently specialized checkpoints. Across nine climate NLP benchmarks, our best model achieves 76.3 average F_1, improving significantly over a vanilla ModernBERT baseline by 2.8 points. Within the climate NLP setting, the results show that academic climate corpora provide the strongest adaptation signal among the evaluated sources, while parameter-space merging improves over joint multi-source training and better preserves complementary information from heterogeneous climate corpora. We release all Climate-ModernBERT variants and training checkpoints to support future research in climate NLP and domain-adaptive pretraining.

This paper proposes an open-set ego-noise separation framework for legged-robot audition via annotation-free adaptation and pretrained-model transfer. The framework removes robot-specific ego-noise while preserving environmental sounds whose classes are not specified in advance. Acoustic sensing provides cues about a robot's surroundings beyond the visual field, but walking-induced ego-noise from footstep impacts, joint-backlash rattling, and motor noise severely contaminates the recordings. The framework first uses RecurGraph to select ego-noise-dominant clips from the unlabeled recordings by aggregating clip embeddings into an embedding centroid and propagating scores over an audio-embedding graph. The selected clips are mixed with diverse environmental sounds from a large-scale sound-event dataset to provide paired mixture--target supervision for open-set separation. Transfer-DiT then adapts a general-purpose zero-shot neural separator to achieve high-fidelity open-set ego-noise separation for the target robot. Experiments with bipedal and quadrupedal robots show reliable clip selection and improvements in separation quality and downstream task performance over baseline separators. These results demonstrate the feasibility of annotation-free adaptation without separately recorded ego-noise-only data or manual clip-level annotations.

We introduce Xiaomi-TabLDM, a tabular large data foundation model for classification and regression via in-context learning, which delivers superior prediction accuracy without requiring task-specific fine-tuning. Pretrained exclusively on synthetic data generated from structural causal models (SCMs), our model enables more flexible context utilization and more efficient capacity scaling. i) A new performance standard. Strong regression performance across benchmarks: Xiaomi-TabLDM ranks 1st on OpenML-CTR23 and 2nd on regression across TALENT, TabArena, and BCCO, demonstrating consistently strong regression performance across four complementary benchmark suites. Favorable performance--efficiency trade-off: Xiaomi-TabLDM combines strong predictive performance with substantially lower computational cost. For example, on TabArena regression, it achieves the second-highest Elo while using 82% less training time and 68% less prediction time than the top-ranked TabFM. ii) Large-scale synthetic pretraining. Xiaomi-TabLDM expands the coverage and diversity of synthetic tabular data used for pretraining. We also adopt a three-stage training strategy together with dual-stream feature grouping, lightweight Attention Residual, and sparse Mixture-of-Experts, enabling Xiaomi-TabLDM to learn richer feature interactions and expert specialization across diverse tabular tasks. iii) Test-time scaling. Xiaomi-TabLDM further extends tabular prediction through test-time compute scaling, where allocating additional computation at inference time consistently improves predictive performance over the base model.

Action-conditioned video models require large-scale visual data paired with control signals that are temporally aligned with the resulting scene transitions. Such supervision is difficult to obtain from ordinary real-world video because the actions that caused each visual change are typically unknown. We present a large-scale synthetic data production pipeline built on Unreal Engine for generating action-conditioned, multi-view video. To accommodate the different execution requirements of real-time physics and high-quality offline rendering, the pipeline executes trajectory generation and final rendering in two stages: Stage I runs real physics in PIE and records per-frame character states, control inputs, and camera states into an intermediate trajectory representation; Stage II replays those trajectories in a new engine process and renders them offline with Movie Render Queue (MRQ). Around this core, we develop a distributed production system with cache-aware task partitioning, node-local slot scheduling, automated scene screening, aesthetic and luminance filtering, partial-output recovery, asynchronous upload, and continuous cluster health monitoring. The production cluster contains 25 servers with eight NVIDIA RTX 5090 GPUs per server. From 2,384 asset packs, 429 levels were retained for production together with a pool of 40 humanoid characters. The pipeline has produced 2,691 hours of 1080p video and 6,076 hours of 720p video. We describe the system architecture, the implementation decisions that emerged from production failures, and the limitations of using perceptual quality proxies for world-model data curation. The pipeline described in this report constitutes the Unreal Engine synthetic-data production component used in EchoWM.

Self-supervised pre-training via cross-view completion learns strong features for 3D vision from co-visible regions of image pairs. However, the reference view provides little information for reconstructing non-co-visible patches, implicitly yielding a monocular training signal in these regions. We introduce Gekko, which turns this limitation into a useful signal. The relative improvement of the cross-view reconstruction error over a masked-autoencoder error is a self-supervised proxy for co-visibility: large improvements indicate co-visible regions, negligible ones non-co-visible areas. Gekko is a network, trained from scratch, that jointly performs cross-view completion, masked autoencoding, and per-pixel prediction of this relative improvement, providing an additional binocular signal for all masked regions without any ground-truth 3D annotation. Under identical architectures and training data, Gekko consistently outperforms CroCo on zero-shot correspondence estimation, relative pose estimation, and pointmap regression, with up to 6 times higher accuracy at the strictest relative-pose threshold and a 22% drop in end-point error on ETH3D. The extra channel it learns is itself a strong co-visibility detector on unseen scenes, and Gekko's frozen features outperform released cross-view backbones of comparable or larger size. It can also be trained directly from raw videos with a simple stride-based curriculum, removing the cumbersome 3D preprocessing prior methods require while matching models trained on curated data. Code and pre-trained models are publicly available.

User representation pre-training has become a fundamental paradigm for alleviating data sparsity in downstream personalization tasks. However, existing studies predominantly focus on single-app or app-level behaviors, overlooking the inherently cross-source and multi-granular nature of user activities on mobile devices. At the device level, user intent emerges from complex interactions among heterogeneous behavior sources and hierarchical action structures, posing challenges that cannot be addressed by conventional app-centric modeling. To tackle this issue, we propose CM-PTM, a novel Cross Multi-source Behavior Pre-Training Model tailored for mobile game user representation learning on device-level behavioral logs. CM-PTM employs hierarchical cascaded mask-then-predict proxy tasks that first infer the source of the next behavior and then progressively refine predictions at the app-action level. This design enables unified modeling of cross-source dependencies and fine-grained behavioral dynamics within a single pre-training paradigm. Extensive experiments on large-scale real-world mobile datasets demonstrate that CM-PTM effectively captures users' endogenous interests and consistently delivers significant performance gains on downstream mobile game recommendation tasks.

We propose EEG-VID, a task-guided latent predictive pretraining framework for EEG decoding under session and subject shifts. EEG-VID predicts future latent EEG states from recent history using an exponential-moving-average target encoder and weak task guidance, followed by supervised fine-tuning. Across VIG-48 and BCI Competition IV-2a/IV-2b, Stage 1 improves mean accuracy in 41 of 42 matched backbone-dataset-protocol comparisons, including all 12 leave-one-subject-out settings, with a maximum gain of 16.22 percentage points. On the 48-region cross-day VIG-48 task, EEG-VID achieves 6.52% Top-1 and 30.50% Top-5 accuracy. In a separate six-participant offline robot-scene study, candidate-constrained target selection reaches 40.24% versus a 25% chance level after subject-specific calibration. These results support task-guided latent prediction as a transferable pretraining strategy for EEG decoding and scene-constrained assistive target selection.

Interactive lesion segmentation in whole-body PET/CT requires a model to provide a strong initial prediction while also responding efficiently to sparse corrective scribbles during inference. This setting is particularly challenging because tracer distributions, physiological uptake patterns, lesion appearance, and acquisition characteristics differ substantially between FDG and PSMA studies. We present TRIAGE, Tracer-aware Refinement via Interactive Anatomy-Guided sEgmentation. The core backbone is a 3D STU-Net initialized through masked autoencoding pre-training with an asynchronous masking strategy, aiming to learn transferable anatomical and cross-modal representations before task-specific fine-tuning. In parallel, we train an auxiliary organ segmentation model whose predictions provide explicit anatomical context and help distinguish physiological uptake from malignant lesions. A dedicated tracer classifier first routes each study to an FDG- or PSMA-specific branch. Within each branch, a first-stage segmentation model consumes CT, PET, and organ context to generate an initial lesion mask. The initial prediction is then combined with cumulative foreground/background scribbles and refined by a second interactive segmentation network. The FDG and PSMA branches share the same overall processing pipeline but are trained independently to account for tracer-specific appearance and error modes. We additionally employ curriculum-style training and model ensembling to improve robustness across interaction steps and heterogeneous cohorts. Experiments are conducted using the official AutoPET V data and ten-fold split; quantitative results, ablations, and final test-set performance are left as placeholders to be completed after the challenge evaluation. Code: https://github.com/Liiiii2101/AUTOPET2026-MEDAI.

We compare six self-supervised pretext tasks for infant cry analysis under a fixed budget, meaning the same compact encoder of 1.17M parameters, the same 115 hours of license-verified public pretraining audio, and the same evaluation protocol for every candidate. On cry detection the reconstructive objectives dominate, and a linear probe over a masked-spectrogram encoder reaches 0.988 AUC with subject-wise splits even though the encoder never observed a cry during pretraining. On cry-reason classification over donateacry, the de facto public benchmark for cry reasons, every encoder performs at chance (0.38 to 0.54 macro AUC over 5 classes), and neither domain adaptation on 1.8 hours of real cries nor end-to-end fine-tuning moves the result. Since a frozen HuBERT-base with 80 times more parameters shows the same pattern, the bottleneck must sit in the labels and not in model capacity. We then reproduce the 90%+ accuracies of the donateacry literature on our own system by changing nothing but the evaluation protocol: clip-wise splits raise accuracy to 85.2% (barely above the 83.8% majority-class baseline), and applying augmentation before splitting raises it to 97.9%, matching the reported state of the art, from the same model that measures 0.49 macro AUC under subject-wise splits. Under leakage-free splits, a twentyfold augmentation of the labeled set (vocoder speaker perturbation and noise mixing, 21 hours) leaves cross-subject AUC unchanged: for this task the effective sample size is the number of infants. We release code, seeds and per-clip license manifests.

Global food security and sustainable climate action increasingly rely on robust, scalable agricultural monitoring. Earth observation foundation models have emerged as powerful, label-efficient tools across general remote sensing domains, yet early attempts to deploy them for agricultural applications have yielded surprisingly poor results. We hypothesize that this performance gap stems from the extreme heterogeneity of agricultural landscapes and the inherent inability of current earth observation foundation models to adapt to task-specific nuances. In this work, we systematically evaluate two critical bottlenecks hindering the deployment of foundation models in agricultural tasks, benchmarking two earth observation foundation models, a foundation model designed for tabular data, and conventional supervised baselines across seven real-world agricultural datasets spanning yield prediction, phenology estimation, and crop classification. First, we identify a pretraining-deployment modality gap: agricultural downstream tasks frequently require diverse, non-imagery data modalities that earth observation foundation models are architecturally unequipped to ingest, while a foundation model built for tabular data handles this heterogeneity more naturally. Second, we formalize the agricultural task space across five structural axes to demonstrate why current models fail to generalize reliably, resulting in highly unstable model rankings across evaluation settings. By characterizing these structural and modal gaps, our insights highlight the friction between general-purpose architectures and specialized agricultural downstream data, providing a strategic roadmap for developing the next generation of domain-aware foundation models.

A recent line of synthetic-data work reconstructs the thinking behind existing text rather than rewriting the text itself, but it operates on short web passages, recovers only local thoughts, and leaves the structure of whole documents untouched. Scientific papers are written to a clear and largely uniform structure and make a natural substrate for lifting this paradigm to the document level. We present a pipeline that unfolds each paper into a multi-turn generation trajectory in which a teacher model reconstructs the writing process of the whole paper: a writing request, a global plan, and pre-writing deliberation for each section. All section texts and the abstract are kept verbatim from the source paper. We apply the pipeline to 1.8M quality-filtered arXiv papers and obtain a 60B-token corpus for continued pre-training (CPT) that is roughly twice the source text. The same reverse construction extends to instruction data and evaluation. We build an SFT dataset of 200K samples using answers derived from paper text. We also use held-out papers to construct PAW-Bench, a benchmark of 2,940 academic writing tasks with per-task rubrics and checklists. In controlled experiments, CPT on our corpus followed by SFT on public datasets improves writing performance while preserving general reasoning and improving long document reading. Replacing part of the writing SFT data with our synthetic instruction data further improves performance on PAW-Bench.

We ask whether language-model pre-training can be decomposed into smaller, independently trainable jobs that can later be recomposed into a coherent larger model. We introduce Mixture of Training (MoT), a scaffolded modular pre-training procedure that partitions a target Transformer into contiguous layer blocks, trains each block inside a frozen pretrained aligner scaffold, and then recomposes the trained blocks with an optional short end-to-end adaptation pass. On a 1.3B-parameter Gemma-style model trained on C4, MoT provides a small-scale proof of mechanism: independently trained depth slices can be recomposed into a usable language model, and a quality-parity schedule reaches the same reported perplexity as the monolithic baseline. This parity setting processes more aggregate tokens and has a shorter idealized layer-equivalent critical path after aligner preparation; its effective compute advantage depends on reusing the aligner across runs. We therefore present MoT not as a general replacement for monolithic pre-training, but as a small-scale framework for studying whether scaffolded sub-runs can act as reusable training units.

Time series foundation models (TSFMs) have advanced primarily through architectural innovation, while training regimes for large-scale heterogeneous corpora remain under-explored. As a result, pre-training distributions are often poorly controlled with respect to domain imbalance, context requirements, prediction horizons, and missingness. We introduce ORBIT (Omni-Range Bootstrap Incremental Training), a training paradigm that makes this distribution explicit and controllable. ORBIT combines Bootstrap Multi-Level Sampling, which controls dataset exposure and samples records, target variables, context windows, and prediction horizons, with Omni-Range Incremental Training, which varies context lengths and prediction horizons throughout a single training stage. Under ORBIT, we train Falcon-2.0, a simple univariate encoder-only Transformer with missingness-aware triple-channel patch tokenization and parallel patch prediction. We further introduce Rank-Guided Cross-Depth Alignment, a training objective that uses late-layer representations as stop-gradient teachers for shallow layers without additional inference cost. Evaluations on GIFT-Eval and fev-bench demonstrate strong zero-shot forecasting performance across diverse domains and frequencies.

Automating empirical research is a long-standing direction of AI. Recent automatic research (AutoResearch) agents bring this goal within reach, as modern LLMs show the capability to independently implement solutions and learn from the execution outcomes. Behind these gains, post-training (especially RL) plays a central role. In this paper, we identify a fundamental tension when scaling RL for these agents: the two components of every AutoResearch trajectory (agent generation and environment execution) scale in very different manners, since all generation shares compute through batching, while each execution occupies its exclusive sandbox and real machine time. As a result, the environment execution dominates the training cost and becomes the bottleneck as trajectories grow. To resolve this tension, we propose World Model RL (WMRL), which replaces environment execution with a world model to remove this bottleneck. Additionally, the world model can be imperfect, as its rewards are corrupted by bias and noise. Therefore, we further equip WMRL with two mitigations, Online Debiasing and Inverse-Variance Denoising, which offset the bias and suppress the noise respectively. Theoretically, we prove that both mitigations of WMRL strictly improve the convergence guarantee. Empirically, WMRL accelerates training by 3-4x on various tasks at different agent scales, while exceeding the performance of standard RL baselines. Moreover, our post-trained 4B and 9B agents outperform much larger open-weight agents of 48B and 120B on held-out benchmarks. Beyond AutoResearch, WMRL also transfers to post-training embodied VLA policies, which demonstrates the generalizability of our method.

Test-time adaptation (TTA) aims to enhance the cross-domain performance of pre-trained models by adapting to unlabeled test data. While most existing TTA methods rely on backpropagation (BP) for finetuning, BP-free methods such as zeroth-order (ZO) methods are more desired in practical on-device scenarios. ZO methods rely only on forward computation, which can largely reduce the complexity and memory overhead of on-device deployment. However, ZO methods suffer from much higher variance compared with first-order methods in estimating the gradient. To address this, we propose an improved ZO method to substantially boost the performance of ZO optimization based TTA. First, we provide an observation to reveal the persistent low-rank Hessian structure of the loss during the adaptation process. Based on this insight, we then propose a loss-landscape curvature-aware zeroth-order (CAZO) method, which leverages a sliding-average estimation of the diagonal Hessian to construct a covariance matrix for anisotropic perturbation sampling. CAZO operates by freezing pretrained weights and optimizing minimal adapter parameters via forward-only passes based gradient estimation, which can substantially reduce the memory overhead compared to BP-based methods. Extensive experiments demonstrate that CAZO significantly outperforms existing TTA methods, achieving state-of-the-art performance while maintaining an excellent balance between accuracy and memory efficiency. Code is available at https://github.com/Hollyming/CAZO.

Continual pre-training (CPT) has been widely adopted as a method for domain adaptation in large language models. However, CPT has consistently been accompanied by challenges, such as the difficulty of acquiring large-scale domain-specific datasets and high computational costs. In this study, we propose a novel method called Test-Enhanced Learning for Language Model Enrichment (TELLME) to alleviate these issues. TELLME leverages the TestEnhanced Learning (TEL) principle, whereby the model's training efficiency is improved using quizzes during training. It integrates this principle with CPT, thereby promoting efficient domain-specific knowledge acquisition and long-term memory retention. Experimental results demonstrate that TELLME outperforms existing methods by up to 23.6% in the financial domain and achieves a 9.8% improvement in long-term memory retention.

Reinforcement learning with verifiable rewards (RLVR) spends most of its compute generating groups of long reasoning trajectories. Recent allocators reduce this cost by assigning budgets to prompts, rollouts, or tokens according to a pointwise notion of difficulty or utility. We identify a statistical mismatch: the unclipped leave-one-out group-relative score gradient is not a sum of independent point contributions, but a second-order U-statistic over pairs of rollouts. Completing one rollout therefore reveals contrast with every other completed rollout, and adaptive endpoint selection changes which pair terms are observable. We introduce PAIR (Pairwise-Aware Inclusion Reweighting), which treats short rollout prefixes as vertices and pair-gradient terms as edges of a contrast graph. A prefix-only predictor estimates correctness and remaining token cost; a convex design chooses positive continuation probabilities under an expected suffix-token budget; and each edge induced by completed vertices is inverse-weighted by its logged joint inclusion probability. Under conditionally independent on-policy rollouts and an unclipped, unstandardized objective, the resulting estimator is design-unbiased for the complete candidate-pair gradient. Across compute-matched RLVR runs on Qwen3-1.7B/4B, PAIR improves average accuracy by +1.2 and +1.4 over the strongest pointwise allocator while using 51% and 52% fewer generated tokens than full-group GRPO. A frozen-population estimator audit confirms that unweighted adaptive selection is biased, whereas pair-inclusion correction recovers the complete-pair target at matched suffix cost.

Sign-language research for resource-constrained languages is often limited by the cost of dense linguistic labels such as glosses, temporal boundaries, and sign order. Broadcast news offers a practical alternative by pairing continuous signing with spoken-language transcripts, but this supervision is weak since text and signing are loosely aligned. Morphologically rich languages such as Turkish add further difficulty, as the same lexical meaning can appear in many inflected forms while some derived forms should remain distinct. We study whether weak transcript-based supervision can pretrain a reusable sign encoder in this setting, where poor text normalization can fragment pseudo-gloss targets and weaken representation learning. Unlike prior pseudo-gloss pipelines designed mainly to improve translation, we test whether the pretrained encoder transfers as a reusable representation for cross-dataset sign spotting. We pretrain on TSL-News, a new Turkish broadcast corpus, using pseudo-gloss labels derived from transcripts rather than manual annotation, comparing rule-based morphological lemmatization with constrained LLM-assisted normalization over a fixed vocabulary. We evaluate the learned representations via cross-dataset sign spotting on a new TSL Spotting Benchmark built from the TSL Dictionary corpus. The LLM-assisted encoder raises top-5 temporal localization mean IoU from 0.235 to 0.465, with 56.2% of examples reaching an IoU of at least 0.50; a frequency analysis suggests this gain is not mainly driven by memorizing frequent pseudo-gloss labels. In a downstream translation check, the same pretraining improves BLEU-4 from 9.60 to 11.04 and ROUGE from 23.48 to 27.43. These results show that loosely aligned broadcast data can provide effective weak supervision for learning sign representations that capture both lexical content and temporal structure.

In large-scale pretraining, the algorithm, architecture, and systems decisions are conventionally made in disconnected stages. A scaling law stage selects an architecture and training recipe, optimizing loss under compute constraints, and a separate systems stage then optimizes the implementation for hardware efficiency. In this work, we develop MOSAIC, which formulates model architecture and systems co-design as an optimization problem. MOSAIC couples a predictive scaling law with a calibrated performance model that estimates Model FLOPs Utilization (MFU), communication cost, memory footprint, and the best parallel layout. We instantiate the framework for sparse Mixture-of-Experts (MoE) language models, where expert count, routing sparsity, and other MoE layer dimensions affect both the loss and systems efficiency. We fit a scaling law on sparse MoE models trained on text data, whose scaling dimensions include the sparsity factor, which is the fraction of model parameters inactive per token in a forward pass. The scaling law sweeps in our work span active parameters from $104$ million to $2.7$ billion and total model sizes reaching $79$ billion parameters. We show that, within the calibrated sparsity range, an efficiency-agnostic model-FLOPs budget admits no interior optimal sparsity. The fitted loss decreases monotonically with sparser models and the compute optimum lies at the upper boundary of the data support. An optimal sparsity in MoE models instead emerges under the cluster's systems constraints, as captured by MOSAIC. Our results argue for a shift towards unified architecture and systems co-design for frontier language model training.

Proteolysis-targeting chimeras (PROTACs) induce protein degradation by recruiting a target protein to an E3 ubiquitin ligase, making degradation a joint outcome of the degrader molecule and its biological context. Although public databases contain thousands of structured molecule-target-E3 records, degradation measurements are available for only a small fraction of them. Existing supervised approaches therefore leave most recorded chemical-biological relationships unused. We introduce DegradeQuery, a context-aware prediction framework that converts these label-missing records into a pretraining signal. Its counterfactual tuple pretraining objective contrasts recorded tuples with alternatives formed by replacing the target, the E3 ligase, or both, enabling the model to learn contextual associations without assigning activity pseudo-labels. The resulting representation is then fine-tuned to predict degradation from the complete molecule-target-E3 context. On the official PROTAC-8K benchmark, DegradeQuery achieves an area under the receiver operating characteristic curve of 0.9065 and an accuracy of 0.8500, outperforming the compared methods. Controlled analyses further show that the improvement is primarily attributable to tuple-level pretraining, can be recovered using only label-missing records, and remains complementary to protein language model representations. These findings demonstrate that incompletely labeled PROTAC databases contain useful relational supervision and provide a practical route for learning context-aware degradation predictors from scarce experimental labels.

Vision Transformers (ViTs) are widely believed to require more labeled data than CNNs for industrial dense prediction. Through controlled experiments on four industrial datasets, we show that the data-efficiency gap stems from pretraining incoherence, which refers to the statistical mismatch between ImageNet-pretrained ViT backbones and COCO-pretrained CNN necks, rather than from inherent self-attention deficits. We characterize the cross-architecture feature gap and propose a lightweight AlignBlock family for pyramid-level feature recalibration. Our core finding empirically identifies a data-efficiency frontier: for domain-proximal scenes with >= 200 samples, Swin-Graft surpasses YOLOv11x (terminal 703-shot: 0.973 vs 0.956 mAP@50); for domain-distant scenes, CNNs retain advantage (hook 141-shot: 0.900 vs 0.600 mAP@50). Grafted neck weights yield up to 2.5x the mAP of a randomly initialized neck.

Offline-to-online (O2O) reinforcement learning aims to leverage policies pretrained on static datasets while improving them through online interaction. However, directly reusing an offline-trained critic can hinder online fine-tuning: as the policy and data distribution change rapidly, value estimates inherited from offline training may become misaligned with the online environment, leading to inaccurate policy improvement and inefficient exploration. To address this problem, we introduce Critic-Free Pretraining: an efficient paradigm that completely abandons the approach of offline critic training, allowing a freshly initialized critic to adapt without inheriting biased estimates. CFP is compatible with various mainstream O2O algorithms and consistently matches or improves upon conventional O2O algorithms across a diverse set of tasks, with particularly pronounced gains on several challenging tasks.

Vision-language-action (VLA) models are a widely adopted paradigm for embodied policies. They excel at efficient closed-loop control but do not explicitly model how physical scenes evolve as a task unfolds. Recently emerging world-action models (WAMs) leverage pretrained video world models to capture spatiotemporal evolution, yet retaining future generation or a large video backbone in the control loop substantially increases inference cost. We introduce World Tokens, an embodied policy architecture built around a World Adapter that bridges visual-language understanding, world-dynamics modeling, and action generation. It uses world modeling during training to enhance the action policy while preserving efficient deployment. Specifically, the World Adapter transforms VLM features into a fixed set of world tokens, which condition a jointly fine-tuned future-video denoiser and simultaneously serve as the action expert's sole visual-language context. This shared conditioning allows gradients from future-video denoising to directly shape the representation used for action prediction, while exclusive routing prevents the policy from bypassing that representation. At deployment, the world-model branch is removed, leaving only the VLM, World Adapter, and action expert, with no online video-model inference. With a 2B backbone and no embodied action pretraining, World Tokens is highly competitive on LIBERO, attains the best reported averages on SIMPLER, substantially improves real-world R1 Pro success over a matched action-only baseline, and generates each action chunk at VLA-level latency.

Face presentation attack detection (PAD) aims to reliably detect a wide range of presentation attacks. While PAD methods achieve strong performance within individual datasets, their performance degrades under cross-dataset evaluation. Variations in sensors or lighting conditions can reduce the effectiveness of detectors from near-perfect to nearly random. Foundation models (FMs) have emerged as a promising alternative because typical PAD datasets, such as the MCIO benchmarks (MSU-MFSD, CASIA-FASD, Replay-Attack, and OULU-NPU), are small relative to the scale used for web-based pretraining. However, existing PAD systems primarily focus on CLIP-based foundation models, while overlooking other FMs with different architectures and training procedures. This study addresses this question by systematically evaluating 32 FMs. Zero-shot prompting achieves performance near chance across model families and scales. The vision encoders, when low-rankadapted (LoRA) with fewer than 1% trainable weights, achieve below 2% intra-dataset ACER in most cases, while cross-dataset ACER is substantially higher. LoRA primarily refines the decision boundary within a dataset, suggesting that pretrained representations and the adaptation dataset play a larger role in cross-dataset generalization than the evaluated lightweight adaptation strategy.

Autoregressive language models align training and use: generation conditions on a clean prompt, and training predicts future tokens from clean left context. Diffusion language models offer parallel denoising, but native dLLM pretraining can randomly corrupt prompt and continuation tokens together, weakening the clean-prefix interface needed for prompt-conditioned generation. We identify this mismatch for prompt continuation and propose PCD (Prefix-Conditioned Diffusion), a pretraining objective that combines AR prefix supervision with no-shift suffix denoising. At the training-objective level, PCD changes the attention mask, corruption mask, and label construction in continued pretraining; it does not require an autoregressive decoder, verifier, or new inference mode. By supervising the clean-prefix side autoregressively and applying diffusion only to the unknown continuation, PCD makes the local training interface resemble how block-diffusion models are queried at evaluation time. We further separate intra-sample prefix conditioning from inter-sample objective mixing, allowing us to identify the local alignment signal separately from the optional batch-level mixing knob. Across LLaDA2-Mini and Qwen-1.7B backbones, PCD consistently improves over same-family native dLLM stable baselines, reaching a 4.2% relative gain on the main LLaDA2-Mini six-benchmark average (+2.56 points) and a 14.2% relative gain in the primary Qwen mechanism comparison (+4.86 points). These results suggest that aligning the pretraining context distribution with prompt-conditioned generation can recover a measurable part of the dLLM continuation gap without changing inference.

Data-Free Knowledge Distillation (DFKD) transfers knowledge from a pretrained teacher model to a compact student model by synthesizing semantically informative data, eliminating the need for access to the original training dataset. Existing DFKD methods rely heavily on architecture-specific statistical priors (e.g., Batch Normalization statistics) to guide data synthesis, however, such architecture-dependent priors are often absent in modern architectures such as Vision Transformers (ViTs), resulting in degraded semantic quality of the synthesized data and consequently catastrophic performance degradation. In this paper, we propose \emph{UniDFKD}, a unified data-free knowledge distillation framework that replaces architecture-specific statistics with explicit, architecture-agnostic semantic priors. \emph{UniDFKD} governs the entire synthesis-distillation pipeline along three dimensions: (1) Categorical Semantic Conditioning (CSC) defines \emph{what} to synthesize by persistently modulating the generator with language-derived embeddings to capture semantic diversity; (2) Spatial Semantic Anchoring (SSA) dictates \emph{where} evidence belongs by anchoring the teacher's spatial attributions to a Gaussian prior; and (3) Spatial Semantic Distillation (SSD) controls \emph{how} knowledge is transferred by explicitly aligning teacher-student spatial evidence alongside predictions. Extensive experiments across CNNs and ViTs demonstrate that UniDFKD establishes a new state-of-the-art, outperforming existing methods by an average absolute margin of over 20% in both homogeneous and heterogeneous settings.

Magnetic field strength is a major source of domain shift in magnetic resonance imaging (MRI), affecting signal-to-noise ratio, tissue contrast, spatial detail, and the visibility of anatomical boundaries. The MRIxFields 2026 challenge investigates this problem through cross-field MRI translation across acquisitions at 0.1T, 1.5T, 3T, 5T, and 7T. Its three tasks, Any-to-7T, 0.1T-to-High, and Any-to-Any synthesis, require the generation of target-field image characteristics while preserving subject-specific anatomy. This problem is particularly challenging because paired acquisitions of the same subject across multiple field strengths are rarely available for training. We propose a 3D unpaired cross-field MRI translation framework based on field-conditioned content-style pretraining. The proposed framework first learns controllable field-to-field translation across all available field strengths by disentangling anatomical content from field-dependent contrast characteristics. The pretrained backbone is then adapted to task-specific target domains. Our model comprises a 3D content encoder, a 3D style encoder, a field-conditioned style generator, an AdaIN-modulated decoder, and a multi-field discriminator. Adversarial learning encourages realistic target-field appearance, while cycle-consistency, identity, content, style, and diversity constraints promote anatomical fidelity and controllable translation. We evaluate the proposed method on MRIxFields data spanning five field strengths and three MRI modalities. Experiments on paired test data demonstrate that the framework can adapt to the three challenge settings while preserving three-dimensional anatomical structure in the synthesized volumes. The implementation code is publicly available at https://github.com/Idea89560041/3D-MRI-Field-Translation.

Node ranking is a fundamental problem in graph information retrieval, measuring the relative importance of nodes and supporting a wide range of applications such as influence analysis, recommendation, and graph-based retrieval augmented generation. However, exact computation of graph-based ranking measures is often computationally prohibitive at scale. Existing GNN-based ranking methods provide scalable approximations, but they are typically tailored to individual ranking criteria and require retraining for each downstream task, which limits their transferability and efficiency. Recent graph pre-training approaches aim to enable knowledge transfer across tasks, yet their learning objectives are largely misaligned with node ranking, resulting in suboptimal adaptability to ranking-oriented applications. To address these limitations, we propose PreGress, the first ranking-native pre-training and prompting framework for supporting a wide range of node ranking tasks. PreGress performs multi-task pre-training using our carefully designed objectives, including degree centrality prediction and attribute reconstruction, to jointly capture structural and attribute information. To support heterogeneous ranking criteria, we design lightweight, task-specific prompt modules that adapt a frozen ranking backbone to downstream tasks without full retraining. Experiments on six public graphs and two real-world query-to-item benchmarks---Yelp2018 and MovieLens-100K---together with a controlled five-criterion graph-access study demonstrate strong ranking quality with low task-specific state overhead.

Pretraining LLMs on artificial languages ("pre-pretraining") is a technique that could reportedly increase token efficiency by 33%, i.e., save up to 33% of training tokens needed to reach a certain performance. We validate this prior result for English on a larger set of natural languages across four language families, using two different tokenizers and varying model sizes. We also relate the observed gains (or losses) in token efficiency to quantified linguistic properties of the languages, such as sentence length, morphological richness, and features of dependency syntactic trees (tree depth, number of children, number of crossing dependencies). Our empirical results indicate that the reported gains depend heavily on the experiment setup and the choice of random seed, although we can confirm the trend of stable gains with 128-Dyck pretraining of small models with the Llama tokenizer for most of the examined languages. On a general note, we argue that multiple training runs should be carried out at least for a subset of experiments to avoid the community adopting unstable approaches.

Field-based agricultural computer vision is important for precision agriculture, yet it largely depends on expensive annotations and costly adaptation of large pretrained models. We introduce AgriField-40K, a field-centric dataset curated from 17 public resources and covering diverse crops, weeds, and field conditions. Building on this, we present AgriMAE, a parameter-efficient continual pretraining baseline that adapts a masked autoencoder pretrained on natural images by training only lightweight adapters. We further explore semantic feature reconstruction as an alternative pretraining objective and evaluate transfer across multiple tasks. AgriMAE consistently improves downstream performance and can match or even outperform full fine-tuning while using up to $9\times$ fewer trainable parameters, showing that AgriField-40K is a practical resource for continual pretraining in agricultural vision. Project page: https://dtu-pas.github.io/agrifield40k/

Stochastic cortical self-reconstruction (SCSR) enables personalized mapping of gray matter atrophy, a hallmark of neurodegenerative disorders such as Alzheimer's disease (AD), onto high-resolution cortical surfaces. Unlike conventional normative modeling approaches, which typically operate at a coarse regional level and remain inherently constrained by the covariates included during training, SCSR estimates an individualized healthy reference directly from the observed cortical thickness at the vertex level. This allows the detection of subtle, subject-specific deviations from healthy cortical shape. In this work, we investigate the generalization and transferability of SCSR, originally trained on UK Biobank (UKB) data, to an independent Chinese population dataset. Specifically, we evaluate the ability of SCSR-derived Z-scores to discriminate between healthy scans, individuals with mild cognitive impairment (MCI), and patients with AD, while also assessing model robustness across the lifespan. We compare four training strategies: direct application of the UKB-trained model, fine-tuning on Chinese data, training from scratch, and joint training on UKB and Chinese cohorts. As reconstruction backbones, we consider both a multilayer perceptron (MLP) and a Spherical UNet (SUNet). Our results demonstrate that SCSR provides robust detection of cortical atrophy in the Chinese population across all evaluated models. The highest discriminative performance was achieved by the fine-tuned SUNet model (average pairwise AUC = 0.848), followed closely by the UKB-trained SUNet. Moreover, reconstruction errors remained low across the lifespan, even when the training population exhibited a substantially narrower age distribution, indicating strong cross-population transferability.

Large-scale pretrained time-series models achieve strong results through large-scale pretraining and task-agnostic representation learning, but they rely on abundant, diverse data that industrial and scientific domains often lack. We therefore propose the FSD-RM (Family of Small-Data Representation Models) paradigm as a practical alternative for limited, domain-specific telemetry. Rather than relying on large-scale pretraining, we focus on capacity-controlled representation learning using established encoder architectures (CNN1D, LSTM, GRU, Transformer), selected for their suitability in small-data settings and interpretability. These encoders are trained unsupervised on multivariate telemetry data and integrated into a two-stage pipeline for downstream lifetime prediction. To systematically examine architectural trade-offs under data constraints, we employ \textbf{dimension-aware neural architecture search (NAS)} to jointly optimize model capacity and input dimensionality. Experiments on cryocooler telemetry show that the proposed approach achieves competitive predictive performance while reducing training cost and model complexity. The contribution lies in combining established representation learning techniques within a coherent, NAS-driven framework tailored to small-data regimes, with explicitly defined parameter settings and design choices. The results indicate that effective representation learning can be achieved without large-scale pretraining when appropriate inductive bias and capacity control are applied.

Pansori is a traditional Korean vocal genre whose mode system (jo) is defined not by scale alone but by the entanglement of pitch collection, microtonal ornament (sigimsae), and vocal timbre. In this study, we introduce a 46-hour frame-level pansori mode annotation, expert-labeled across all five canonical batang, and evaluate four complementary input representations (mel spectrogram, F0 contour, MIDI piano roll, and a multi-cultural SSL encoder) under two split strategies designed to detect shortcut learning. Across the three well-represented modes, performance degrades by only 2.1--3.6 points of F1 when entire works are held out, indicating that the models learn mode-relevant features rather than memorizing repertoire. Per-class results further show that source separation removes the percussion cue on which changjo depends, and that generic multi-cultural pre-training fails specifically on the Ujo--Gyemyeonjo distinction. Qualitative analysis of cross-modal disagreement recovers musicologically documented phenomena and agrees with published score-based analyses of modern changjak pansori.

Routine CT interpretation is inherently comprehensive, capturing incidental findings across the entire scan volume. 3D CT foundation models could assist this process by providing generalizable representations of anatomy and pathology. To evaluate their diagnostic breadth, we benchmark ten frozen CT encoders across three cohorts of thoracic CT scans, including an unseen internal clinical dataset, using $k$-nearest neighbors, zero-shot prompting, and linear probing. We find no universal state-of-the-art, with rankings fluctuating significantly depending on the evaluation context. While models combining fine-grained image tokenization with vision-language alignment generally perform best, a lightweight supervised encoder remains highly competitive, demonstrating that explicit labels can effectively substitute for scale. Crucially, rather than model architecture, we observe that the primary determinant of performance is a physical bottleneck: a finding's detectability scales with its contrast against surrounding tissue and its spatial extent. Through controlled within-organ comparisons, we empirically demonstrate that widespread or high-contrast abnormalities, such as devices and effusions, are reliably recovered. Conversely, small, low-contrast focal lesions remain a persistent challenge across all evaluated encoders. We attribute this to the inherent limitations of globally pooled embeddings, suggesting that accurately representing small, low-contrast structures will require region- or lesion-level pretraining.

Mainstream World-Action Models (WAMs) adapt pretrained video generation models (VGMs) for robot control, transferring their learned dynamics prior for action prediction. These VGMs are typically trained in a variational autoencoder (VAE) latent space. However, the VAE latent space is optimized for pixel reconstruction, which rewards fine appearance detail and leaves the action prediction fragile under visual shifts. Recent works build WAMs in semantic latent space, which are more robust to appearance shifts. However, these models cannot leverage the large-scale VGM pretraining that exists only in VAE space. To overcome this dilemma, we propose Robust-WAM, a general post-training method for video-generation-based WAMs that preserves the VAE-based generative path and adds a lightweight semantic foresight alignment objective on the action stream. This retains the large-scale VGM pretraining while grounding actions in appearance-invariant dynamics that stay reliable under illumination shifts and other visual out-of-distribution conditions. Specifically, we employ learnable query tokens to bring future-scene semantics into the action stream by aligning their output hidden states with the semantic foresight of future ground-truth frames. To establish the temporal correspondence between each query and the future step it describes, we give it the positional encoding of the matching action tokens. Experiments on out-of-distribution generalization simulation benchmarks and a real-robot setup show that our Robust-WAM consistently improves the success rates of multiple WAM baselines without sacrificing in-distribution performance.

Scientific images are essential for communicating experimental observations, quantitative evidence and conceptual knowledge. Unlike natural images, their quality depends on both visual clarity and scientific informativeness, making assessment challenging. In this work, we present SciQNet, a two-stage multimodal adaptation framework for scientific image quality assessment. The first stage performs domain-adaptive pretraining on scientific document images and the second stage conducts task-specific fine-tuning with joint scoring and understanding supervision. For scoring-oriented supervision, we combine instruction tuning with a Huber loss derived from rating-word logits, while understanding-oriented supervision is formulated as multiple-choice visual question answering. Experiments show that using a 40% stratified subset of the domain-adaptive data gives the best performance among the evaluated pretraining fractions, suggesting that pretraining-data relevance may be as important as pretraining-data scale. The final model achieves an SIQA-S score of 92.21, an SIQA-U score of 47.38 and a combined score of 69.80. This work presents our solution to the ICME 2026 Scientific Image Quality Assessment Challenge, which ranked 2nd in the scoring track.

Dialectal variation remains a major challenge for multilingual language models. Perturbation-based continued pre-training (CPT) has emerged as a promising approach to improving robustness, yet existing work largely evaluates individual perturbation strategies in isolation and provides limited insight into why they work. We present a systematic study of perturbation-based CPT for multilingual dialect robustness in LLMs, comparing six training conditions across nine German, Italian, and Arabic dialect tasks. Perturbation-based CPT, especially character-noised CPT, consistently improves zero-shot dialect robustness while largely preserving standard variety performance. More importantly, we show that methods with similar downstream performance induce distinct mechanisms of robustness, exhibiting different patterns of language model adaptation, representational alignment, and prediction repair. Our results provide a more complete understanding of how synthetic surface variation improves robustness and offer practical guidance for selecting CPT strategies in multilingual and dialectal settings.

Recent advances in machine learning have enabled training of wireless foundation models, which aim to support tasks such as channel estimation, beam prediction, and localization based on wireless signals. Existing wireless foundation models typically pretrain on channel tensors using masked reconstruction over subcarriers, antennas, or time but ignore the physical characteristics of wireless propagation. In this work, we propose to instead use multipath propagation as the fundamental pretraining object. We present MultiPathFormer, an autoregressive foundation model that represents each transmitter-receiver link as an ordered sequence of continuous-valued path tokens and pretrains with next-path prediction. We introduce an Environmental RAG (retrieval-augmented generation) mechanism and a first-path codebook on top of the transformer backbone, leveraging environment knowledge to improve path statistics estimation like delay and power by up to 59%. MultiPathFormer pretrained on 27 environments transfers to unseen users and, after scenario-specific fine-tuning, outperforms training the corresponding models from scratch in new environments. Across downstream tasks, it outperforms SOTA channel-based foundation models, achieving 5.57 m mean localization error, 0.914 top-3 beam accuracy, 0.994 line-of-sight classification accuracy, and 0.561 channel estimation NMSE. These results show that path-level pretraining can learn reusable representations of wireless propagation.

Decoding continuous motor trajectories from neural activity is essential for developing practical brain-computer interfaces (BCIs). However, current neural decoders are constrained by the limited scale and heterogeneity of neural recordings. In contrast, behavioral data can be collected more readily and at substantially larger scale from humans, animals, simulations, and robotic systems. Here, we introduce NeuroPB, a framework that scales neural decoding by transferring knowledge from pretrained behavioral representations. NeuroPB first pretrains a motor encoder on large-scale motor behavior data and then aligns neural activity with the resulting behavioral representation space using a limited set of paired neural-behavioral recordings. A neural encoder and lightweight motor decoder are subsequently optimized to reconstruct continuous movement from the aligned neural representations. Across multiple macaque motor datasets, behavioral pretraining improves trajectory decoding, including an 11% $R^2$ increase on center-out and 8% on random-target compared with training the motor encoder from scratch. Notably, pretraining on robotic trajectories achieves performance comparable to pretraining on macaque trajectories, demonstrating that transferable kinematic structure is shared across biological and artificial models. Moreover, decoding performance improves as the scale and diversity of robotic pretraining data increase, when the amount of neural data is fixed. Pretraining also enhances generalization across recording sessions, subjects, and motor tasks, with only 10% calibration needed to match training from scratch. Overall, these results establish behavioral pretraining as a scalable source for neural decoding and provide a promising route toward high-performance and calibration-efficient BCIs under limited neural data.

Existing studies on self-supervised learning for white-box networks typically decouple the derivation of white-box networks via optimization algorithms from self-supervised learning paradigms. In this work, we instead revisit the two components from a joint perspective. The LeJEPA-based self-supervised framework assumes an isotropic Gaussian distribution as the optimal embedding distribution for downstream tasks, which is conceptually equivalent to the expansion term $R(Z)$ in the sparse rate reduction objective guiding white-box Transformer optimization. Building on this observation, we use the LeJEPA self-supervised paradigm to optimize $R(Z)$, and derive the remaining terms $R^{c}(Z\mid U_{[K]})+λ\lVert Z\rVert_{0}$ via the alternating direction method of multipliers (ADMM) into an attention-only Transformer that dispenses with the ISTA structure or MLP layers of the original design. Experimental results demonstrate that our attention-only white-box Transformer achieves classification accuracies of $88.88\%$ on CIFAR-10 and $63.54\%$ on CIFAR-100 at the Base scale under the LeJEPA self-supervised paradigm, while the original white-box Transformer CRATE achieves classification accuracies of $89.18\%$ on CIFAR-10 and $63.56\%$ on CIFAR-100. Our model achieves competitive performance while reducing the parameter count by roughly $31\%$. Beyond the white-box setting, we further investigate standard ViTs and find that replacing all MLP blocks with ReLU activations under knowledge distillation removes approximately 66% of the parameters while preserving competitive accuracy, motivating further investigation into the potential redundancy of MLP modules in standard ViT architectures.

Pre-pretraining language models (LMs) on symbolic data can accelerate and improve natural language acquisition. However, existing pre-pretraining tasks, such as Dyck and procedural algorithms, rely on narrow primitives that fail to capture the expressive capacity of natural language. Moreover, prior studies remain restricted to relatively small token budgets, offering limited insight into skill emergence and representational dynamics. To address these limitations, we propose logic pre-pretraining (Logic-PPT) as a principled initialization strategy, leveraging formal derivations to impart richer structural and linguistic biases. Formal derivations require abstract mechanisms that are central to natural language, simultaneously binding variables, connecting quantifiers and relational dependencies, and composing predicate-argument structures over long contexts. Scaling our evaluation to a 100B-token regime, logic pre-pretraining substantially accelerates skill acquisition in LMs, achieving 80% accuracy on linguistic tasks with 36B fewer tokens than standard initialization, and outperforming alternative pre-pretraining baselines. Mechanistically, formal derivations induce persistent structural reorganization, distinctively characterized by a lower-rank, spectrally concentrated representation space. Crucially, we show that this internal geometry enables improved model compressibility via pruning, matching the dense baseline performance even at $\approx$33% sparsity.

A companion paper studies when activation patching and weight-space ablation agree, inside an idealized model where a conditional computation is carried additively through a residual stream. For the one composition in that model where two carriers are architecturally dependent, an attention head and its own layer's normalization-MLP composition, it derives an exact first-order interaction formula, zero when only the MLP is ablated and second-order bounded when the head is also ablated. That result is confined to a single residual block and checked only on small transformers on a synthetic task. This paper extends the result past both limits. First, the interaction from ablating carriers spanning several layers decomposes exactly into same-block terms, one per touched layer, plus a cross-layer remainder on which the decomposition makes no claim of smallness. Second, we isolate that remainder exactly, for two layers, as a double integral of a mixed second derivative, and name the missing ingredient needed to bound it: a Jacobian bound for the attention sub-block. We derive this bound in closed form and verify it, without a single violation, against Qwen2.5-1.5B-Instruct's real weights, though we do not yet chain it across layers. We also give, in closed form, the curvature constant the companion paper's bound leaves unexhibited. Third, on that same model, we search for and find an emergent circuit for indirect object identification, never designed into it, using the original activation-patching method for this task, and test collapse, dissociation, and interaction on it. The result is mixed: a shared carrier emerges across all five tested instances, collapse and dissociation hold on most but not all, and a nonzero interaction is measurable on three of five, at layer pairs outside the same-block case the companion theorem covers.

Pretraining has shown strong potential for learning transferable representations, yet it remains underexplored for electron-density-based molecular learning. Electron density provides a continuous three-dimensional description of molecular electronic structure, capturing both local spatial patterns and global physical quantities. This raises a key question: can electron-density fields be used for self-supervised pretraining to learn a shared representation that transfers across diverse electronic-structure-related tasks? We propose ED-DiT, a physics-guided Diffusion Transformer for self-supervised pretraining on electron-density point clouds. ED-DiT learns reusable representations by reconstructing corrupted and partially masked log-density fields across diffusion noise levels. An electron-number consistency constraint is further introduced to preserve the total electronic mass. The pretrained encoder can be adapted to property prediction, open-/closed-shell classification, molecule-electron-density retrieval, and molecule-conditioned electron-density prediction. Experiments on six EDBench tasks show that ED-DiT consistently outperforms the same architecture trained from scratch, especially under limited supervision. For molecule-conditioned electron-density prediction, it reduces RMSE from 2.2474 to 1.3753 and surpasses the available baseline. With only 10% labels, it improves orbital energy prediction RMSE from 0.0293 to 0.0138. These results demonstrate the effectiveness of physics-guided electron-density pretraining for learning transferable molecular representations.

Accurate detection of sedentary behavior is important for studying health risks related to prolonged sitting, but posture-based classification remains challenging with wearable sensors, especially at the wrist. We study whether a deep learning model trained on hip-worn accelerometer data can transfer to wrist-worn accelerometer data for sitting versus non-sitting classification. We use CHAP, a CNN-BiLSTM model originally developed for hip accelerometers, and evaluate its zero-shot performance on wrist data as well as its adaptation through finetuning with varying amounts of labeled wrist data. Experiments are conducted on the iWatch dataset with ground-truth posture labels derived from wearable cameras. The hip-trained model performs strongly on hip data without retraining, but accuracy drops on wrist data due to sensor placement shift. Finetuning CHAP provides consistent advantages over transformer models trained from scratch. These findings suggest that hip-based pretraining provides a useful starting point for wrist deployment, while highlighting the need for wrist-specific adaptation to handle higher signal variability.

Latents from vision foundation models (VFMs) are semantically rich and well suited for visual understanding. Recent representation autoencoder methods such as RAE have shown that they can provide promising latent spaces for image generation. However, VFM latents remain difficult to model directly: DiT-generated latents exhibit spectral mismatch with encoder latents, especially in high-frequency components. Our channel-wise spectral analysis further reveals that these high-frequency components are diffusely distributed across latent channels and entangled with semantic information, making the latent space difficult for DiT to model. To address these challenges, we propose SPAE, latent adaptation framework for generation. Specifically, SPAE employs a compact bottleneck to distill stable semantic information while suppressing high-frequency components, thereby improving the alignment between DiT-generated latents and encoder latents. In addition, we apply a channel-wise masking strategy to promote the decoupling of semantic information and high-frequency details across bottleneck channels. Experiments show that SPAE achieves a favorable balance among visual understanding, generation quality, and reconstruction fidelity.

The rapid growth of single-cell transcriptomic data has enabled the development of foundation models pretrained primarily by reconstructing masked expression values. This objective encourages these models to learn gene dependencies but does not directly optimize whole-cell representations, which are essential for many downstream tasks. To bridge this gap, we propose a contrastive pretraining framework that learns cell representations through complementary transcriptomic views. Since standard contrastive learning is not readily applicable to single-cell pretraining, we introduce specific adaptations along three dimensions --- co-expression-guided gene partitioning, expression-aware contrast-set construction, and competence-gated contrastive onset. Specifically, we first construct two complementary views of each cell by partitioning its genes according to their co-expression structure. Then, to prevent the model from using gene-set identity as a shortcut, we construct hard negatives by permuting expression values while keeping gene identities unchanged. Finally, we introduce a competence-aware controller to determine how the contrastive objective is applied. Experiments on cell-type annotation and gene regulatory network inference demonstrate competitive transfer under the evaluated protocols. In the six-network GRN evaluation, our method records the highest mean AUROC and AUPRC point estimates among the compared variants, while the highest-scoring variant differs across individual networks. These results establish complementary-view contrastive learning as an effective direction for single-cell pretraining beyond gene reconstruction.

Despite the widespread adoption of foundation models as feature extractors for medical imaging, relatively little is understood about how different pretraining strategies influence the transferability of learned representations to weakly supervised ophthalmic imaging tasks. We investigate this question in ultra-widefield (UWF) retinal imaging by evaluating foundation model representations within a patch-based multiple instance learning (MIL) framework for disease classification on UWF images. We compare Vision Transformer encoders pretrained with supervised, Masked Autoencoder (MAE), and self-distillation objectives, while keeping the downstream aggregation architecture unchanged. Within a controlled comparison of ViT-B encoders pretrained on ImageNet-1k, the choice of pretraining objective substantially influenced frozen representation transfer, with supervised and self-distillation-based models outperforming MAE. A contemporary DINOv3 model pretrained at a larger scale achieved the strongest overall performance, with a quadratic weighted kappa of 0.863 for five-class diabetic retinopathy grading, comparable with DINOv1. Attention analysis further revealed distinct patch-aggregation behaviours associated with the different pretrained representations, while partial fine-tuning substantially reduced the performance gap for MAE. These findings suggest that pretraining strategy influences both representation transferability and the subsequent aggregation of patch-level evidence within MIL, resulting in differences in downstream classification performance.

Deep research agents are often trained on expensive, environment-grounded tool-use trajectories that require repeated retrieval, document inspection, and report evaluation. We introduce Deep Research Pretraining (DRP), an offline framework that derives predictive navigation supervision from naturally occurring evidence structures. Given a citation-bearing or hyperlinked passage, DRP constructs a proxy research objective, recovers linked evidence and graph-related alternatives, and converts them into search-open-write trajectories. This teaches models what to search for, which documents to inspect, and how to synthesize evidence, without a live retrieval environment or executed policy rollout. We instantiate DRP on scholarly citation graphs (DRP-Paper) and Wikipedia hyperlinks (DRP-Web), continually pretrain separate Qwen3-14B-Base models on 1B tokens, and fine-tune them on controlled fractions of 13K agent trajectories. Across five independently sampled subsets at each low-data budget, both variants consistently outperform matched no-DRP models on DeepResearch Bench. With one quarter of the SFT data, DRP-Web even surpasses a fixed no-DRP full-data checkpoint, with gains transferring to ResearchQA, WebWalkerQA, and SimpleQA. Starting from matched low-data SFT checkpoints, the DRP-Web advantage also persists through subsequent agentic RL. Source-matched and evidence-mismatch controls indicate that these improvements arise from evidence-conditioned navigation rather than domain exposure or agent-format imitation. DRP thus provides a promising complementary approach to trajectory-based agent training.

Logical rule induction seeks interpretable rules that transfer across propositional schemas. This requires respecting symmetries: atom naming, example order, polarity flips, and label swap. Enforcing exact symmetry by construction lets one trained inducer scale beyond its training schemas. Our central contribution is a canonical export that decodes a discrete rule from literal scores. It needs no retraining and is exactly equivariant whenever those scores respect the symmetries. We instantiate it on the Neural Rule Inducer, a disjunctive-normal-form (DNF) foundation model that natively respects only example order. We restore the remaining symmetries through architecture, inference, and training. On synthetic stress tests, accuracy on the support labels stays stable at much larger schemas, and rule fidelity on fresh inputs remains above the unmodified model. On real data, accuracy improves most on larger schemas. The exported rule is exact on synthetic full-group tests and on schema-valid real-data tests. This is a mathematical property of the export rather than of a specific model, and we validate it empirically only on the NRI. Enforcing symmetry by construction turns this small-data pretrained model into a reusable, interpretable inducer that transfers to larger schemas.

Neural Combinatorial Optimization (NCO) techniques have emerged as a highly efficient alternative to traditional exact algorithms for solving routing problems such as the Traveling Salesman Problem (TSP). However, the generalization capabilities of these Reinforcement Learning-based models are severely hindered when scaling to high-dimensional instances. This issue has been mitigated in other domains, like computer vision and natural language processing, by adopting a self-supervised pre-training strategy. Nevertheless, its application to routing graphs, which lack complex topological attributes beyond 2D spatial coordinates, remains a challenge. In this paper, we propose a geometric self-supervised pre-training framework specifically designed to capture spatial invariance and global relative distance distributions. By applying isometric transformations, such as rotations and axial reflections, the model learns robust structural representations prior to the policy optimization phase. Empirical results demonstrate that this strategy consistently outperforms models trained from scratch (baselines), achieving a 7.23% improvement in tour length for massive zero-shot extrapolation scenarios (TSP1,000). Furthermore, the proposed model exhibits remarkable computational efficiency, delivering speedups of up to two orders of magnitude over the exact solver Concorde at massive scales. The source code and pre-trained models are publicly available at https://github.com/davidaguadocosano/TSP-GeoPretrain.git.