Model Pretraining

The current pretraining paradigm for large language models relies on massive compute and internet-scale raw text, creating a significant barrier to foundational research. In contrast, biological systems demonstrate highly sample-efficient learning through multi-timescale processing, such as the functional organization of the frontoparietal loop. Taking this as inspiration, we introduce HRM-Text, which replaces standard Transformers with a Hierarchical Recurrent Model (HRM) that decouples computation into slow-evolving strategic and fast-evolving execution layers. To stabilize this deep recurrence for language modeling, we introduce MagicNorm and warmup deep credit assignment. Furthermore, instead of standard raw-text pretraining, we train exclusively on instruction-response pairs using a task-completion objective and PrefixLM masking. Serving as an empirical existence proof of efficient pretraining, a 1B-parameter HRM-Text model trained from scratch on only 40 billion unique tokens and $1,500 budget achieves 60.7% on MMLU, 81.9% on ARC-C, 82.2% on DROP, 84.5% on GSM8K, and 56.2% on MATH. Despite utilizing roughly 100-900x fewer training tokens and 96-432x less estimated compute than standard baselines, HRM-Text performs competitively with 2-7B parameter open models. These results demonstrate that co-designing architectures and objectives can radically reduce the compute-to-performance ratio, making pretraining from scratch accessible to the broader research community.

Learning to generalise from limited data is a fundamental challenge for both artificial and biological systems. A common strategy is to extract reusable structure from abundant unlabelled data, enabling efficient adaptation to new tasks from limited labelled data. This two-stage paradigm is now standard in modern training pipelines, where pretraining is followed by fine-tuning or linear probing. We provide an analytical model of this process: structure extraction is formalized as principal component analysis on unlabelled data, and downstream learning as linear regression on a separate labelled dataset. In the high-dimensional regime, we derive exact expressions for training and generalisation error showcasing their dependence on representation dimensionality, unlabelled and labelled sample sizes, and task alignment. Our results show that pretrained representations strongly influence downstream generalisation, and we characterize the optimal representation size as a function of task parameters: with abundant pretraining data but scarce downstream data, maximally compressed representations are optimal, whereas with limited pretraining data, higher-dimensional representations generalise better. Furthermore, we establish an exact trade-off between pretraining and supervision, quantifying how much unlabelled data is required to replace a single labelled sample. Beyond our idealised model, we observe similar phenomenology in autoencoders and pretrained LLMs. Altogether, we highlight that optimising representation size is critical, giving conditions for when compression during pretraining improves generalisation.

The success of self-supervised learning (SSL) in vision and NLP has motivated its rapid adoption for time series. However, research has focused primarily on Generative paradigms and forecasting tasks, leaving the broader utility of learned representations unquantified. We establish a controlled framework to evaluate the "pre-training dividend": the value added by SSL across diverse temporal tasks. We systematically compare Generative paradigms against Latent Alignment architectures, introducing adaptations of LeJEPA and DINO for time series. These adaptations utilize Discrete Wavelet Transform (DWT) augmentations to enforce invariance to local fluctuations. Our analysis reveals that the pre-training dividend is highly asymmetric: SSL yields gains of up to 375% for anomaly detection and classification, yet remains marginal for forecasting. We demonstrate that representational utility is non-universal, governed by a precision-invariance trade-off where the specific signal resolution required by the task must align with the objective. Finally, we show that representation quality is largely independent of data origin and saturates at moderate architectural depths, suggesting a path to scaling via massive synthetic generation. Our code is available at: https://github.com/noammajor/Models

We investigate data filtering for large model pretraining via new scaling studies that target the high compute, data-scarce regime. In spite of an apparently common belief that filtering data to include only high-quality information is essential, our experiments suggest that with enough compute, the best data filter is no data filter. We find that sufficiently trained large parameter models not only tolerate low-quality and distractor data, but in fact benefit from nominally ``poor'' data.

Deep learning methods have demonstrated promising results in predicting BI-RADS scores from mammography images. However, the interpretation of these images can vary, leading to discrepancies even among radiologists. Given the inherent complexity of mammograms, training classification models solely on image labels often yields limited performance. To address this challenge, we curated 2313 mammogram images and their corresponding captions from two mammography atlases. Our proposed approach employs a multi-modal model that uses a pretrained PubMedBERT as the language component. By training this model on image-text pairs with contrastive learning, we enable the vision encoder to absorb the rich information contained in the captions, thereby improving its understanding of mammography findings. We then fine-tune the vision encoder on two datasets for BI-RADS prediction, achieving superior performance compared with models trained without this pretraining, particularly when labeled samples are scarce. The improvement in the 3-class average F1 score ranges from +1% to +14%: a +1% increase with 40K training samples, and a +14% increase with 1K samples. Furthermore, our experiments reveal that 2K image-text pairs from mammography atlases can be more informative than 2K labeled samples for label prediction, with an average margin of +1.1% when more than 10K training samples are available. Overall, our work provides a vision-language model for mammography and highlights the value of textual information from mammography atlases. In addition, we publicly release preprocessed mammography images of the TEKNOFEST dataset. The training code, pre-trained model weights, data extraction scripts, and the released dataset are publicly available at: https://github.com/igulluk/MAM-CLIP

Synthetic PET images are valuable for quantitative imaging workflow development, scalable virtual imaging trials, and deep learning model training, but conventional physics-based simulation approaches are computationally intensive, limited in anatomical variability, and often fail to capture heterogeneous PET uptake. This study developed a pretrained domain-adapted diffusion (PAD) model for anatomy-conditioned PET synthesis from uniform organ activity maps. PAD adopts a natural-image pretrained text-to-image decoder with an upstream conditioning encoder and a downstream PET-domain adapter. A two-phase training strategy was used, with the first phase learning coarse uptake distributions and the second refining local image details. Uniform organ activity maps were generated from CT-based segmentations by assigning each organ its mean uptake from the paired PET image. Evaluation included quantitative accuracy, noise assessment, radiomic analysis, tumor segmentation performance, and a human observer study. PAD-generated images achieved high quantitative accuracy, with concordance correlation coefficients above 0.92 between organ mean SUVs and assigned activity values. The synthesized images showed noise levels and texture characteristics similar to target PET images and produced comparable tumor segmentation performance. In a two-alternative forced-choice observer study, four readers achieved approximately 50% accuracy, indicating visual indistinguishability between synthesized and target images. PAD also generated realistic PET images from XCAT-derived activity maps, demonstrating compatibility with phantom-based anatomical priors. Overall, PAD provides a diffusion-based framework for generating clinically relevant heterogeneous PET images from uniform organ activity maps derived from clinical segmentations or digital phantoms, supporting data augmentation and downstream imaging studies.

Fine-tuning adapts a pre-trained model to downstream tasks using a small amount of labeled data. Low-Rank Adaptation (LoRA) is an efficient fine-tuning method that reduces memory and computation costs while often achieving performance close to full fine-tuning. Despite its widespread use, the theoretical behavior of LoRA is not yet well understood. In this paper, we study LoRA in a simple linear regression setting and compare its excess risk with that of full fine-tuning. Our analysis identifies regimes in which LoRA achieves lower excess risk than full fine-tuning in both overdetermined and underdetermined settings. Specifically, our theory predicts that LoRA can outperform full fine-tuning when the difference between the pretraining and the downstream tasks is effectively low-rank. We further show how the choice of LoRA rank affects generalization performance, explaining why using a very small rank can improve test accuracy in certain settings, even though it limits model expressivity. Finally, we support our theoretical results with experiments on practical tasks, suggesting that the identified tradeoffs and insights extend beyond linear regression.

Task vector composition has emerged as a promising paradigm for editing pre-trained models, enabling model merging through addition and unlearning through subtraction. Fine-tuning in the tangent space of a pre-trained model (linear fine-tuning) has proven effective, as it produces task vectors that are naturally disentangled and resistant to interference. However, linearized models suffer from limited expressivity during training and incur higher computational costs at inference time, which restrict their practical applicability. In this work, we bridge the gap between linear and standard non-linear fine-tuning. We show that linearity with respect to weight perturbations, a property defined in parameter space, can be enforced through constraints in activation space during training. Concretely, we distill hidden representations from a curvature-regularized linearized teacher into a non-linear student trained via conventional fine-tuning. We find that the resulting model inherits key properties of linearized models for task arithmetic, enabling effective composition of task vectors and achieving strong performance across vision and language benchmarks without incurring any inference-time overhead.

Classifying protein topology is essential for deciphering biological function, but progress is held back by the lack of large-scale benchmarks that avoid duplicates and by models that do not scale well. We introduce TEDBench, a large-scale, non-redundant benchmark for protein fold classification constructed from the Encyclopedia of Domains (TED) and Foldseek-clustered AlphaFold structures. We show that on TEDBench, current protein representation learning methods either require very large models or fail to deliver strong performance. To address this challenge, we propose Masked Invariant Autoencoders (MiAE), a self-supervised framework for protein structure representation learning. MiAE uses an extremely high masking ratio of up to 90% with an $\mathrm{SE(3)}$-invariant encoder and a lightweight decoder that reconstructs backbone coordinates from the latent representation and mask tokens. MiAE scales well and outperforms supervised counterparts and state-of-the-art baselines on TEDBench, establishing a strong recipe for protein fold classification. To test transfer beyond AlphaFold structures, we further benchmark on a curated dataset from experimental structures of CATH v4.4. TEDBench is available at https://github.com/BorgwardtLab/TEDBench.

We present TabH2O, a foundation model for tabular data that performs classification and regression in a single forward pass via in-context learning. TabH2O builds on the TabICL architecture with several key modifications: (1) unified training, a single model handles both classification and regression via a dual-head architecture, eliminating the need for separate models and reducing total pretraining cost; (2) single-stage pretraining, training stability improvements (bounded scalable softmax, inter-stage normalization, learnable residual scaling, logit soft-capping) eliminate the need for multi-stage curriculum learning, enabling training with full-length sequences from the start; and (3) noise-aware pretraining, synthetic datasets include explicit noise dimensions to teach the model robustness to irrelevant features. We evaluate TabH2O v1 (29.2M parameters) on the TALENT benchmark (300 datasets), where it achieves an average rank of 2.55 out of 6 evaluated methods, outperforming tuned CatBoost (4.07), H2O AutoML (4.18), and LightGBM (5.08), competitive with TabPFN v2.6 (2.74), and behind TabICL v2 (2.12), while placing in the top-3 on 81% of the testing datasets across classification and regression tasks.

LLM pretraining is shifting from a compute-bound to a data-bound regime, where available human (organic) text falls far short of scaling demands. However, reaching the data-bound regime does not mean the model has fully utilized its organic corpus. In this paper, we introduce SynPro, a synthetic data generation framework that helps LLMs more thoroughly learn from limited organic data. SynPro applies two operations, rephrasing and reformat, that present the same organic source in diverse forms to facilitate deeper learning without introducing external information. Both generators are optimized via reinforcement learning with quality, faithfulness, and data influence rewards, and are continuously updated as pretraining plateaus to target content the model has yet to absorb. We pretrain 400M and 1.1B models with 10% of their Chinchilla-optimal tokens (0.8B and 2.2B) from DCLM-Baseline, reflecting a realistic data-bound regime in frontier pretraining. Our results reveal that organic data is significantly underutilized by standard repetition: SynPro unlocks 3.7-5.2x the effective tokens of repetition, even surpassing the non-data-bound oracle that trains on equivalent unique data at the 1.1B scale. Analyses confirm that faithful, model-aware synthesis sustains data-bound scaling without causing distribution collapse. We open-source our code at https://github.com/cxcscmu/SynPro.

Muon has recently emerged as a competitive alternative to AdamW for large-scale pre-training, with orthogonalization via Newton-Schulz (NS) iterations as its core operation. Existing Muon variants apply a uniform NS schedule to all parameter matrices, overlooking possible differences in orthogonalization difficulty and its impact on performance. Through a systematic empirical study, we show that this per-matrix heterogeneity is pervasive and largely determined by matrix geometry, which evolves dynamically across operator types, training stages, and network depths. As a result, uniform NS schedules can lead to uneven orthogonalization quality across the model. Motivated by these findings, we propose Adaptive Muon Orthogonalization (AMO), an observe-then-commit method that measures weight geometry by operator type early in training and then uses these signals to allocate the NS budget for the remainder of training. AMO delivers consistent improvements over uniform-schedule Muon across standard, prolonged, and continual pre-training, surpassing the strongest baseline by +0.76 on Llama3.1-1.4B and +0.51 on Qwen3-1.7B in average downstream performance of 12 evaluation tasks.

It is widely believed that stochastic gradient descent (SGD) performs significantly worse than adaptive optimizers such as Adam in pre-training Large Language Models (LLMs). Yet the underlying reason for this gap remains unclear. In this work, we attribute a large part of the discrepancy to SGD's inability to sustain learning rates comparable to Adam's much larger effective learning rates. Through empirical and theoretical analysis of LLM pre-training dynamics, we identify that training is characterized by small gradient norms and large weight-to-gradient ratios, an effect that becomes more pronounced with larger batch sizes typical in pre-training, necessitating such large effective learning rates. However, we find that output-layer gradient magnitudes become highly uneven across token classes, and that large gradient spikes frequently occur during training. Together, these effects severely restrict the admissible learning rate of SGD. Guided by this understanding, we show that simple clipping mechanisms that stabilize SGD at large learning rates enable it to recover most of Adam's performance. In our large-scale experiments, the validation loss gap between large-learning-rate SGD and Adam shrinks from more than 50% to only about 3.5% when pre-training a 1B-parameter LLaMA model with a 1M-token batch size.

Music search at the scale of Amazon Music presents a unique challenge: queries frequently deviate from indexed metadata due to misspellings, transpositions, and phonetic variations, yet the retrieval system must operate under strict millisecond-level latency constraints. Our existing learning-to-retrieve system, the High Confidence Index (HCI), learns query-entity associations from customer behavior, relying on continual ``exploration'' to choose candidates. Traditional n-gram matching enables this exploration but suffers from poor semantic robustness and high noise, limiting the system's ability to learn from long-tail queries. In this work, we present a \textbf{robust neural sparse retrieval system} designed to maximize exploration efficiency. We adapt a state-of-the-art \textbf{inference-free} sparse retrieval architecture to the music domain, combining it with an effective \textbf{domain-specific granular subword tokenization strategy}. Our approach utilizes short-length token constraints (max 3 chars) to enforce the learning of surface-form robustness over lexical memorization. By pre-computing the neural embeddings and term expansions during the offline indexing phase, online processing is reduced to minimal tokenization and IDF weighting, achieving effectively zero latency overhead for query encoding. Evaluations on a 6M-document production corpus show an aggregate \textbf{91.4%} recall@10 (vs. \textbf{57.7%} for trigrams) at comparable throughput. Simulation of the HCI feedback loop demonstrates improved exploration efficiency, with \textbf{+0.8%} higher stabilized recall than production trigrams. Ablation studies indicate that our sparse training methodology drives the performance gains, while domain-specific pretraining provides a cost-effective alternative to large-scale general-purpose pretraining.

Continual test-time adaptation adapts a source-pretrained model to non-stationary, unlabeled target streams while retaining past competence, yet texture-biased backbones risk error accumulation and catastrophic forgetting. Drawing inspiration from the process of decoupling shape and texture in the human visual system, we introduce MoASE, a plug-in mixture-of-experts that disentangles domain-agnostic structure from domain-specific texture using Activation Sparsity Experts with Spatial Differentiable Dropout, forming complementary high- and low-activation pathways, while high- and low-rank bottlenecks diversify representations. The Activation Sparsity Gate produces input-adaptive SDD thresholds for precise token selection, and the Domain-Aware Router assigns per-sample expert weights using texture-sensitive cues. To curb confirmation bias on unlabeled streams and stabilize supervision, we then introduce Domain-Adaptive On-Policy Distillation to constitute MoASE++, with an EMA-anchored on-policy reverse KL distillation and an augmentation policy conditioned on entropy and confidence that aligns predictions across the same views and improves the robustness-plasticity balance. Extensive experiments on classification (CIFAR-10/100-C, ImageNet-C) and semantic segmentation (Cityscapes->ACDC) demonstrate consistent state-of-the-art performance, offering a principled, controllable approach to continual adaptation in dynamic visual environments.

Large language models pretrained on general-domain corpora often exhibit tokenization inefficiencies when applied to specialized domains. Although continual pretraining for domain adaptation partially alleviate performance degradation, it does not resolve the fundamental vocabulary mismatch. To address this gap, we introduce a targeted parameter-efficient domain adaptation approach that combines vocabulary adaptation with pretraining for LLM-based text summarization. Our unified framework augments pretrained tokenizers with domain-specific tokens while selectively replacing under-trained and unreachable tokens to limit parameter growth. We evaluate our approach on Llama-3.1-8B and Qwen2.5-7B across legal and medical summarization tasks on a challenge-oriented evaluation protocol focused on expert-driven text and summaries which typically has higher concentration of over-fragmented Out-of-Vocabulary (OOV) words. The vocabulary adaptation algorithm enhances the overall quality of the summarization model by improving semantic similarity between the generated summaries and their references. In addition, the adapted model produces summaries that incorporate more appropriate novel and domain-specific words, leading to improved coherence, relevance, and faithfulness. We further observe that our proposed approach significantly reduce training time by $35-55\%$ over continual pretraining and reduce parameter counts up to $37\%$ w.r.t expansion-only methods. We make the codebase publicly available at https://github.com/gb-kgp/VocabReplace-Then-Expand.

Time series foundation models rely on large-scale pretraining over diverse datasets across domains, yet their heterogeneity in temporal patterns could hinder the effectiveness of training and learning transferable time series representations. Inspired a fundamental concept, normalized power spectral density (PSD) in signal processing, we assume harmonizing datasets via PSDs in the spectral domain could reduce mismatches and enhance pretraining. We then go beyond the direct intractable minimization optimization and innovatively reformulate it as a principled harmonization approach. Specifically, we propose Harmonizer, a module that reshapes spectral structures and implicitly harmonizing PSDs across datasets, which theoretically corresponds to a shared reparameterization of second-order temporal correlations. Our theoretical analysis further reveals token interactions with Harmonizer can be efficiently mediated by a compact set of resonators, motivating a HarmonicAttention design that performs self-attention in a low-dimensional interaction space. Then, we propose Olivia, a novel time series foundation model built upon these harmonization mechanisms. Extensive experiments on two large-scale benchmarks (TSLib and GIFT-Eval) and extra 6 datasets from GluonTS, demonstrate Olivia consistently achieves state-of-the-art performance under zero-shot, few-shot, and full-shot forecasting scenarios. Our code is available at https://github.com/TSTS13/Olivia.

While scaling laws have established a fundamental framework for foundation models in natural language processing, their applicability to electrocardiogram (ECG) models remains poorly characterized. Indeed, recent studies do not always yield consistent downstream gains as one increases the model size or pre-training dataset size of ECG models, leaving the exact roles of architectural inductive biases, pre-training paradigms, and expected improvements with size largely unanswered. In this work, we systematically investigate neural and loss-to-loss scaling laws within the ECG domain. By pre-training over $120$ models (ranging from $20$K to $200$M parameters) on the large-scale CODE dataset ($2.3$M records), we decouple the effects of model architecture (ResNet vs. Transformer) and pre-training paradigm, namely supervised learning (SL) versus self-supervised learning (SSL). We found that (i) SL models are data-bottlenecked in-distribution, whereas SSL models scale robustly across both model and data sizes; (ii) for out-of-distribution (OOD) generalization, ResNets are $1.3$ to $2.5$ times more parameter-efficient than Transformers, while SSL is up to $16$ times more data-efficient and achieves up to $7.6$ times higher transfer efficiency than SL on unseen clinical tasks; (iii) across the observed scales, ResNet-based models generally achieve the lowest OOD loss, with SSL dominating on unseen clinical tasks and self-supervised Transformers overtaking at very large model sizes. Our results suggest that the path to effective ECG foundation models lies in the strategic alignment of architecture and paradigm rather than brute-force scaling.

Electrocardiogram (ECG) foundation models pretrained on typical diagnostic 10-second ECG segments, have demonstrated strong transferability across a range of clinical applications. However, many real-world applications produce recordings that are typically longer, and are varied in duration during inference time. These 10-second models have no built-in way to combine information across time. Extending them to longer horizons introduces two challenges: structural incompatibilities arising from input-length disparities, and semantic challenges that limit meaningful temporal aggregation. We propose a parameter-efficient framework that extends pretrained ECG foundation models to longer and variable-length ECGs without retraining the backbone. Guided by a frozen pretrained 10-second model, we introduce a lightweight plug-in module that extends the model in two complementary ways: (i) structurally compatible long-sequence processing and (ii) semantically informed temporal modeling. Experiments on multiple long-horizon ECG tasks, datasets, and foundation model backbones demonstrate that our method enables robust long-horizon extension from pretrained snapshot models, consistently outperforming sliding-window and pooling-based baselines with strong parameter efficiency.

Cross-entropy pretraining and preference alignment update the same transformer weights, but leave geometrically distinct traces. We characterise this asymmetry with a relative-subspace-fraction probe that tracks how weight deltas align with residual-stream activation subspaces and with the prediction subspace defined by the unembedding. Alignment deltas concentrate in the read pathway ($W_Q$, $W_K$), along principal directions of attention-input activations, while remaining near-isotropic in the write pathway ($W_O$, $W_2$) relative to the prediction subspace. We explain this pattern through anisotropic gradient accumulation: updates to a matrix $W$ are sums of outer products $δ_t a_t^\top$, and inherit directional structure from whichever side has concentrated covariance. For read-pathway matrices, this side is the input activation $a_t$, whose covariance is spiked in trained transformers and therefore produces objective-agnostic concentration. For write-pathway matrices, the relevant side is the upstream gradient $δ_t$, whose anisotropy depends on the loss. Cross-entropy supplies the canonical sharp per-sample signal, inducing write-pathway prediction geometry during pretraining; alignment objectives typically add little further write-side concentration. We support this explanation with a within-checkpoint trajectory, a graded contrastive-objective control, and a closed-form rank-1 intervention with matched direction controls, providing causal evidence for the proposed weight-space geometry.

Timeseries regression models often struggle to leverage large volumes of labeled multimodal data, particularly when the data are irregularly sampled or contain missing values. This is common in domains like healthcare and predictive maintenance, where data are collected from unreliable sources, and labeling requires expert knowledge or costly equipments. Transformer-based large language models have proven effective on structured data such as text through self-supervised learning (SSL) and generative pretraining (GPT) frameworks. However, such models lack the flexibility to efficiently process irregularly sampled multimodal timeseries data. In this paper, we introduce ITGPT, an attention-based architecture designed for handling multimodal, irregularly sampled timeseries by allowing training with both SSL losses and GPT-like objectives. We evaluate its performance on a healthcare task with the TIHM dataset, and a predictive maintenance task with the CompX dataset. Our results demonstrate that ITGPT achieves state-of-the-art performance without requiring resampling, feature fusion or explicit data imputation. Furthermore, when labels are scarce, ITGPT effectively leverages unlabeled data through SSL and GPT training, outperforming the purely supervised approach. This represents an important step towards efficiently using large and unstructured timeseries datasets for practical inference tasks.

Efficient transfer learning of vision-language models (VLMs) commonly suffers from a Base-New Trade-off (BNT): improving performance on unseen (new) classes often degrades accuracy on known (base) classes. Addressing how to boost recognition of unseen classes without sacrificing known-class performance remains a central challenge. Existing work often simplistically attributes the BNT to overfitting on known classes. We observe an interesting phenomenon: VLMs frequently exhibit asymmetric confusion on certain downstream data, i.e., samples of class A are systematically mispredicted as class B, while the reverse confusion (B to A) rarely occurs. For known classes, this kind of bias can be mitigated by tuning using a cross-entropy loss, but for unseen classes, such pretraining-induced bias persists and harms generalization. Motivated by this, we propose NeRP, a plug-and-play prompting correction strategy that improves discrimination on unseen classes without modifying model parameters. NeRP leverages neutral text prompts and reference images to measure class-wise prior preferences along the pre-trained inter-class geometry, and combines them with the sample likelihood to obtain the model's surrogate score. If, for a given sample, the prior strongly favors the current prediction while the observed evidence is clearly insufficient, we perform a local flip between easily confusable class pairs, thereby correcting prior-dominated mispredictions. Extensive experiments across multiple backbones and 15 few-shot and cross-domain benchmarks show that NeRP substantially improves accuracy on unseen classes while preserving known-class prediction performance.

Through an analysis of sequences generated by open-weight large language models (LLMs), we demonstrate that LLMs are heavily biased toward English. While continual pre-training is commonly used to adapt LLMs to a target language, we show that it does not offer a cost advantage over training from scratch, even for improving cultural understanding in the target language. These findings suggest that dedicated per-language investment may become increasingly important for future LLM development, rather than relying primarily on the expansion of English-centric resources.

Extreme low-data fine-grained classification is common in expert domains where labeling is expensive, yet practitioners still need principled guidance for selecting pretrained encoders. We study emerald inclusion grading with a custom dataset of labeled images across three classes and ask: under matched backbone capacity, how does pretraining objective affect downstream representation quality? We compare four frozen ViT-B/16 encoders trained with supervised classification, contrastive learning (SigLIP2), masked reconstruction (MAE), and self-distillation (DINOv3), and evaluate them with leave-one-out cross-validation using linear and nonlinear probes. To control statistical noise in the low-N regime, we use permutation testing (N=1000) on macro one-vs-rest AUC. Supervised and contrastive encoders provide the strongest linear separability (logistic AUC: 0.768 and 0.735; SVM AUC: 0.739 and 0.697), while MAE improves under nonlinear probes (XGBoost AUC: 0.713). We find that DINOv3 underperforms across probe families in this domain. These results support a practical recommendation for extreme low-data FGVC: prioritize margin-enforcing pretraining objectives when data scarcity restricts probing to linear decision rules, and consider reconstruction-style encoders when nonlinear classifiers are feasible given dataset constraints.

Continually pre-training a large language model on heterogeneous text domains, without replay or task labels, has remained an unsolved architectural problem at LLM scale. Existing methods rely on replay buffers, task identifiers, regularization penalties that scale poorly, or sentence-classification-scale evaluation. We introduce TFGN, an architectural overlay for transformer language models that produces input-conditioned, parameter-efficient updates while leaving the rest of the transformer unchanged. On six heterogeneous text domains (Prose, Python, Math, Biomedical, Chinese, JavaScript) at 1B tokens per phase across three model scales (~398M, ~739M, ~9B) and two regimes (From-Scratch and Retrofit), TFGN achieves backward transfer of -0.007 at LLaMA 3.1 8B Retrofit, HellaSwag retention 0.506/0.504/0.510, and >=99.59% L2-orthogonal gradient separation between domain pairs - with no replay, no task IDs, no Fisher penalty. The same matrices show positive cross-domain forward transfer: held-out JavaScript PPL drops 26.8% at LLaMA-8B Retrofit and 62.0% at GPT-2 Medium From-Scratch purely from Python training. Two extensions on the same substrate close further open problems. A closed-loop meta-control layer (Extension A) reduces forgetting by an additional 81% at ~398M, mapping onto the System A and System M roles of Dupoux et al. (arXiv:2603.15381). An operator-level plan vector (Extension B) reshapes forward-pass behavior at 99.96% cosine fidelity over 30 source->target pairs. The architectural insight is a Read/Write decomposition: the forward pass is fully dense, while cross-domain parameter updates are structured so prior-domain subspaces are not written to. To our knowledge, TFGN is the first architecture that simultaneously closes catastrophic forgetting at LLM scale, realizes a closed-loop autonomous-learning meta-controller, and carries an operator-level latent planner.

Medical image segmentation remains challenging in low-data regimes, where scarce annotations often yield poor generalization and ambiguous boundaries with missing fine structures. Recent self-supervised pretraining has improved transferability, but it often exhibits a texture bias. In contrast, accurate segmentation is inherently geometry-aware and depends on both topological consistency and precise boundary preservation. To address this problem, we propose a two-stage framework that couples structure-aware encoder pretraining with boundary-oriented decoding. In Stage-1, we aim to learn structure-aware representations for downstream segmentation in low-data regimes. To this end, we propose Mixed-Domain MeanFlow Pretraining, which aligns images and binary masks in a shared latent space through latent transport regression, where masks act as conditional structural guidance rather than prediction targets, making the pretraining task-agnostic. To further improve training stability under scarce supervision, we incorporate a lightweight Dispersive Loss to prevent representation collapse. In Stage-2, we fine-tune the pretrained encoder with a lightweight decoder that combines Direct Attentional Fusion for adaptive cross-scale gating and Frequency-Directional Dynamic Convolution for high-frequency boundary refinement under appearance variation. Experiments on ISIC-2016, Kvasir-SEG, and GlaS demonstrate consistent gains over state-of-the-art methods, with improved robustness in low-data settings and sharper boundary delineation.

Inferring spatially resolved gene expression from histology images offers a cost-effective complement to spatial transcriptomics (ST). However, existing methods reduce this task to a simple morphology-to-expression mapping, where visual similarity does not guarantee molecular consistency. Meanwhile, single-cell data has amassed rich resources far surpassing the scale of ST data, yet it remains underexplored in vision-omics modeling. Furthermore, current approaches commit to a monolithic paradigm with bottlenecks, unable to balance expressive flexibility with biological fidelity. To bridge these gaps, we propose DUET, a novel dual-paradigm framework that synergizes parametric prediction and memory-based retrieval under cellular inductive priors. DUET implements a parallel regression-retrieval paradigm, adaptively reconciling the outputs of its complementary pathways. To mitigate aleatoric vision ambiguity, we incorporate large-scale single-cell references to impose molecular states as biological constraints for faithful learning. Building upon structural refinement, we further design a lightweight adapter to dynamically assign branch preference across spatial contexts to achieve optimal performance. Extensive experiments on three public datasets across varied gene scales demonstrate that DUET achieves SOTA performance, with consistent gains contributed by each proposed component. Code is available at https://github.com/Junchao-Zhu/DUET

In current Large Language Models we can trust the production of smoothly flowing prose on the basis of the principles of machine learning. However, there is no comparably principled basis to justify trust in the content of the text produced. It appears to be conventional wisdom that addressing this issue by adding more principled reasoning is not computationally affordable. Here we propose a principled method of reasoning that is efficient enough to be practical for large language models. Further, the method allows the retention of much of the currently used software and hardware base. Our method for improving the functioning of large language models consists of a first stage of preprocessing that recodes the data to a Unary Relational Integracode that is more explicit about the relationships among the objects described in the text, followed as a second stage by a standard but possibly streamlined machine learning process that then also learns to predict these relationships. The method may be viewed as realizing a world model and applying beyond natural language, to vision and actions, for example, where the multiple properties of an object referred to in an input are brought together explicitly, rather than remaining distributed in the various references to it in the input. We articulate its advantages in terms of Robust Logic, a system for performing principled chaining on learned, and hence uncertain, information. We show that this recoding has the surprising and fortuitous property that, while succinct, it makes the task of learning a core subset of relational rules that hold in the world described in the training data polynomial time learnable in a defined sense, the polynomial depending on the complexity of the rule. This gives support for sound reasoning within each single call of the learned classifier as well as between multiple calls.

Bioacoustic recognition requires fine-grained acoustic understanding to distinguish similar-sounding species. However, many large-scale data repositories such as iNaturalist are weakly annotated, often with only a single positive species label per recording, making supervised learning particularly challenging. Inspired by advances in computer vision, recent approaches have shifted toward self-supervised learning to capture the underlying structure of audio without relying on exhaustive annotations. In particular, masked autoencoders (MAE) have shown strong transferability on massive audio corpora, yet their effectiveness in more modest bioacoustic settings remains underexplored. In this work, we conduct a systematic study of MAE pretraining for species classification on iNatSounds, analyzing the impacts of pretraining data scale, domain specificity, data curation, and transfer strategies. Consistent with prior work, we find that models pretrained on diverse general audio data achieve the best transfer performance on iNatSounds. Contrary to observations from large-scale audio benchmarks, we find that (1) additional masked reconstruction pretraining on domain-specific data provides limited benefits and may even degrade performance relative to off-the-shelf models, and (2) selective data filtering offers a negligible advantage when the overall data scale is limited. Our results indicate that, in moderate-sized fine-grained bioacoustic settings, pretraining scale dominates objective design. These findings further clarify when MAE-based pretraining is effective and provide practical guidance for model selection under limited supervision.

Subword regularization methods such as BPE dropout are typically applied only during fine-tuning, while pretraining is usually done with deterministic tokenization. This creates a potential segmentation mismatch between pretraining and fine-tuning. We investigate whether applying BPE dropout during pretraining improves downstream performance in low-resource NLP. We train monolingual and bilingual BERT models on downsampled subsets of English, German, French, Spanish, Kiswahili, and isiXhosa, and evaluate them on XNLI, PAWS-X, PAN-X, and MasakhaNER 2.0. Across tasks, the best results are typically obtained when stochastic tokenization is applied during both pretraining and fine-tuning, whereas applying BPE dropout only during fine-tuning can underperform deterministic tokenization in smaller-data settings. This disadvantage diminishes as fine-tuning data increases, while the benefits of pretraining-time BPE dropout are largest when either pretraining or fine-tuning data is scarce. The benefits of BPE dropout are often attributed to better compositional representations, especially for rare words. To examine this, we measure morphological boundary alignment under BPE dropout and find only modest improvements in expected alignment, while better-aligned segmentations remain rare. This suggests that fine-tuning alone may provide limited exposure to such segmentations, whereas stochastic tokenization during pretraining exposes the model to them more consistently. We further show that selectively introducing morphologically aligned segmentations during fine-tuning improves performance mainly for models pretrained without BPE dropout. Overall, these findings suggest that exposure to better-aligned segmentations may contribute to the downstream benefits of applying BPE dropout during pretraining.

Density estimation in high-dimensional settings is an important and challenging statistical problem.Traditional methods based on kernel smoothing are inefficient in high dimensions due to the difficulties in specifying appropriate location-adaptive kernels. In this work, we introduce pre-training, a key idea behind many cutting-edge AI technologies, to the context of non-parametric density estimation. By establishing a pre-trained neural network that can recommend an appropriate location-adaptive kernel for each sample point, efficient density estimation with adaptive kernels is achieved in high dimensions. A wide range of numerical experiments show that this strategy is highly effective for improving density-estimation accuracy, when the target distribution is close to the distribution family for pre-training. When the target distribution is substantially different from the pre-training distribution family, the benefit from the proposed pre-training strategy may be diluted, but can be reactivated by an additional fine-tuning procedure.

As language models scale, the amount of data they require grows -- yet many target data sources, such as low-resource languages or specialized domains, are inherently limited in size. A common strategy is to mix this scarce but valuable target data with abundant generic data, which presents a fundamental trade-off: too little target data in the mixture underexposes the model to the target domain, while too much target data repeats the same examples excessively, yielding diminishing returns and eventual overfitting. We study this trade-off across more than 2,000 language-model training runs spanning multiple model and target dataset sizes, as well as several data types, including multilingual, domain-specific, and quality-filtered mixtures. Across all settings, we find that repetition is a central driver of target-domain performance, and that mixture training tolerates much higher repetition than single-source training: scarce target corpora can be reused 15-20 times, with the optimal number of repetitions depending on the target data size, compute budget, and model scale. Next, we introduce a repetition-aware mixture scaling law that accounts for the decreasing value of repeated target tokens and the regularizing role of generic data. Optimizing the scaling law provides a principled way to compute effective mixture configurations, yielding practical mixture recommendations for pretraining under data constraints.

The continued improvements in language model capability have unlocked their widespread use as drivers of autonomous agents, for example in coding or computer use applications. However, the core of these systems has not changed much since early instruction-tuned models like ChatGPT. Even advanced AI agents function on message exchange formats, successively exchanging messages with users, systems, with itself (i.e. chain-of-thought) and tools in a single stream of computation. This bottleneck to a single stream in chat models leads to a number of limitations: the agent cannot act (generate output) while reading, and in reverse, cannot react to new information while writing. Similarly, the agent cannot act while thinking and cannot think while reading or acting on information. In this work, we show that models can be unblocked by switching from instruction-tuning for sequential message formats to instruction-tuning for multiple, parallel streams of computation, splitting each role into a separate stream. Every forward pass of the language model then simultaneously reads from multiple input streams and generates tokens in multiple output streams, all of which causally depend on earlier timesteps. We argue that this data-driven change remedies a number of usability limitations as outlined above, improves model efficiency through parallelization, improves model security through better separation of concerns and can further improve model monitorability.

When adapting an encoder to a new domain, the standard approach is to continue training with Masked Language Modeling (MLM). We show that temporarily switching to Causal Language Modeling (CLM) followed by a short MLM decay improves downstream performance. On biomedical texts with ModernBERT, this CLM detour outperforms MLM baselines trained on identical data and compute across 8 French and 11 English biomedical tasks, by +1.2-2.8pp and +0.3-0.8pp respectively, depending on model size. We investigate the reasons for these gains. We find that CLM's dense supervision impacts low transformer layers (0-7) far more than MLM does. Freezing low layers during CLM eliminates the downstream benefit; freezing mid layers preserves it. The representational changes persist through the MLM decay phase, even when it matches the CLM phase in length, and they scale with model capacity. We release ModernCamemBERT-bio and ModernBERT-bio as state-of-the-art biomedical encoders in Base and Large sizes.

Zero-shot anomaly detection aims to identify defects in unseen categories without target-specific training. Existing methods usually apply the same feature transformation to all samples, treating normal and anomalous data uniformly despite their fundamentally asymmetric distributions, compact normals versus diverse anomalies. We instead exploit this natural asymmetry by proposing AVA-DINO, an anomaly-aware vision-language adaptation framework with dual specialized branches for normal and anomalous patterns that adapt frozen DINOv3 visual features. During training on auxiliary data, the two branches are learned jointly with a text-guided routing mechanism and explicit routing regularization that encourages branch specialization. At test time, only the input image and fixed, predefined language descriptions are used to dynamically combine the two branches, enabling an asymmetric activation. This design prevents degenerate uniform routing and allows context-specific feature transformations. Experiments across nine industrial and medical benchmarks demonstrate state-of-the-art performance, achieving 93.5% image-AUROC on MVTec-AD and strong cross-domain generalization to medical imaging without domain-specific fine-tuning. https://github.com/aqeeelmirza/AVA-DINO

Transfer learning is a machine learning technique that uses previously acquired knowledge from a source domain to enhance learning in a target domain by reusing learned weights. This technique is ubiquitous because of its great advantages in achieving high performance while saving training time, memory, and effort in network design. In this paper, we investigate how to select the best pre-trained model that meets the target domain requirements for image classification tasks. In our study, we refined the output layers and general network parameters to apply the knowledge of eleven image processing models, pre-trained on ImageNet, to five different target domain datasets. We measured the accuracy, accuracy density, training time, and model size to evaluate the pre-trained models both in training sessions in one episode and with ten episodes.

Selective layer-wise updates are essential for low-cost continued pre-training of Large Language Models (LLMs), yet determining which layers to freeze or train remains an empirical black-box problem due to the lack of interpretable guidance. To address this issue, we propose LayerTracer, an architecture-agnostic diagnostic framework that reveals the evolution patterns of layer-wise representations and stability by locating task execution positions and quantifying layer sensitivity. Analysis results reveal that deep layers act as critical regions for task execution and maintain high stability against disruptive updates. Guided by this finding, we conduct three controlled continued pre-training trials to compare diverse freeze-train strategies, demonstrating that training shallow layers while freezing deep layers consistently outperforms full-parameter fine-tuning and the opposite allocation on both C-Eval and CMMLU benchmarks. We further present a hybrid model case study, which validates that placing high-quality pre-trained modules in deep layers effectively preserves inherent knowledge of the model. This work delivers a low-cost and interpretable solution for resource-constrained teams, offering actionable guidance for layer-wise parameter allocation in continued pre-training and hybrid model construction.

Tabular data forms the backbone of high-stakes decision systems in finance, healthcare, and beyond. Yet industrial tabular datasets are inherently difficult: high-dimensional, riddled with missing entries, and rarely labeled at scale. While foundation models have revolutionized vision and language, tabular learning still leans on handcrafted features and lacks a general self-supervised framework. We present MaskTab, a unified pre-training framework designed specifically for industrial-scale tabular data. MaskTab encodes missing values via dedicated learnable tokens, enabling the model to distinguish structural absence from random dropout. It jointly optimizes a hybrid supervised pre-training scheme--utilizing a twin-path architecture to reconcile masked reconstruction with task-specific supervision--and an MoE-augmented loss that adaptively routes features through specialized subnetworks. On industrial-scale benchmarks, it achieves +5.04% AUC and +8.28% KS over prior art under rigorous scaling. Moreover, its representations distill effectively into lightweight models, yielding +2.55% AUC and +4.85% KS under strict latency and interpretability constraints, while improving robustness to distribution shifts. Our work demonstrates that tabular data admits a foundation-model treatment--when its structural idiosyncrasies are respected.

We investigate whether the training protocol can induce spatial locality in the early layers of a Vision Transformer (ViT) trained from scratch, without large-scale pretraining. Keeping the architecture and optimization procedure fixed, we compare a Baseline protocol with a Modern protocol (AutoAugment/ColorJitter, CutMix, and Label Smoothing) on CIFAR-10, CIFAR-100, and Tiny-ImageNet, characterizing each attention head via Mean Attention Distance (MAD) and normalized entropy. Across all three datasets, the Modern protocol produces more local and more concentrated attention in early layers; on CIFAR-100, the minimum MAD drops from 0.316 (Baseline) to 0.008 (Modern). To identify the source of this effect, we conduct an ablation study on CIFAR-100 by adding or removing each component individually. The results identify CutMix as the determining component within our experiments: all conditions with CutMix exhibit MAD 0.024, while all conditions without CutMix remain at MAD 0.210. AutoAugment and Label Smoothing show no independent effect on locality. Taken together, these findings suggest that the pressure to classify from partial image regions, induced by CutMix, can promote the emergence of local attention in Vision Transformers.

Large-scale pretraining has made Vision-Language-Action (VLA) models promising foundations for generalist robot manipulation, yet adapting them to downstream tasks remains necessary. However, the common practice of full fine-tuning treats pretraining as initialization and can shift broad priors toward narrow training-distribution patterns. We propose PriorVLA, a novel framework that preserves pretrained priors and learns to leverage them for effective adaptation. PriorVLA keeps a frozen Prior Expert as a read-only prior source and trains an Adaptation Expert for downstream specialization. Expert Queries capture scene priors from the pretrained VLM and motor priors from the Prior Expert, integrating both into the Adaptation Expert to guide adaptation. Together, PriorVLA updates only 25% of the parameters updated by full fine-tuning. Across RoboTwin 2.0, LIBERO, and real-world tasks, PriorVLA achieves stronger overall performance than full fine-tuning and state-of-the-art VLA baselines, with the largest gains under out-of-distribution (OOD) and few-shot settings. PriorVLA improves over pi0.5 by 11 points on RoboTwin 2.0-Hard and achieves 99.1% average success on LIBERO. Across eight real-world tasks and two embodiments, PriorVLA reaches 81% in-distribution (ID) and 57% OOD success with standard data. With only 10 demonstrations per task, PriorVLA reaches 48% ID and 32% OOD success, surpassing pi0.5 by 24 and 22 points, respectively.

The quality of open-weight language models has dramatically improved in recent years. Sharing weights greatly facilitates model adoption by enabling their use across diverse hardware and software platforms. They also allow for more open research and testing, to the extent that users can use them as checkpoints, fine-tune them according to their needs, and potentially redistribute them. In some cases, however, concerns on modifying these weights towards unauthorized uses may outweigh the pros of giving users such a freedom. Defending against such adaptation is non-trivial: since an adaptive attacker can observe all weights and architectures by definition, they can reverse simple structural defenses, and use optimization to defeat the simplest locking mechanisms. In this work, we exploit the inference-training asymmetry of automatic differentiation as a novel defense axis. We propose DLR-Lock, a method where the purveyor of the model purposely replaces each pretrained MLP in their model with a deep low-rank residual network (DLR-Net) of comparable parameter count, forcing activation memory that grows linearly with depth during backpropagation. DLR-Nets are efficiently trained via module-wise distillation. We show that, beyond this memory overhead, DLR-Lock results in architectural mismatches that complicate the optimization landscape of standard fine-tuning, and a backward pass that incurs disproportionately more overhead than the forward pass. Our defense succeeds in withstanding adaptive attackers with full knowledge of the defense strategy while preserving the original model's capabilities. Experiments on LLM validate these claims.

Molecular Graph Neural Networks (GNNs) are increasingly common in drug discovery, particularly for Quantitative Structure-Activity Relationship (QSAR) studies; yet, their superiority compared to classical molecular featurisation approaches is disputed. We report a general strategy for improving GNNs for QSAR by pre-training to predict Extended-Connectivity Fingerprints (ECFP). We validate our approach with statistical tests and challenging out-of-distribution (OOD) splits. Across five out of six Biogen benchmarks, we observed a statistically significant improvement in standard performance metrics over all evaluated baselines when using ECFP pre-trained GNNs. However, for more heterogeneous datasets and more complex endpoints, such as binding affinity prediction, pre-trained GNNs underperformed in OOD settings. Importantly, we investigated the impact of substructure-level data leakage during pre-training on downstream performance. While we identified scenarios where pre-training on ECFPs was less effective, our findings show that ECFP-based pre-training can enhance downstream OOD performance on a diverse set of practically relevant QSAR tasks.

Continual Pre-Training (CPT) is essential for enabling Language Models (LMs) to integrate new knowledge without erasing old. While classical CPT techniques like data replay have become the standard paradigm, the mechanisms underlying how LMs acquire and retain facts over time, termed as continual Factual Knowledge Acquisition (cFKA), remain unclear. In this work, we present a theoretical framework that characterizes the training dynamics of cFKA using a single-layer Transformer, offering a unified explanation for the behavior of representative CPT methods. Our analysis reveals that regularization-based methods merely adjust the convergence rate of parameters without altering the inherent forgetting tendency, whereas data replay methods succeed in shifting convergence dynamics and stabilizing pretrained knowledge. Building on these insights, we propose a novel generative data replay approach, called \textbf{S}electing \textbf{T}okens via attenti\textbf{O}n \textbf{C}ontribution~(STOC), which identifies influential factual snippets to guide replay data generation. Extensive experiments on both synthetic and real-world datasets validate our findings and demonstrate that STOC effectively enhances cFKA by mitigating catastrophic forgetting.

A causal-decoder block is hierarchical: lower layers build the residual basis that upper layers attend over. We identify a failure mode in GPT pretraining: upper layers commit to sharp attention patterns before lower-layer features stabilize. We call this premature upper-layer attention specialization. Temporarily slowing only upper-layer Q/K projections during early training improves final perplexity and downstream accuracy without altering other parameters; it prevents upper attention from collapsing onto an immature residual basis. In LLaMA-style blocks, the same intervention is nearly unnecessary. Through ablations, we isolate multiplicative gated FFNs (not RMSNorm or bias removal) as the component that suppresses the upstream residual writes driving the failure. A pathwise analysis unifies both findings: the learning-rate intervention reduces a step-size factor, while gated FFNs reduce a residual-energy factor on the same growth pathway. Our results identify upper-layer Q/K timing as a concrete interaction point between decoder architecture and optimization.

Large language models (LLMs) rely on web-scale corpora for pre-training. The noise inherent in these datasets tends to obscure meaningful patterns and ultimately degrade model performance. Data curation mitigates but cannot eliminate such noise, so pre-training corpora remain noisy in practice. We therefore study whether a lightweight pre-pre-training (PPT) stage based on synthetic data with learnable temporal structure helps resist noisy data during the pre-training (PT) stage. Across various corruption settings, our method consistently improves robustness to noise during PT, with larger relative gains at higher noise levels. For a 1B-parameter model, a synthetic PPT stage with only 65M tokens achieves the same final loss as the baseline while using up to 49% fewer natural-text PT tokens across different noise levels. Mechanistic analyses suggest PPT does not immediately suppress attention to noisy tokens. Rather, PPT-initialized models gradually downweight attention between corrupted tokens during noisy PT. This indicates that synthetic PPT inhibits noise self-modeling and shapes the subsequent optimization trajectory. Code is available at https://github.com/guox18/formal-language-prepretraining.

Post-training (via supervised fine-tuning) improves instruction-following, but often induces semantic mode collapse by biasing models toward low-entropy fine-tuning data at the expense of the high-entropy pretraining distribution. Crucially, we find this trade-off worsens with scale. To close this semantic diversity gap, we propose annotation-anchored training, a principled method that enables models to adopt the preference-following behaviors of post-training without sacrificing the inherent diversity of pretraining. Our approach is simple: we pretrain on documents paired with semantic annotations, inducing a rich annotation distribution that reflects the full breadth of pretraining data, and we preserve this distribution during post-training. This lets us sample diverse annotations at inference time and use them as anchors to guide generation, effectively transferring pretraining's semantic richness into post-trained models. We find that models trained with annotation-anchored training can attain $6 \times$ less diversity collapse than models trained with SFT, and improve with scale.

We present MambaNetBurst, a compact tokenizer-free byte-level sequence classifier for network burst classification based on a Mamba-2 backbone. In contrast to most recent strong traffic-classification and intrusion-detection approaches, our method operates directly on raw packet bytes, avoids tokenization, patching, and heavy engineered multimodal representations, and does not require any self-supervised pre-training stage. Given a packet flow, we form a fixed-length burst from the first few packets, embed the resulting byte sequence appending a learnable CLS token, and process it with a stack of residual pre-normalized Mamba-2 blocks for end-to-end supervised classification. Across six public benchmarks spanning encrypted mobile app identification, VPN/Tor traffic classification, malware traffic classification, and IoT attack traffic, MambaNetBurst achieves consistently strong results and is competitive with, or outperforms, substantially heavier and often pre-trained baselines. Our ablation study shows that preserving byte-level temporal resolution is critical, that early downsampling through striding is consistently harmful, and that moderate state sizes are sufficient for robust generalization. We further show that Mamba-2, despite its more constrained transition structure relative to Mamba-1, remains highly effective for packet-byte modeling while providing clear efficiency advantages, particularly in training speed. Overall, our results demonstrate that direct undiluted byte-to-classification learning with compact selective state space models is a practical, effective and novel direction for efficient, deployable traffic analysis that bypasses the complexity of pre-training pipelines even over highly optimized linear attention architectures.

Robotic deployment in real-world environments depends on rich, domain-specific action data as much as on strong model architecture. General-purpose robot foundation models show modest performance in complex unseen tasks such as manipulation in a retail domain when applied out of the box. The root cause is a data gap: retail environments are structurally absent from general robot pretraining distributions, and the path to filling that gap through teleoperation is prohibitively expensive, logistically constrained, and difficult to scale. We introduce SABER, a high-fidelity retail robotics action dataset built from over 100 hours of natural in-store capture across multiple real grocery environments. Egocentric footage from head-mounted cameras records fine-grained hand activity at the point of interaction, while exocentric 360-degree scene footage from DreamVu's ALIA camera simultaneously observes all actors and activities across the entire space. This combination yields a uniquely complete picture of human retail behavior: dexterous hand activity, whole-body motion, and scene dynamics, all captured without staging, scripting, or teleoperation overhead. The SABER corpus contains 44.8K training samples across three action representation streams: 25K latent action sequences via LAPA-style encoding, 18.6K dexterous hand-pose trajectories retargeted to robot joint space, and 1.2K whole-body synchronized motion sequences retargeted to a humanoid embodiment. When applied to GR00T N1.6 via a shared-backbone multi-task post-training recipe, SABER yields a mean success rate of 29.3% across ten retail manipulation tasks -- more than 2.19x over fine-tuning baselines (13.4%). SABER demonstrates that the path to capable retail robots runs through better data, which can be collected today, at scale, without a robot in the loop. The dataset and code are available at https://dreamvu.ai/saber

Real-world model deployment across multiple domains requires multimodal models to operate under two complementary regimes: (1) multi-task pretraining, tasks are co-available at design time where related tasks could borrow representational strength from one another, (2) continual adaptation, in which new tasks emerge after deployment with previously unseen modality combinations. However, neither regime alone suffices: the pretraining task set is never exhaustive, while bypassing joint training forfeits the transfer gains and efficiency among co-trainable tasks. Sparse Mixture-of-Experts (MoE) is a natural fit for this dual requirement: sparse activation enables modular capacity expansion as new tasks arrive, while routing decouples modality-level computation from task-level composition. In this work, we propose a scalable MoE framework for multitask pretraining and continual learning across flexible modality combinations. The framework is designed to support training on multimodal tasks with diverse modality configurations by leveraging modality-specific routers that process tokens from each modality across tasks. Furthermore, it enables continual learning over sequential multimodal tasks within a fixed-capacity MoE by compressing accumulated expert knowledge into low-rank memory subspaces, while expanding only the lightweight routers. We validate the effectiveness of our method on multiple healthcare multimodal benchmarks. It demonstrates competitive multitask pretraining performance while alleviating catastrophic forgetting and improving parameter efficiency.

The scaling laws guiding modern model training were calibrated for a single regime: data-rich, single-epoch pretraining. The dominant such scaling law form, Chinchilla's $L = E + A/N^α+ B/D^β$, has three structural limitations outside that regime: it diverges as unique data shrinks instead of saturating at the uninformed baseline; it cannot represent overfitting when capacity exceeds the data; and it conflates total examples seen with unique examples available. We propose a closed-form extension, $L(N, D, T) = E + (L_0 - E)\,h/(1+h)$ with $h = a/N^α+ b/T^β+ c\,N^γ/D^δ$, that decomposes loss into undercapacity, undertraining, and overfitting terms. It saturates between the irreducible loss $E$ and an uninformed baseline $L_0$ fixed by the loss type, and reduces to Chinchilla in the data-rich, single-epoch limit. We validate it on four multi-epoch experiments spanning four architecture families (MLPs, ResNets, Fourier neural operators, and transformers) across vision, scientific ML, and language domains, and refit it to five published LLM scaling-law grids. Extrapolating to higher compute and larger unique data than seen at fit time, our form achieves state-of-the-art RMSE on every published LLM grid we evaluate and on most cells of our constructed experiments. Once calibrated, the form admits a cost-aware allocation that recovers Chinchilla's optimum when data is free and shifts toward smaller corpora and more epochs as data grows expensive.

We introduce a predictive model that estimates the pre-training loss of large models from model size (N), batch size (B) and number of weight updates (K). This is the first loss prediction model that can handle changing batch size. The model outperforms Chinchilla's loss model, a model of the test loss using the batch size and number of tokens, in terms of projecting the loss at extrapolated compute budgets (up to 1000 folds). A natural use of the model is to find optimal N, B, K configurations under explicit and compound resource constraints like time, memory and compute. In our experiments, the model-selected configurations are close to ground-truth optimal. Our work advocates for loss prediction as a better alternative to heuristic-based laws, which are growing in complexity. The implementation is available on https://github.com/chuningxdy/Noisy-Quadratic-System.

Deep learning techniques have revolutionised medical imaging, improving diagnostic accuracy and enabling both more accurate and earlier disease detection. However, the relationship between pre-training strategies and downstream performance in medical imaging models requires further exploration. Here, we systematically compare convolutional neural networks and transformers, examining various pre-training approaches, including supervised and self-supervised learning, as well as different initialisations and data modalities. Models are evaluated on natural images, chest X-rays, chest CT and retina OCT images, considering the effects of matching pre-training data with target modalities. Our findings indicate that only pre-training on data closely matching the target modality significantly improves downstream performance. While self-supervised learning can outperform supervised methods, its effectiveness varies with context. The study underscores the importance of pre-training strategies to enhance the reliability and effectiveness of deep learning models in medical imaging. By addressing these key factors, our research aims to contribute to the development of more accurate and dependable diagnostic tools, ultimately improving patient outcomes in clinical settings.

Pretraining large language models (LLMs) with next-token prediction has led to remarkable advances, yet the context-dependent nature of token embeddings in such models results in high intra-class variance and inter-class similarity, thus hindering the efficiency of representation learning. While similarity-based regularization has demonstrated benefit in supervised fine-tuning and classification tasks, its application and efficacy in large-scale LLM pretraining remains underexplored. In this work, we propose the SimReg, an embedding similarity regularization loss that explicitly encourages token representations with the same ground-truth label within each sequence to be more similar, while enforcing separation from different-label tokens via a contrastive loss. Our analysis reveals that this mechanism introduces gains by enlarging multi-classification margins, thereby enabling more efficient classification. Extensive experiments across dense and Mixture-of-Experts (MoE) architectures demonstrate that SimReg consistently accelerates training convergence by over 30% and improves average zero-shot downstream performance by over 1% across standard benchmarks. Further ablation studies and analyses offer practical insights into hyperparameter tuning and loss effectiveness.

Structured pruning and knowledge distillation (KD) are typical techniques for compressing large language models, but it remains unclear how they should be applied at pretraining scale, especially to recent mixture-of-experts (MoE) models. In this work, we systematically study MoE compression in large-scale pretraining, focusing on three key questions: whether pruning provides a better initialization than training from scratch, how expert compression choices affect the final model after continued training, and which training strategy is most effective. We have the following findings: First, across depth, width, and expert compression, pruning a pretrained MoE consistently outperforms training the target architecture from scratch under the same training budget. Second, different one-shot expert compression methods converge to similar final performance after large-scale continual pretraining. Motivated by this, we introduce a simple partial-preservation expert merging strategy that improves downstream performance across most benchmarks. Third, combining KD with the language modeling loss outperforms KD alone, particularly on knowledge-intensive tasks. We further propose multi-token prediction (MTP) distillation, which yields consistent gains. Finally, given the same training tokens, progressive pruning schedules outperform one-shot compression, suggesting that gradual architecture transitions lead to better optimization trajectories. Putting it all together, we compress Qwen3-Next-80A3B to a 23A2B model that retains competitive performance. These results offer practical guidance for efficient MoE compression at scale.

Single-cell representation learning (SCRL) from gene expression data offers a way to uncover the complex regulatory logic underlying cellular function. Inspired by large language models in natural language modeling, several single-cell pretrained models have recently been proposed that treat genes as tokens and cells as sentences. However, these models are fundamentally limited by the long-tailed nature of cell-type distributions and struggle to generalize under covariate shifts in gene expression data. While fine-tuning is often used to mitigate these issues, we observe that performance remains bounded. To address this challenge, we introduce CellRefine, a post-pretraining method that operates between the pretraining and fine-tuning stages of a single-cell foundation model. CellRefine uses a multi-faceted objective that incorporates marker-gene sets as structural priors to guide post-pretraining and refine the latent embedding manifold of cells. Across multiple computational biology tasks, empirical results show that CellRefine consistently improves downstream performance, yielding gains up to 15%.

Modern deep models are often pretrained on large-scale data with missing labels using composite objectives, where the relative weights of multiple loss terms act as hyperparameters. Tuning these weights with random search or Bayesian optimization is computationally expensive, as it requires many independent training runs. To address this, we propose a gradient-based bilevel method that learns pretraining loss weights online by aligning the composite pretraining gradient with a downstream objective. By exploiting the structure of the loss, the method avoids the multiple backward passes typically required by truncated backpropagation through the full model, reducing the overhead of hyperparameter tuning to approximately 30% above a single training run. We evaluate the approach on event-sequence modeling and self-supervised computer vision, where it matches or improves upon carefully tuned baselines while substantially reducing the cost of hyperparameter tuning compared to random or Bayesian search.

Sparse Mixture-of-Experts (MoE) models scale capacity by routing each token to a small subset of experts. However, their routers exhibit a fundamental trade-off: strong load balancing can suppress expert specialization, while aggressive diversity often causes routing collapse. We propose Hi-MoE, a grouped MoE framework that decomposes routing control into two coupled levels: (i) inter-group balancing that enforces fair traffic across expert groups, and (ii) intra-group specialization that promotes complementary expert behaviors while preventing within-group collapse. Our analysis provides a principled explanation of how our hierarchical objectives reshape the router, thereby promoting stable specialization and mitigating collapse. We observe consistent improvements over recent sparse-routing and grouped-MoE baselines across NLP and vision benchmarks, and confirm robustness via scaling studies (model size, expert count) and targeted ablations. In large-scale pre-training on 58B tokens, Hi-MoE-7B achieves a 5.6% perplexity reduction and a 40% improvement in expert balance over OLMoE-7B across diverse evaluation domains.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a pivotal technique for enhancing the reasoning capabilities of Large Language Models (LLMs). However, the de facto practice of mainstream RL algorithms is to treat all tokens of one response equally and assign the same optimization objective to each token, failing to provide granular guidance for the reasoning process. While in Chain-of-Thought (CoT) reasoning, different tokens usually play distinct roles. Therefore, the current RL algorithms lack an effective mechanism to dynamically balance the exploration-exploitation trade-off during learning. To this end, we propose Hierarchical Token-level Objective Control Policy Optimization (HTPO), a novel RL algorithm that takes the divide-and-conquer idea to hierarchically partition the response tokens into specific functional groups from three aspects (i.e., prompt difficulty, answer correctness, and token entropy). Within each group, according to the contributions to exploration or exploitation, we design specialized optimization objectives to facilitate the effective execution of each token's expected functionality. In this way, HTPO can achieve a more balanced exploration-exploitation trade-off. Extensive experiments on challenging reasoning benchmarks validate the superiority of our HTPO algorithm, which significantly outperforms the strong DAPO baseline (e.g., +8.6% and +6.7% on AIME'24 and AIME'25, respectively). When scaling test-time compute, the HTPO-trained model maintains a consistent performance advantage over the DAPO baseline, and the gap widens as the sampling budget increases, validating that our adaptive token-level control method fosters effective exploration without sacrificing exploitation performance. Code will be at https://github.com/xcyao00/HTPO.

Finetuning pretrained models occurs in a low-dimensional subspace of the full parameter space. Prior work has focused on characterizing this optimization subspace, but largely ignored the complementary question: why do certain directions remain unexplored during finetuning? Are these stable directions irrelevant to downstream tasks, or do they already encode task-relevant structure that requires no further adjustment? Answering this question is central to understanding how pretrained knowledge transfers. Through systematic spectral analysis across vision and language models, we show that the leading singular vectors of pretrained weight matrices remain highly stable under finetuning and are shared across unrelated downstream tasks, revealing that pretraining establishes a reusable spectral coordinate system. Models pretrained on larger datasets exhibit greater spectral stability under distribution shift or task change, directly linking pretraining scale to geometric transferability. Motivated by these findings, we propose a parameter-efficient method that freezes pretrained singular vectors and optimizes only leading spectral coefficients, achieving competitive performance on GLUE with 0.2% trainable parameters. Our results reveal that the stable directions encode transferable structure rather than irrelevant noise: successful pretraining discovers spectral bases that downstream tasks inherit and operate within.

Dataset distillation compresses a large training set into a small synthetic set that preserves downstream training utility. While most existing methods target training networks from scratch, modern visual transfer learning often uses frozen pre-trained encoders followed by lightweight linear probing. Existing distillation methods for this setting either unroll iterative linear-probe updates with trajectory-based gradient matching, or rely on closed-form formulations originally designed for from-scratch training with neural-tangent-kernel (NTK) approximations. Neither route exploits the fact that frozen-feature linear probing admits a closed-form solution determined directly by the pre-trained features themselves, with no infinite-width approximation and no inner-loop trajectory. We propose Closed-Form Linear-Probe Dataset Distillation (CLP-DD), a bilevel formulation that computes the linear probe induced by the synthetic set with a sample-space kernel ridge solver. The synthetic images are then updated by evaluating this induced classifier on real features through a temperature-scaled softmax cross-entropy, where the classifier columns act as learned class anchors in feature space. We further show that the choice of outer objective is decisive: pairing the closed-form inner solver with a standard MSE outer loss substantially underperforms trajectory-based methods, while the discriminative outer loss closes most of the gap. On ImageNet-100 with four pre-trained backbones, CLP-DD substantially improves over LGM without DSA and approaches LGM with DSA at a fraction of the computational cost. On ImageNet-1K, CLP-DD matches or surpasses LGM with DSA on three of four backbones while running roughly $14\times$ faster and using less than one-eighth of the GPU memory.