Model Pretraining

The environmental impact of training large language models (LLMs) is increasingly scrutinised, yet most published estimates focus on operational energy and disclose little about manufacturing (embodied) emissions, water consumption, or the underlying highperformance computing (HPC) infrastructure. We present a life cycle assessment (LCA) of the pre-training of Lucie 7B, an open-source multilingual Foundation Model developed by the OpenLLM-France consortium and trained on the NVIDIA H100 partition of the Jean Zay supercomputer operated by IDRIS (CNRS). The assessment is framed by the AFNOR SPEC 2314 "Frugal AI" reference and applies the Labos 1point5 methodology for greenhouse gas(GHG) accounting in computing. The study scope extends from data preparation to model validation, and integrates the full life cycle of the hardware infrastructure: manufacturing (including raw-material extraction), use (compute, temporary storage, system administration, cooling), and end-of-life. We report (i) an annual footprint of 417.5 tCO2eq for the Jean Zay H100 partition, split almost equally between manufacturing and operation; (ii) an effective intensity of 36.7 gCO2eq per H100 GPU-hour; (iii) a total training footprint of 21 tCO2eq for Lucie 7B (574 564 H100 GPU-hours), inclusive of amortised hardware manufacturing; (iv) on-site water consumption of approximately 76m3 for the training campaign and an annual Water Usage Effectiveness (WUE) of 0.07 L/kWh for IDRIS; (v) a heat-reuse factor (ERF) of 0.37 thanks to waste-heat recovery into the urban heating network. The study contributes one of the few publicly documented LCAs of an LLM training campaign that explicitly couples operational data with embodied emissions decomposed by subsystem (compute, storage, power chain, cooling), and discusses the implications for the design of frugal-by-construction AI systems in Europe.

Online task-free continual learning (TFCL) requires intelligent agents to sequentially accumulate knowledge from an unbounded, non-stationary data stream under strict single-pass constraints and without any explicit task identifiers. Existing online TFCL paradigms primarily rely on parameter-efficient prompt tuning or dynamic structure expansion driven by training-coupled optimization dynamics, such as empirical loss fluctuations or evolving latent distances. As a result, these training-coupled solvers remain agnostic to the structural origins of distribution drift, mechanically enforcing a fixed strategy across fundamentally distinct streaming variations. To address this gap, we propose LargeMonitor, a framework that leverages large pretrained foundation models to autonomously orchestrate task-free continuous adaptation. Specifically, LargeMonitor introduces a decoupled detection module utilizing the frozen, stable representation space of large vision models (LVMs) to achieve robust, zero-shot drift detection without training-dependent interference or brittle threshold tuning. Upon a confirmed drift, the framework activates a context-aware diagnostic module driven by large multimodal models (LMMs) to interpret the precise semantic etiologies of the stream variation (e.g., novel class emergence vs. environmental domain shift). This dual-stage capability empowers the continuous learner to dynamically deploy adaptive and shift-specific optimization strategies. Extensive experiments across multiple TFCL settings and benchmarks demonstrate that LargeMonitor achieves precise, robust detection and diagnosis of complex data streams while consistently improving the performance of existing online TFCL algorithms.

Large Language Models fail at implicit multi-hop reasoning: a model answers "When was $X$ born?" and "Who is $Y$'s closest friend?" correctly but fails on "When was $Y$'s closest friend born?" in a single forward pass, even when both facts are perfectly memorized and individually retrievable. We study this failure in a controlled natural language setting with a strict separation between individuals exposed to compositional contexts during pretraining and those that never appear in any such context. We confirm that compositional failure persists even at 97% 1-hop accuracy, establishing the gap as a pretraining failure rather than a knowledge absence. We propose and test nine data-centric augmentation formats and find that compositional pretraining transfers to unseen questions for exposed individuals, but never to individuals absent from compositional pretraining, suggesting that exposure to compositional contexts during pretraining is a necessary condition for implicit multi-hop reasoning. Code is available at https://github.com/ykrmm/composition-bound .

Parkinson's disease (PD) is a progressive neurodegenerative disorder that frequently causes speech impairments associated with hypokinetic dysarthria. As speech production relies on the precise coordination of complex neuromuscular mechanisms, speech analysis has emerged as a promising non-invasive and cost-effective biomarker for early PD detection. Recent deep learning approaches have shown encouraging results; however, most existing methods rely on a single speech representation, potentially overlooking complementary pathological information encoded across different feature spaces. In this work, we propose a multi-branch deep learning framework for automatic PD detection from speech. Each recording is segmented into 5-second chunks and represented using three complementary modalities: Log-Mel spectrograms, MFCCs, and HuBERT embeddings extracted from raw waveforms. The spectrograms are processed using a pre-trained ResNet-18 encoder, MFCC sequences are modeled through a BiLSTM network, and raw speech is encoded using a pre-trained HuBERT model. To effectively integrate these heterogeneous representations, we introduce a context-guided cross-modal attention mechanism that dynamically weights temporal HuBERT embeddings according to the global acoustic context derived from the spectrogram and MFCC branches. Experiments conducted on the publicly available Spanish PC-GITA corpus under strict speaker-independent 5-fold cross-validation demonstrate the effectiveness of the proposed approach. The proposed architecture achieves an accuracy of 91.51%, an F1-score of 91.24%, and an AUC of 95.97%. Furthermore, ablation studies confirm the contribution of both the proposed context-guided cross-modal attention mechanism and the integration of complementary speech representations. These findings highlight the potential of heterogeneous speech modeling for robust and clinically reliable PD detection.

EEG foundation-model releases are usually audited one endpoint at a time: raw-reconstruction, membership inference, identity linkage, or DP-SGD on the downstream head. We audit the same released embeddings under all four endpoints jointly, on BIOT, LaBraM, and EEGPT, and show that each single-endpoint audit clears releases that still leak spectral attributes. The decisive evidence is a cross-encoder transfer audit: a single ridge attribute decoder learned from one frozen encoder transfers, via a fitted linear bridge, to held-out-subject test splits of every other encoder, with subject-disjoint matched-control 95% CI lower bound at least 0.081 across all six BIOT/LaBraM/EEGPT directions. We prove a sufficient condition: two encoders sharing a nontrivial attribute-coordinate projector overlap beta admit a chained ridge bridge attacker with centered-gain lower bound sqrt(beta/(1+tau^2)) - eps_br - rho_0, and back-solve beta in [0.008, 0.198]. To turn the joint audit into a deployment-readable decision rule we introduce an audit-endpoint disagreement score (AEDS), prove sufficient conditions for its positivity, and bootstrap-calibrate it per cell; AEDS is positive in all eight matched-CI cells (BIOT/LaBraM/EEGPT on EEGMMI; LaBraM on Sleep-EDF, 54-channel LIMO, CHB-MIT pediatric scalp EEG) with p<0.001, while a head-level Carlini LiRA membership audit reaches AUC only 0.50-0.70. Standard defenses fail under audit: a Wiener-style noise-aware adaptive attacker, the LiRA audit, and DP-SGD at every utility-preserving epsilon in {4,8} leave the attribute channel essentially unchanged. The contribution is an audit framework that turns scattered single-endpoint defenses into a joint release decision, supported by a cross-encoder bridge theorem and adaptive-attacker, LiRA, and DP-SGD baselines; the audit licenses release-blocking, not raw-waveform exfiltration or held-out-subject identity recovery.

Muon replaces a matrix gradient $G=UΣV^\top$ by its polar factor $UV^\top$. This keeps the singular directions selected by the gradient, but makes the update spectrum flat. We study the optimization bias created by this operation. Under explicit alignment assumptions, we prove that the polar update is the one-step entropy-maximizing choice among bounded updates that use the gradient singular directions and do not adapt to the current weight spectrum. In an underdetermined regression model, we derive exact singular-value dynamics for continuous-time Muon and identify a measurement-dependent condition under which the normalized spectrum moves toward equal nonzero singular values. This geometry also rules out a common low-rank interpretation: at fixed Frobenius norm, Muon's distinguished state has a flat spectrum, whereas nuclear-norm minimization favors spectral concentration. Controlled matrix-sensing experiments separate the effect from simple gradient rescaling, show that norm-matched gradient descent does not reproduce Muon, and recover the predicted flattening trend across broad ablations. In small NanoGPT pretraining, Muon preserves stable rank, has a broad learning-rate plateau, and improves validation loss relative to AdamW; in a matched small-ViT control, the ranking reverses. The resulting picture is regime-dependent: Muon is not universally superior, but its flat-spectrum bias can help when many spectral directions need to remain active.

MRI preprocessing defines the input distribution seen by brain MRI foundation models, yet it is usually treated as routine data cleaning rather than a modeling choice. We ask how much preprocessing is worth its computational cost for self-supervised 3D MRI pretraining. Keeping the corpus, 3D ViT backbone, masking protocol, and downstream evaluations fixed, we compare a graded P0-P7 preprocessing spectrum for masked autoencoding (MAE) and joint-embedding predictive learning (JEPA) on 20,000 heterogeneous brain MRI volumes, then transfer the encoders to IDH prediction, MCI classification, brain age regression, and GLI/PED tumor segmentation. The results do not support a simple "more is better" rule. P0/P1 are numerically unstable, making P2 the lowest-cost feasible level; beyond P2, choosing the best feasible preprocessing level improves aggregate utility by only 3.4 percentage points for MAE and 1.8 percentage points for JEPA, with most paired gains statistically unresolved. Stronger preprocessing is beneficial only in selected regimes: IDH improves modestly, AGE and GLI/PED are often near or best at P2, and MCI shows the clearest empirical P7 gain. Cross-level MCI transfer further shows that much of the P7 advantage can be recovered by applying stronger preprocessing downstream, without requiring P7 throughout pretraining. These findings recast MRI preprocessing as a downstream-aware cost-utility decision rather than a default escalation pipeline. Code is available at https://github.com/PangJiangShuan/PreBrain.

Choosing the wrong synthetic generator for time-series foundation model pretraining is costly: under identical training budgets, the best and worst generators produce up to a $2\times$ gap in forecasting error, yet the field has no principled way to make this choice. The problem is compounded by the fact that generator rankings are not stable across architectures: across 11 generator families evaluated on Chronos-T5-Mini and Moirai-Small trained from scratch, we find that which generators are useful depends on the model architecture. Rather than solving the generator selection problem, we sidestep it: a simple equal-weight mixture of all generators matches or beats the best individual generator for both architectures, and composing this mixture with real data yields the strongest pretraining corpora overall. Synthetic pretraining is therefore a corpus composition problem, not a generator selection problem, and composition choices should be validated per model family rather than assumed to transfer.

Pretraining is fundamental to the development of Large Language Models (LLMs), yet the opacity of pretraining data complicates model analysis and raises ethical, legal, and fairness concerns. Detecting whether specific datasets were used during pretraining is, therefore, critical. Existing state-of-the-art methods typically rely on access to model probability distributions, making them unsuitable for closed-source LLMs that provide only input-output interfaces. To address this limitation, we introduce Masked Corpus-level Pretraining Data Detection (MC-PDD), a novel method inspired by the masked language modeling paradigm. MC-PDD masks highly specific tokens in each text and prompts the LLM to predict the missing content. It then assesses whether the difference in prediction hit rates between a candidate corpus and a reference non-member corpus is statistically significant. Based on this comparison, MC-PDD determines whether the candidate texts were likely included in the model's pretraining data. Experimental results demonstrate clear and consistent differences in prediction hit rates between pretrained and unseen data across three datasets, for both open-source and closed-source LLMs. Despite operating under a stricter black-box setting, MC-PDD achieves performance comparable to existing detection methods. Our approach enables practical applications such as model auditing and data copyright verification using only standard API access. Upon acceptance, we will publicly release the code and datasets.

Scalable data attribution methods typically assign isolated utility scores to individual training examples. This prevalent additive assumption fundamentally fails to capture critical subset dynamics, including data redundancy and complementary coverage. In this work, we reframe attribution as subset-level counterfactual utility prediction and introduce GRASP, an interaction-aware surrogate. Grounded in a theoretical smoothness lower bound, GRASP explicitly models subset interactions through a quadratic geometric penalty. To achieve pretraining-scale efficiency without relying on hidden oracle tuning, we couple low-dimensional feature sketches with a strictly finite lower-confidence bound selection protocol. Extensive subset-retraining evaluations demonstrate that GRASP decisively outperforms existing scalable baselines. It more than doubles the task-level rank correlation for counterfactual subset fidelity while reducing upfront artifact construction costs by nearly an order of magnitude. Downstream diagnostics further show that this scoring mechanism transfers to language model curation and cross-domain vision selection, establishing a robust foundation for optimizing massive pretraining corpora.

Classical scaling laws for language model pretraining balance model size against training dataset size under a fixed compute budget, assuming abundant data and a single pass over the corpus. As training compute grows faster than the supply of natural language data, pretraining is likely to enter a data-constrained, compute-rich regime where models train for multiple epochs over a finite dataset. We study data-constrained pretraining along two axes, regularization and scaling. For regularization, we study masked-input regularization (MIR), an auxiliary next-token prediction loss on randomly masked inputs. MIR tests whether the random masking central to diffusion language models can benefit autoregressive pretraining without architectural changes or inference overhead. Across 72M to 1.4B parameter models, we find that MIR added on top of strong weight decay improves validation loss over autoregressive strong-weight-decay-only models, with downstream gains at 1.4B. For scaling, we propose SoftQ, a scaling law that couples model size and data size to capture their interaction under repeated data. Classical alternatives such as the Chinchilla law use an additive form that decouples these terms, making them misspecified in the data-constrained regime. We find that SoftQ fits data-constrained experiments substantially better than these alternatives, and estimates MIR's gains as equivalent to roughly 1.3 times as much unique training data. We release our code at https://github.com/yixinw-lab/dc_pretrain.

Egocentric human video offers a scalable alternative to robot data for pretraining, yet models pretrained on such video consistently underperform those pretrained on robot data. We attribute this gap to a missing signal, the active perception behavior in egocentric videos, where humans continuously reposition their viewpoint during manipulation, inducing camera motion that standard pipelines treat as noise. To address this, we present ActiveMimic, a pretraining framework that recovers synchronized camera and wrist trajectories from a single body-worn RGB camera, models camera motion as a viewpoint action, and jointly learns active perception and manipulation from in-the-wild egocentric human video before adapting to a target robot. Empirically, real-world experiments across tasks with diverse active perception demands show that ActiveMimic consistently surpasses baselines pretrained on human video and matches state-of-the-art models pretrained on robot data. Further analysis provides evidence that active perception capability originates from egocentric human video pretraining rather than robot-specific fine-tuning, confirming active perception as the key to unlocking egocentric human video for robot pretraining.

Modern language models rely on pretraining filters to remove undesirable content from training corpora and inference-time guardrails to suppress undesirable outputs during deployment. In this paper, we examine how these filtering and moderation decisions produce forms of epistemic erasure and reveal tensions both across automated systems and between these systems and human judgment. We audit four pretraining filters and three inference-time guardrails on Common Crawl sentences containing gender and regional-origin mentions, together with a manually annotated subset of 500 sentences. Our analysis shows that filtering and guardrail decisions are strongly associated with blocklist-based lexical cues, while frequently failing to flag content containing private information or explicit hate speech. At the same time, marginalized groups, particularly transgender people, women, and Central Americans, are significantly over-flagged across systems. Human annotators, by contrast, would retain 88.5% of filter-flagged and 91.3% of guardrail-flagged content, often recognizing representational harms arising from tensions of content removal that current systems fail to capture. Taken together, our findings document a form of epistemic erasure in which mentions of marginalized groups are disproportionately removed before pretraining and additionally suppressed again at inference time.

Longitudinal treatment decisions from multivariate time-series data require predicting potential outcomes under future treatment sequences in the presence of time-varying confounding, heterogeneous patient dynamics, and limited domain-specific data. Existing longitudinal causal estimators typically address this problem by training a new model for each cohort or simulator. We introduce Causal Longitudinal Prior-Fitted Networks (CausalLongPFN), a prior-fitted network for time-series causal inference in longitudinal treatment-response data and zero-shot in-context counterfactual outcome prediction. The model is pretrained entirely on synthetic episodes sampled from a broad prior over temporal structural causal models, exposing it to treatment-confounder feedback, latent heterogeneity, nonlinear state evolution, delayed effects, and cumulative treatment responses. At test time, CausalLongPFN remains frozen and is used zero-shot: it conditions on support trajectories, a query history, and a planned future treatment sequence, and returns a predictive distribution over future outcomes without gradient updates or propensity-model fitting. Multi-step predictions are obtained by recursively applying the one-step predictor under the specified treatment sequence. We evaluate the model on branchable cancer, HIV, and warfarin benchmarks with ground-truth counterfactual labels, and on factual-only rolling-origin prediction in MIMIC-III ICU trajectories. CausalLongPFN is competitive with domain-trained longitudinal baselines on counterfactual benchmarks and performs strongly on factual MIMIC-III prediction, suggesting that broad synthetic causal pretraining can provide a frozen, amortized alternative for zero-shot longitudinal treatment-response prediction when repeated domain-specific training is costly or impractical.

The efficacy of continued pre-training for Large Language Models (LLMs) hinges upon hyperparameter configurations, such as learning rate and batch size. However, current practices often rely on heuristics or grid searches, leading to training instability and excessive costs. In this work, we first empirically discover that optimal hyperparameters follow stable and predictable scaling laws throughout the continued pre-training process. Leveraging these insights, we propose a novel framework to establish quantitative relationships between compute budget and optimal hyperparameters for a given checkpoint. Our approach has two stages: (1) \textit{Empirical Law Discovery}, where we train small-scale proxy models to derive functions mapping compute budget to optimal hyperparameters via standard loss-compute scaling laws; and (2) \textit{State-Aware Hyperparameter Prediction}, where we evaluate an initial checkpoint's validation loss and use the inverse scaling law to estimate its \textit{equivalent pre-training compute} -- the compute needed to achieve the same loss from scratch. Combining this with the planned compute budget, we predict optimal hyperparameters for the target run. Empirical results demonstrate that our method reduces the hyperparameter search overhead by up to 90% while achieving comparable or superior performance relative to baselines. This model-agnostic framework generalizes across architectures, providing a principled and efficient methodology for diverse continued pre-training scenarios starting from any given point.

Every new clinical imaging device creates a domain shift where dense gland masks are expensive yet cheap clinical signals -- eyelid outlines, Pult grades, morphometric ratios -- are routinely recorded. We present TopoPult-SSL, a two-stage framework for cross-device meibomian gland segmentation. Stage 1 adapts a source-trained model without target gland masks in the training loss, using four weak-prior anchors driven by target eyelid masks and clinical metadata only. Stage 2, when target gland masks are available, distils complementary Stage-1 teachers into a single compact student via supervised self-distillation. We develop and validate the technique on the public MGD-1k to CAMG research benchmark (1,000 to 100 images, different device), where the distilled model achieves Dice 0.716+/-0.006 (best 0.726), surpassing UA-MT (0.710) and the ensemble teacher (0.720) -- with a single pass. The gland-mask-free Stage-1 variant reaches Precision 0.694 vs. 0.30-0.34 for SAM/MedSAM (p<0.001), enabling deployment without dense gland contouring. Code and reproducibility scripts are released.

Tabular foundation models, such as TabPFN, achieve strong performance on tabular datasets with numerical and categorical data, but do not natively handle high-cardinality text features. Standard pipelines, therefore, embed text with a language model and compress the resulting vectors with PCA into a small number of scalar features before inputting them into TabPFN. This creates an information bottleneck: most embedding dimensions are discarded, and the compressed representation must then be expanded again by TabPFN's feature encoder. End-to-end alternatives can avoid PCA, but they require large amounts of pretraining data containing text cells and usually perform subpar compared to tabular foundation models that were pretrained on large amounts of synthetic data. Inspired by modality-alignment approaches like LLaVA (vision-to-LLM token projection) and TableGPT-style systems (table-to-LLM token projection), we introduce the TabPFN Text Adapter (text-to-TFM token projection). We freeze both the sentence encoder and TabPFN, and train only a lightweight adapter that maps text embeddings into a short sequence of tokens in TabPFN's embedding space. This design removes the PCA bottleneck, preserves TabPFN's numerical strengths, and is more efficient to train than end-to-end text-tabular pipelines.

The standard LLM training pipeline applies reinforcement learning (RL) only after pre-training and supervised fine-tuning (SFT). We question this status quo by training a LLM from scratch and applying RL, SFT, and SFT followed by RL directly to intermediate pre-training checkpoints. We find that RL is effective very early, and often matches the full SFT$\to$RL pipeline early as well. Through experiments on harder problems, we find that targeted pre-training data composition is a strong lever for RL effectiveness, even more so than model scale. Beyond reasoning accuracy, applying RL directly to base checkpoints expands the model's distribution; the sharpening effect reported in recent work arises only when RL follows SFT. The general capabilities of the model remain essentially unchanged by RL, while they degrade following SFT. Finally, we merge RL and SFT objectives by parallel averaging, which outperforms across all other training methods discussed, across metrics, while preserving general capabilities. Together, these results suggest that LLM training might benefit from an expanded use of RL.

Data samples used for training often differ from those encountered during fine-tuning and deployment, and while ML models show promise, their performance remains limited when only small annotated datasets are available. Performance often degrades under distribution shifts caused by diverse sensors, populations, and application settings. Although pre-training helps, models frequently encounter out-of-distribution (OOD) data in real-world settings, leading to reduced robustness. Existing adaptation methods usually assume fixed distribution shifts and struggle when multiple types or severities occur. In particular, they overlook shift severity, for example treating adaptation to a large familiar dataset the same as adaptation to a small dataset with a new task, which limits generalisation. To address this, we propose ADAPTOOD, a novel framework that leverages data uncertainty to quantify distribution shift severity and guide fine-tuning for time series. This uncertainty measures how strongly samples from the target deployment distribution deviate from the pre-training distribution, providing a direct signal of OOD severity. Our framework combines this uncertainty with low-rank model updates and adaptive hyperparameter optimisation to improve adaptation. We show that ADAPTOOD achieves up to 7% higher accuracy and 12.9% higher precision than existing methods in OOD tasks, maintaining strong performance as distribution shift severity increases.

Multi-epoch training is becoming the standard now that compute is growing faster than the supply of high-quality text. But pretraining a single model saturates within a few passes, long before the compute budget is exhausted. We argue this calls for a conceptual shift from training a single model toward exploring a population of models and aggregating their predictions. We introduce hyper-epoch pretraining (q0), which turns a multi-epoch budget into a population of diverse models whose combined predictions reach a lower validation loss than a single refined model. q0 reduces to three core primitives. A cyclic schedule with anti-correlated learning rate and weight decay collects diverse models from a few parallel trajectories. Chain distillation trains each model against its predecessor so that model quality compounds across the population. A learned prior, fit on a held out set, selects and weights members for any inference budget. On a 1.8B-parameter model trained on 100M FineWeb tokens, q0 matches a strong 256-epoch ensemble baseline using only ~56 epochs (~4.6x fewer), or ~67 epochs (~3.8x fewer) when matched to the baseline's ensemble size, and continues to improve beyond it. These gains reach cumulative ~12.9x data efficiency under the Slowrun setting and transfer to downstream benchmarks. Crucially, the optimal allocation shifts with the budget, so we give prescriptive recipes for how to spend a given epoch budget to maximize generalization, from a single epoch up to the largest budgets.

High-quality pretraining data is a central ingredient in modern language models, but German-language resources remain far less developed than their English counterparts: they are often smaller, less carefully curated, weakly documented, and rarely validated through controlled training experiments. We introduce KletterMix, a high-quality German corpus for language model pretraining and annealing, designed as a reusable dataset artifact for the natural language processing and modeling community. KletterMix is built by translating a state-of-the-art English pretraining corpus into German while preserving document boundaries, metadata, source structure, and topical diversity. This construction yields a German corpus with the scale and diversity of a modern pretraining dataset, while enabling direct comparison to its English source. We document the dataset through a broad set of corpus-level analyses, including translation quality, document length distributions, topic coverage, source composition, and geographic metadata. Using COMETKiwi, we show that the translated documents achieve strong quality across diverse domains, suggesting that careful translation can preserve much of the semantic and stylistic richness of the original corpus. Beyond dataset construction, we evaluate KletterMix as training data. Through controlled pretraining and annealing ablations against established German corpora, we show that models trained on KletterMix achieve measurable improvements on German-language downstream evaluations. These results demonstrate that carefully curated translated data can substantially strengthen the German pretraining data ecosystem.

Pretraining cost is a major bottleneck for research on tabular foundation models, slowing the iteration cycle for new architectures, priors, and optimization ideas. Yet the community lacks a simple way to compare and accumulate pretraining speedups. We introduce a community speedrun for nanoTabPFN: contributors modify a single-file training script and compete to reach a fixed downstream ROC AUC target on subsampled TabArena using one NVIDIA L40S GPU. The current best record reaches the target in 0.92 minutes, an 81x speedup over the 74.32 minute baseline while using 22x fewer synthetic datasets. The speedrun format provides a simple protocol for the community to add, verify, and stack pretraining improvements, with the leaderboard open to contributions. Code and records are available at https://github.com/borawhocodess/modded-nanotabpfn.

Digital healthcare generates vast amounts of clinical text that can support AI-assisted applications, yet German biomedical language models remain limited by older architectures or restricted training data. We present ChristBERT (Clinical- and Healthcare-Related Issues and Subjects Tuned BERT), a family of domain-specific German RoBERTa-based language models trained on a 13.5GB corpus of scientific publications, clinical texts, health-related web content, and translated clinical resources. To investigate the impact of domain adaptation strategies in German clinical NLP, we compare continued pre-training, training from scratch, and domain-specific vocabulary adaptation. The resulting models are evaluated on three medical named entity recognition tasks and two text classification tasks. ChristBERT consistently outperforms existing general-purpose and medical German language models on four of five benchmarks and establishes a new state of the art for German clinical language modeling. Our results show that the optimal adaptation strategy is task-dependent: in our evaluation, training from scratch is particularly effective for highly specialized clinical texts, whereas continued pre-training performs well on more commonly written medical texts. All models are publicly released to support future research and applications in German medical NLP.

Pre-training has become a fundamental paradigm in modern machine learning, with one of its key empirical benefits being reduced downstream sample complexity as the scale of pre-training data increases. However, existing theoretical frameworks for pre-training do not fully explain this phenomenon. In this paper, we introduce complexity minimization, a novel meta-representation learning framework designed to enable theoretical analysis of this scaling behavior, which learns representations by evaluating the downstream model complexity best suited to each domain and minimizing the worst-case such complexity across source domains. Our end-to-end theoretical analysis, spanning pre-training through downstream regression, shows that this framework provably captures this scaling behavior; in particular, we show that the error rate of few-shot adaptation improves as the amount of meta-training data grows. Empirically, we demonstrate that incorporating complexity regularization into existing meta-learning methods consistently improves downstream sample efficiency.

Research and applications in artificial intelligence have recently shifted with the rise of large pretrained models, which deliver state-of-the-art results across numerous tasks. However, the substantial increase in parameters introduces a need for parameter-efficient training strategies. Despite significant advancements, limited research has explored parameter-efficient fine-tuning (PEFT) methods in the context of transformer-based models for instance segmentation. Addressing this gap, this study investigates the effectiveness of PEFT methods, specifically adapters and Low-Rank Adaptation (LoRA), applied to two models across four benchmark datasets. Integrating sequentially arranged adapter modules and applying LoRA to deformable attention--explored here for the first time--achieves competitive performance while fine-tuning only about 1-6% of model parameters, a marked improvement over the 40-55% required in traditional fine-tuning. Key findings indicate that using 2-3 adapters per transformer block offers an optimal balance of performance and efficiency. Furthermore, LoRA, exhibits strong parameter efficiency when applied to deformable attention, and in certain cases surpasses adapter configurations. These results show that the impact of PEFT techniques varies based on dataset complexity and model architecture, underscoring the importance of context-specific tuning. Overall, this work demonstrates the potential of PEFT to enable scalable, customizable, and computationally efficient transfer learning for instance segmentation tasks.

Real-world domains often contain heterogeneous tables whose headers vary while their underlying attribute semantics are shared, making it difficult to induce domain-specialized semantics from table-local evidence alone. Existing encoders model parts of this problem, but often underuse column-level value distributions and apply uniform objectives across attributes with different semantic roles. We propose NAVI, a segment-centric pretraining framework that treats each header-value pair as the unit for aggregating schema-level structural evidence and column-level distributional evidence. We realize this design through Masked Segment Modeling and Entropy-driven Segment Alignment, which jointly enforce structured header-value coupling and semantic alignment across stable and instance-specific attributes. Experiments on heterogeneous in-domain tables show improved reconstruction, semantic consistency, and downstream utility across evaluation settings overall.

Zero-shot time series forecasting aims to predict future values for previously unseen series, requiring models to generalize temporal dynamics beyond the training distribution. While recent foundation models achieve strong in-domain performance through large-scale pretraining, their effectiveness often relies on broad data coverage and implicit pattern memorization, which can limit generalization when data are scarce or source and target domains are disjoint. In this work, we propose FSA, a feature-to-strategy framework for controlled zero-shot univariate forecasting. Instead of directly modeling raw sequences in the observation space, FSA learns a structured mapping from an interpretable feature space to an autoregressive strategy space. This design introduces explicit inductive biases that disentangle global trends, periodic components, and local temporal dynamics, enabling the model to capture transferable time-series structure with fewer data assumptions. Empirical results show that, under identical pretraining data, training protocol, and comparable parameter budgets, FSA outperforms Transformer-based architectures in our controlled zero-shot setting.

Despite the growing availability of cryo-electron microscopy (cryo-EM) density maps, effectively leveraging them for protein representation remains challenging. First, current methods lack a general-purpose protein pretraining framework tailored for cryo-EM density maps, designed for protein-related property prediction. Second, existing approaches typically partition density maps into local box regions and model them independently, overlooking interactions across boxes which are essential for capturing global structural context in cryo-EM density map. To address these challenges, we propose CryoProt, a protein pretraining framework designed for cryo-EM density maps. CryoProt introduces a Map Encoder based on multi-head latent attention (MLA), where box-level representations interact through a shared latent space, enabling explicit modeling of cross-box dependencies within the density map. Furthermore, we adopt a multi-task pretraining strategy to learn generalizable representations that can be effectively transferred to diverse downstream tasks, such as protein flexibility prediction, where cryo-EM density maps are not required and can be inferred implicitly by the pretrained model. Experimental results demonstrate that CryoProt consistently outperforms existing state-of-the-art methods across multiple benchmarks, achieving up to 12% improvement over the best-performing baselines, highlighting the importance of modeling cross-box interactions in cryo-EM data. The source code is publicly available at https://anonymous.4open.science/r/CryoProt.

Electroencephalography (EEG) supports a variety of brain-computer interface (BCI) tasks ranging from brain-state monitoring to human-LLM interactions. EEG foundation models are emerging, but evaluation remains fragmented due to heterogeneous datasets and nconsistent task protocols. Here, we introduce OmniEEG-Bench, a unified benchmark and downstream task roadmap for EEG foundation models (FMs). It organizes evaluation of EEG FMs into six task families spanning (i) signal reliability, (ii) biometrics and disease, (iii) consciousness and state, (iv) cognition and emotion, (v) naturalistic stimulus decoding, and (vi) motor and interaction, introducing a new generation of tasks not systematically benchmarked in prior EEG FM work. OmniEEG-Bench standardizes model deployment, task definitions, and metrics through a task-card specification, and unifies 54 EEG datasets with consistent evaluation protocols. We benchmark 10 representative EEG foundation models and report a leaderboard that covers diverse evaluation settings. Both pretraining dataset diversity and model size are significantly associated with better average ranks across datasets, revealing scaling-law behavior in EEG foundation models (Figure 1). These results suggest that scaling EEG foundation models requires not only larger architectures but also broader and more diverse pretraining data. The benchmark code is available at https://github.com/ncclab-sustech/omni-eegbench.git.

While continual pretraining~(CPT) is a practical way to extend large language models to new languages, naïve finetuning on targeted data erodes existing capabilities through catastrophic forgetting. Organizing training around language families reduces cross-language interference but cannot alone prevent forgetting of the general knowledge needed for downstream tasks. We link this forgetting to parameter drift in multilingual CPT and present a suite of five layer-aware parameter alignment strategies: hard layer freezing, soft regularization, post-hoc weight reversion, and model merging. We systematically compare our alignment strategies against two unregularized CPT baselines on benchmarks spanning 32 training languages from five language families, plus held-out languages, across four evaluation axes: perplexity, reading comprehension, physical reasoning, and translation. Parameter alignment substantially reduces forgetting at minimal cost to language acquisition: layer freezing and regularization best preserve comprehension, whereas post-hoc reversion yields the strongest translation gains. Together, these results map the acquisition--forgetting frontier for family-expert CPT and offer practical deployment guidelines pairing each strategy to the tasks it best serves.

Grokking, the phenomenon in which neural networks generalize long after fitting their training data, has been studied in supervised settings on many epochs. LLM pre-training instead involves next-token prediction over an unlabeled corpus, with limited data repetition and no explicit train/validation split. To address this, we propose an exposure-based framework that enables the study of grokking-like dynamics during LLM pre-training. We ground our evaluation in BLiMP minimal pairs, which provide controlled grammatical contrasts. For every BLiMP minimal pair, we identify a critical phrase, the smallest continuous span that captures the grammatical contrast and the phenomenon-relevant context. Examples whose critical phrase appears in the pre-training window are assigned to the proxy-train split; the remaining examples are assigned to the proxy-validation split. Across five grammatical phenomena, we observe delayed generalization. Analyzing pre-training checkpoints before and after generalization shows that grammatical concept vectors become more predictive of grammatical acceptability and occupy a higher-dimensional subspace after generalization. We also find that attention from the critical token to the relevant context token is concentrated in a small number of heads.

Modern network intrusion detection systems (NIDS) are caught in a structural contradiction: the protocols carrying the highest threat intelligence are precisely those encrypted under TLS 1.3 and QUIC, where payload inspection yields nothing. We ask a simpler question -- what if the attack signature is not in the bytes, but in the rhythm? -- and answer it by treating network flows as a language whose grammar is written entirely in L3/L4 packet metadata: length, inter-arrival time, TTL, TCP flags, and hashed port numbers. We present PLM-NIDS, which proves three claims in sequence. (1) The grammar exists and is learnable: a RWKV-4 state-space model trained on 344,232 unlabelled Monday flows achieves a causal LM validation loss of 0.204, demonstrating that benign traffic has predictable, statistically consistent structure. (2) Attacks violate this grammar: the per-flow perplexity score cleanly separates benign from attack flows with PR-AUC = 0.93 using zero attack labels at training time. (3) This separation is architecturally nontrivial: an LSTM trained on identical token sequences degenerates to a majority-class predictor (ROC-AUC approximately 0.50, F1 = 0.91 by always predicting "attack"), proving that RWKV's causal pre-training provides an inductive bias unavailable to direct classifiers. Supervised fine-tuning further raises PR-AUC to 0.94 and ROC-AUC to 0.75, with a precision of 97.7% at the calibrated operating threshold. The RWKV backbone's O(T) recurrent inference enables per-packet streaming without flow buffering, making PLM-NIDS operationally viable at line rate. Because it reads only IP/TCP/UDP headers, it is inherently encryption-agnostic: TLS 1.3, QUIC, and future encrypted protocols are handled transparently.

Pre-training data mixtures are commonly tuned by running small-scale experiments and extrapolating to the target training budget. When high-quality data is scarce and must be repeated, this extrapolation frequently fails, but the source of the failure has not been isolated. We show that a primary culprit is a repetition mismatch: because high-quality datasets are small, their repetition rate changes as the training budget grows, shifting the optimal mixture in ways that small-scale proxy experiments do not anticipate. A subsampling procedure that matches the target repetition rate controls for this effect. In a two-source setting combining limited high-quality data with web crawl, a single repetition-controlled experiment using only 1/16 of the target tokens recovers a mixture within 0.10 of the optimum on Wiki-Text for a 1.17B parameter model, compared to an error of 0.85 without repetition control. Achieving comparable accuracy without repetition control requires multiple training horizons, consuming 19%, 44%, and 94% of the target token budget when using the results from two, three, and four horizons respectively. With three data sources, the larger mixture space requires more than a single experiment to constrain, but the approach remains effective: at the 757M scale, just two repetition-controlled horizons recover the optimal mixture, outperforming baselines that instead require the full two-source experiments to construct. Our results reveal that repetition dynamics, not scale alone, shape whether small-scale mixture experiments generalize. More broadly, they suggest that data repetition deserves treatment as a first-class variable in mixture optimization, rather than an inconvenient side effect of limited data.

Tool calling extends large language models (LLMs) by enabling grounded interaction with external executable interfaces, thereby supporting environment-coupled problem solving. However, mainstream in-context learning (ICL) approaches typically incorporate detailed tool documentation and usage examples directly into the context. This results in substantial inference overhead and heightened risks of hallucination as the context length grows. Conversely, while tuning-based methods improve general tool-calling capabilities, they often fail to effectively internalize the specific details of previously seen tools, thereby retaining a dependency on in-context documentation. To address these limitations, we propose ParaTool, a framework that projects each tool into a dedicated, loadable set of parameters. By equipping a dynamic integration of these parameterized tools, the LLM can perform tool calling without relying on in-context documents or examples. Specifically, our approach consists of three stages: (1) parametric tool pre-training encapsulates the knowledge of different tools into independent parameter modules; (2) soft tool selection employs a gating network to dynamically weigh and aggregate relevant tool parameters; and (3) parametric tool fine-tuning jointly updates tool parameters to align the training and inference processes. Experiments on Stable ToolBench and BFCL demonstrate that ParaTool significantly outperforms strong ICL-based baselines, achieving superior performance while reducing computational complexity.

AI model hubs provide access to a rapidly growing collection of powerful pre-trained models, enabling off-the-shelf mixture-of-experts systems with different routing strategies. However, this rapid growth poses two fundamental challenges: scaling model selection across thousands of experts and continually updating routing mechanisms as new models and tasks are introduced. In this paper, we formalise this setting as Continual Model Routing (CMR) and propose CMRBench, a new large-scale benchmark simulating realistic hub expansion and including over 2,000 candidate models. Finally, we introduce CARvE, a contrastive embedding approach for efficient continual model routing via checkpoint-based anchoring and structured replay. Extensive empirical results and ablations show that CARvE significantly outperforms zero-shot retrieval, fine-tuning, and adapter-merging baselines in model, family, and domain-level accuracy.

The massive scaling of Large Language Models (LLMs) has made pretraining increasingly cost-prohibitive. While low-rank representation and orthonormal weight matrices could in principle reduce parameter counts and computational overhead, most existing methods rely on static rank selection and do not enforce weight orthonormality due to high computational cost. This paper introduces TSVD, a framework that maintains low rank and strict orthonormality throughout the training process. It utilizes a spectral energy-based heuristic for adaptive rank selection, and a caching mechanisms to maintain orthonormality. Theoretical analysis justifies the advantage of the approach in pretraining dynamics and experiments across various model scales demonstrate that it is effective empirically. TSVD matches or exceeds the performance of full-parameter baselines while significantly reducing compute requirements. The approach thus offers a well-founded, practical, and scalable path toward efficient high-performance LLM pretraining.

Fine-tuning large-scale pre-trained models is a recent prevalent paradigm for adapting general representations to specialized tasks. However, when a new version of a pre-trained model becomes available, expertise acquired through fine-tuning cannot be directly reused because it is tied to the parameterization of the original model, requiring another costly fine-tuning. To address this inefficiency, recent work uses task vectors, defined as the parameter difference between a fine-tuned model and its base model, to transfer expertise across models. While existing methods bridge disparate models by matching activations or gradients, a significant performance gap remains relative to direct fine-tuning, suggesting that these partial correspondences are insufficient. In this work, instead of viewing a task vector merely as a parameter offset, we revisit the formation of task vectors and show that they can be derived as accumulated bilinear interactions between input-side activations and output-side gradients. Motivated by this observation, we formulate task-vector transfer as a dual-space alignment problem and propose BiCo, a training-free framework for transferring task vectors through Bilinear Coordinate alignment. BiCo estimates orthogonal Procrustes mappings in both spaces using a single forward-backward pass on a small calibration set, without any parameter update. Across extensive computer vision and natural language processing benchmarks, BiCo consistently outperforms existing transfer methods across models that differ in width, depth, and pre-training configuration.

Predicting how a cell will change its transcriptional state under a developmental signal or a genetic perturbation is the computational core of in-silico biology and the AI Virtual Cell program. Existing approaches either fit static control-to-treated maps that discard time, or solve multi-step ODE / Schrödinger-bridge problems on each dataset independently. We introduce Chreode, a one-step cell world model that predicts action-conditioned cell-state transitions through a structured residual transition operator. It shifts distributional evolution from inference time to training time, enabling single-pass generation while preserving a Waddington-inspired decomposition into downhill landscape flow, rotational in-tangent dynamics, and stochastic spread. The model is pretrained with a shared scVI encoder and a DiT-based dynamics backbone on a 2.4M-cell mouse embryonic atlas spanning 7 datasets. As a fine-tuning initialization, Chreode improves per-target Sinkhorn distance on Weinreb hematopoiesis and Veres islet differentiation over matched scratch models, PI-SDE, and PRESCIENT. As a transferable gene-state embedding for GEARS, the pretrained dynamics representation reduces shared-vocabulary DE20 mean squared error on Norman Perturb-seq from 0.2121 to 0.1858, a 12.4% relative improvement, without changing the GEARS training procedure. We interpret this transfer to perturbation prediction as evidence that pretrained developmental-trajectory dynamics encode differentiation primitives transferable to CRISPR-induced state shifts, since both involve cell-state transitions in a shared latent geometry. The pretrained backbone additionally produces zero-shot clonal fate scores on Weinreb that are competitive with strong dynamic-OT baselines.

With the rapid progress of large language models (LLMs), reliably evaluating the capabilities of pre-trained LLMs has become increasingly important. The challenge is that base pre-trained models are optimized for next-token prediction and often fail to follow instructions or produce well-formed answers under standard prompting and direct decoding. As a result, benchmark performance can conflate model capability with decoding-induced failures to produce task-oriented outputs, while exposing such behavior often relies on costly post-training. Recent decodingonly approaches attempt to reshape output distributions, but such methods can be inefficient and brittle across open-ended tasks. To address these limitations, we propose Energy-Based Decoding (EBD), a training-free, reward-guided framework for activating task-oriented behaviors from frozen pre-trained LLMs across both open-ended and objective tasks. EBD augments decoding with an external lightweight reward model, steering generations toward high-utility responses while anchoring them to the pre-trained model prior through a reward-tilted target distribution. We show that EBD shifts base-model outputs toward more instructionfollowing behavior, increasing behavioral similarity to post-trained counterparts and enabling a fairer inference-time evaluation of accessible pre-trained-model behavior. Empirically, EBD outperforms baselines across five models and six benchmarks, improving Qwen3-8B-Base on AlpacaEval2.0 from 8.8 to 44.5, reducing Mistral-7B Math500 latency by 18.9x relative to prior decoding work, and remaining robust to reward-model size.

The rapid advancement of diffusion-based image generation models has raised serious concerns regarding potential copyright and privacy infringements involving human-created data. Membership inference attacks (MIAs) have emerged as a promising tool for identifying unauthorized data usage during model training. Existing methods typically assess the ability of model to denoise perturbed suspect images as an indicator of membership status. However, the discriminative power of such features is highly dependent on the degree of model memorization and deteriorates significantly when applied to less exposed data (e.g., pre-training data). Although several methods attempt to enhance detection by leveraging internal model features, these features are generally inaccessible in mainstream closed-source image generation platforms, limiting their practicality. In this paper, we demonstrate that analyzing how a black-box diffusion model denoises a target image and corresponding perturbed textual instructions can reveal more distinctive membership cues. Based on this insight, we propose a black-box membership inference attack framework (named SD-MIA) that leverages a cross-modal data perturbation mechanism to detect pre-training data in diffusion models. We conduct extensive experiments on both a public benchmark dataset and a newly constructed dataset, each comprising pre-training membership and non-membership samples with identical distributions. Experimental results demonstrate that SD-MIA achieves superior performance compared to existing baselines, including those with the unfair advantage of accessing internal model features.

Modern time series architectures face a fundamental trade-off: channel-independent models scale well with increasing data volume but ignore critical inter-channel dependencies, while channel-dependent models are expressive but remain ``dimension-bounded'', struggling to generalize across heterogeneous datasets.To bridge this gap, we introduce Unicorn (Universal Correlation Network), a framework for scalable, multi-dataset pretraining on high-dimensional time series. At the core of Unicorn is a latent prototype codebook that decouples correlation modeling from specific channel identities. By projecting heterogeneous channels into a shared latent space, UniCorN learns identity-agnostic, reusable interaction patterns that transfer across domains with diverse dimensionalities and semantics. Extensive experiments show that Unicorn significantly outperforms state-of-the-art forecasting architectures, particularly in few-shot transfer scenarios, offering a scalable path toward multivariate time series foundation models.

Causal discovery from time series is critical for many real-world applications, such as tracing the root causes of anomalies. Existing approaches typically rely on dataset-specific optimization, making it difficult to transfer their causal discovery capabilities to new time series governed by diverse causal mechanisms. In this paper, we propose \textbf{PTCD}, a novel \textbf{P}retraining framework for \textbf{T}ime-series \textbf{C}ausal \textbf{D}iscovery, which improves cross-task generalization through context-conditioned modeling and transferable causal augmentation. To model complex temporal causal dependencies, PTCD employs a dual-scale iterative attention mechanism to capture window-level causal relationships, and a Gaussian mixture with a context-level routing mechanism to handle heterogeneous exogenous distributions. To further address distribution shifts across causal graphs, PTCD adopts a pretraining paradigm on synthetic datasets that integrates intervention-based learning and a causal mixup strategy, promoting stable causal discovery and stronger generalization. Extensive experiments on multiple real-world out-of-distribution (OOD) datasets demonstrate that PTCD excels in both causal discovery and root cause identification.

Large language model pre-training typically exhibits a two-phase trajectory: a fast initial loss drop followed by a prolonged slow improvement. We identify an underlying spectral phenomenon, Stability of Singular Distribution (SoSD), where the trace-normalized singular value spectrum stabilizes early, even as parameter matrices continue to evolve. We demonstrate that synchronization between SoSD and the slow-descent regime is widely observed across diverse architectures (GPT-2, LLaMA) and settings, including various schedules (Step-wise, WSD, Cosine Decay), weight decays, and optimizers (AdamW, Muon). By analyzing a simplified Transformer, we prove that growing weight norms inevitably precipitate an early SoSD threshold, after which the rate of loss decrease becomes theoretically bounded by the variation in the singular distribution. We further interpret strategies like WSD and Muon through their ability to modulate the SoSD scale, offering a spectral lens for understanding efficient pre-training dynamics.

Clinical time-series learning is routinely constrained by small, heterogeneous cohorts and protocol drift, while its downstream use spans both classification (e.g., pathology diagnosis) and regression (e.g., temporal forecasting). These constraints make foundation-model pretraining appealing, but raises an important question of which inductive biases should the pretraining objective impose so that representations transfer across task types and subjects. We study this question in pathological gait analysis for spinal cord injury (SCI) via PathoFM, an encoder-centric transformer pretrained on multivariate gait windows with three complementary objectives: Local Completion (reconstruct contiguous masked spans to enforce local structure), Temporal Continuity (predict a masked mid-horizon continuation from an observed prefix to enforce smoothness and causal consistency), and Unsupervised In-Context Dynamics (support-query reconstruction conditioned on subject exemplar windows via attention). Empirically comparing objective families (grouping/contrastive, dynamics-based, and generative reconstruction), we find that dynamics-centric mixtures produce the most balanced transfer: grouping objectives favor discriminative margins but can degrade magnitude fidelity needed for continuous targets, whereas reconstruction-only objectives preserve waveform structure but may underperform on classification. Overall, combining local reconstruction with temporal continuity, and adding in-context conditioning when exemplar access is realistic, yields robust subject-generalizing representations.

Multi-domain graph pre-training is a crucial step in constructing foundational graph models with cross-domain generalization capabilities. However, existing methods predominantly rely on jointly training all source domain graphs, resulting in high computational costs. Furthermore, it remains unclear whether all source domain graph data contribute equally to effective transfer. This paper empirically reveals significant data redundancy in multi-domain graph pre-training. Based on this finding, we propose the Multi-domain Graph Pre-training Framework, MDGMIX, which combines boundary-aware subgraph mixing with hierarchical discrimination. By selecting boundary nodes to construct challenging mixed-domain subgraphs, MDGMIX employs coarse-grained domain discrimination and fine-grained domain decomposition losses to decouple shared patterns from domain-specific patterns. During adaptation, MDGMIX employs a lightweight prompt weighting mechanism to transfer source domain knowledge. Extensive experiments demonstrate that MDGMIX consistently outperforms strong baselines in few-shot classification tasks while exhibiting superior time and memory efficiency. The code is available at: https://github.com/zhengziyu77/MDGMIX.

Cardiovascular diseases (CVDs) remain a leading cause of death globally, necessitating continuous, accurate non-invasive cardiac monitoring. While non-contact radar-based approaches show great promise, they often employ a single "distortion-driven" or "perception-driven" paradigm, frequently facing a trade-off between "low distortion but weak semantic information" and "high perceptual fidelity but poor interpretability." To address this, we propose a Three-stage Distortion-Perception Pre-Training Model (TriDP-PTM), a radar-based multi-scale fusion dual-path framework that systematically compares the "direct radar-to-task" path against an "indirect radar-to-ECG-to-task" path. By integrating an ECG generator with a feature discriminator to form a composite loss function, our approach effectively incorporates medical priors - such as ECG morphology and rhythm - into downstream tasks. Through empirical analysis, we reveal that this trade-off manifests in three distinct phases (Positive-Sum, Coopetitive, and Negative-Sum), showing optimal downstream clinical accuracy typically emerges in the coopetitive stage. Extensive experiments on a dataset involving 30 subjects across 5 physiological states reveal that the indirect path consistently outperforms the direct path in diverse tasks, achieving 0.80 mean IoU in waveform segmentation, 98.3% average classification accuracy across four tasks, and a 56% MAE reduction in blood pressure regression compared to the strongest baselines. These findings validate our framework and indicate that, within the indirect radar-to-ECG pathway, appropriately weighting distortion and perception losses to operate in the coopetitive regime is critical for achieving both clinically interpretable ECG morphology and strong downstream accuracy in non-contact cardiac monitoring.

In contemporary large language models (LLMs), the swish-gated linear unit (SwiGLU) activation function is widely adopted to regulate the information flow and introduce non-linearity. For large positive inputs, SwiGLU approximates the quadratic function $x^2$, providing strong nonlinearity and expressive capacity. However, this property also causes numerical instability as the input or model scale increases, particularly in low-precision LLM training. The main reason is its approximate quadratic amplification, which enlarges the output range and exacerbates outliers. To address this issue, we propose a stable activation function, Power Linear Unit (PowLU), for large-scale LLM pre-training. Specifically, PowLU employs a rational power function to achieve adaptive nonlinearity, thereby improving representation ability and enabling stable training in spike regions. Moreover, we provide theoretical justification for several key properties of PowLU. Scaling law experiments confirm that the performance is consistent across model sizes, and further experimental results with the Ling architecture (7.9B and 124B total parameters) demonstrate that PowLU achieves competitive results against SwiGLU and SwiGLU-Clip in large-scale training of LLMs. In addition, the experimental results also show that PowLU effectively improves the scalability of the large-scale training of LLMs.

Deep learning models are vulnerable to adversarial perturbations, raising important concerns for safety-critical deployment. Empirical defenses can achieve strong robustness in practice, but lack formal guarantees, motivating the need for certifiably robust classifiers. While certified methods provide formal guarantees, they often yield overly conservative bounds due to their inability to exploit structure in complex data distributions. In this work, we propose a framework for designing certifiably robust classifiers that leverages latent structure in data representations. We first analyze the Gaussian mixture setting, deriving necessary and sufficient conditions for the existence of robust classifiers and constructing a classifier with a closed-form robustness certificate and generalization guarantees. Our main contribution is to show that exact structure is not required: we prove that if a pretrained encoder maps inputs to a latent distribution that is $\varepsilon$-close (in KL divergence) to a Gaussian mixture, then certified accuracy degrades gracefully, with an explicit bound relating robustness under the true and approximate distributions. This result enables the direct use of pretrained models without requiring exact distributional assumptions. Empirically, our method achieves state-of-the-art or competitive certified accuracy on CIFAR-10 and ImageNet, while maintaining strong clean performance and low computational overhead. Overall, our work establishes approximate latent structure as a practical and principled route to certifiable robustness.

Legal NLP benchmarks evaluate models on randomly split data, implicitly assuming that legal language is stationary. We test this assumption by fine-tuning four transformer encoders -- XLM-RoBERTa (base and large) and their legal-domain variants -- on Ukrainian court decisions from three temporal epochs defined by geopolitical disruptions: pre-war (2008-2013), hybrid war (2014-2021), and full-scale invasion (2022-2026). Each model is trained on one epoch and evaluated on all three, producing a 3x3 cross-temporal generalization matrix. Four findings emerge. (1) Forward degradation is severe: models trained on pre-war data lose up to 27.2 percentage points of macro-F1 when applied to full-scale invasion era decisions. (2) The degradation is asymmetric: backward transfer (full-scale to pre-war) is substantially more robust than forward transfer, consistent with the hypothesis that legal language is additive. (3) Legal-domain pretraining (Legal-XLM-R) does not improve absolute performance but reduces forward degradation magnitude and asymmetry. (4) Chronological continual learning eliminates catastrophic forgetting for general XLM-R: pre-war knowledge is fully retained (+1.8 to +6.2 pp) while full-scale performance gains +16.5 to +19.0 pp; reverse-chronological training causes severe forgetting. Cross-jurisdictional pretraining on Swiss Judgment Prediction data improves absolute performance but does not reduce temporal degradation magnitude, confirming that temporal drift is an intrinsic property of legal language evolution. The dataset (428K decisions across three epochs) is publicly available as a LEXTREME contribution.

Automated detection and masking of individual methane plumes from satellite imagery is important for operational emission attribution and quantification. We present a machine learning framework for plume detection from MethaneSAT retrieved column-averaged dry-air mole fractions of methane. We address two core challenges: the scarcity of labeled MethaneSAT data and the need for inference reliability across diverse atmospheric and surface conditions. We first demonstrate that Mask R-CNN with a ResNet-50 backbone outperforms U-Net semantic segmentation on both MethaneAIR (an airborne version of MethaneSAT) and MethaneSAT data, with pixel-level F1 score gains of 10.49 and 5.48 respectively. To address MethaneSAT data scarcity, we evaluate three cross-sensor transfer strategies leveraging MethaneAIR flights and synthetic plumes. Mask R-CNN with ResNet-50 fine-tuned from MethaneAIR pre-trained weights is the most effective strategy, achieving instance-level precision of 0.60 and a near-perfect recall of 0.98 at the baseline operating point. A physics-informed post-processing pipeline converts detections into two operationally distinct modes. The first is a high-sensitivity mode that applies morphological filtering and proximity-based merging for comprehensive emission screening, achieving precision of 0.71 and recall of 0.94. The second is a high-precision mode that additionally applies a distribution-based classifier for confident source attribution, achieving precision of 0.92 and recall of 0.70. Manual review of detections classified as false positives against our wavelet-based ground truth labels reveals that a meaningful fraction of cases correspond to real methane enhancements excluded by conservative labeling criteria, indicating that precision values reported are lower bounds on true detection performance... Our data and code are available at: https://doi.org/10.7910/DVN/FR959H

Most black-box optimization methods require extensive hyperparameter tuning, often limiting their ability to generalize across different optimization domains. Foundation models for black-box optimization that learn optimization principles from a large collection of optimization trajectories offer a promising alternative, with the potential to outperform manually designed methods across diverse problem classes. However, prior work has either relied on non-public datasets or on purely synthetic data, limiting reproducibility and generalization to real-world problems. As a result, progress in this area has been constrained by the lack of large-scale, real-world, publicly available pre-training data. We introduce BBO-Pile, the first open-source dataset comprising over 500K optimization trajectories evaluated across 3095 different black-boxes for different optimizers, which represents by far the largest public dataset for this task. Using this dataset, we train a family of foundation models at multiple scales, ranging from 2M to 80M parameters and from 200M to 2B training tokens, and study their scaling behavior with respect to compute. Our results demonstrate that large-scale pre-training is a viable and effective approach to imitate black-box optimization methods, paving the way for future research in this direction.

Nucleic acids are increasingly recognized as therapeutic targets beyond conventional protein-centered drug discovery, yet accurate and efficient docking of small molecules to nucleic acid structures remains challenging. Physics-based docking methods often show limited accuracy and efficiency, whereas deep learning approaches are constrained by the scarcity of experimentally resolved nucleic acid-ligand complexes. Here, we present NucleoDock, a deep learning framework for nucleic acid-small molecule docking. To address data scarcity, NucleoDock combines physics-guided large-scale pretraining on millions of docking-generated synthetic complexes with fine-tuning on curated experimental co-crystal structures. It further integrates sequence- and structure-informed nucleotide representations with atomistic three-dimensional features to capture both biological context and binding-site geometry. A mixture density network-based geometric scoring head is used to model conditional interaction-distance distributions for pose ranking. On an external benchmark of 125 nucleic acid-ligand complexes, NucleoDock achieved a top-1 success rate of 56 percent at an RMSD cutoff of 2.0 Angstrom, outperforming rDock with 29 percent, while generating 100 poses in approximately 5 seconds per complex. Retrospective virtual screening on the ROBIN benchmark further showed improved early enrichment. NucleoDock represents a step toward bridging the methodological gap between protein- and nucleic acid-directed computational drug discovery.

Relational databases (RDBs) remain the cornerstone of modern data systems and support diverse predictive tasks. Recent relational deep learning (RDL) methods enable end-to-end prediction by converting RDBs into graphs, where rows are represented as nodes and inter-table interactions are represented as edges, and then applying graph-based models for representation learning. Despite the strong capability of RDL, effective self-supervised pre-training for RDBs remains non-trivial. RDB tasks often require multi-faceted information across different perspectives and granularities. For example, user churn classification may rely more on interaction patterns, whereas consumption value prediction requires both user-item behaviors and intrinsic user attributes for fine-grained regression. Such heterogeneous needs challenge RDB representation learning, as pre-training objectives should cover comprehensive information for downstream adaptation. However, existing SSL methods typically derive supervision from a single facet, such as node-level intrinsic attributes or subgraph-level relational structures, providing limited adaptability. To this end, we propose RelPrism, a multi-faceted self-supervised learning framework for RDBs. RelPrism constructs intrinsic, relational, and hybrid attributes from distinct perspectives, and applies multi-granularity clustering to each perspective to form corresponding pseudo-task pools. Pre-training over these pools exposes representations to broader perspectives and granularity levels, yielding a stronger basis for downstream adaptation. Experiments on 14 tasks across 5 real-world datasets show that RelPrism improves ROC-AUC by 4.15% for classification and reduces MAE by 10.75% for regression over state-of-the-art baselines. Our code is available at https://anonymous.4open.science/r/RelPrism.

Large language models (LLMs) are typically trained on shuffled corpora, yielding models whose knowledge is frozen at train time and whose temporal grounding remains poorly understood. In this work, we study the impact of pre-training dynamics on the acquisition of time-sensitive factual knowledge, focusing specifically on data ordering. Our main contributions are twofold. First, we introduce a comprehensive benchmark of over 7,000 temporally grounded questions and an evaluation protocol that enables analysis of whether models correctly associate facts with their corresponding time periods. Second, we pretrain 6B-parameter models on temporally ordered Common Crawl snapshots and compare them against standard shuffled pre-training. Our results show that sequentially trained models match shuffled baselines on general language understanding and common knowledge while consistently exhibiting more up-to-date and temporally precise knowledge. Temporally ordered pre-training yields improved factual freshness, while shuffled pre-training peaks on older data, possibly due to increased factual repetition. These findings, along with the release of our code at https://github.com/kyutai-labs/kairos , checkpoints, and datasets at https://huggingface.co/collections/kyutai/kairos provide a foundation for future research on continual learning for LLMs.

Surface electromyography (sEMG) signal-based activity recognition has attracted increasing research attention in recent years. To develop accurate sEMG signal-based activity recognizers, numerous approaches have been proposed. Some studies focus on designing larger and more expressive model architectures to enhance the representational capacity of sEMG signals, while others aim to enrich model priors through large-scale pretraining, thereby improving recognition performance. Recently, large language models (LLMs) have shown remarkable generalization and reasoning capabilities in natural language processing, whose implicit knowledge, learned from extensive linguistic descriptions of actions, opens new possibilities for interpreting sEMG signals and inferring activity intentions. Motivated by this, we propose LLM-sEMG, a novel framework that leverages LLMs as sEMG activity recognizers. Within this framework, we design a language-oriented mapping mechanism that converts continuous sEMG sequences into sEMG language, integrating several strategies to further facilitate the signal-to-language mapping process. Extensive experiments demonstrate that the proposed framework achieves highly accurate sEMG signal-based activity recognition using large language models.

Large Language Models (LLMs) have become the predominant paradigm in NLP, advancing both research and industry. As model sizes and pretraining data grow, concerns about Pretraining Data Exposure (PDE) increase due to the scale and opacity of training datasets. PDE refers to determining whether specific data appeared in an LLM's pretraining corpus. It is critical for ensuring evaluation integrity and protecting privacy, intersecting two key areas: data contamination and membership inference. Though conceptually related, these areas have often been studied in isolation. This paper offers the first unified survey of both under the PDE framework. We formalize PDE across exposure levels, review attack and defense methods, synthesize empirical findings, and highlight open challenges and future research directions.

Computed tomography (CT) is a central to three-dimensional medical imaging, yet CT-based artificial intelligence remains fragmented across task-specific models for segmentation, classification, registration, and report analysis. Here we present FlexiCT, a family of CT foundation models trained by agglomerative continual pretraining on 266,227 CT volumes from 56 publicly available datasets, forming a large-scale public resource for CT representation learning. FlexiCT uses agglomerative pretraining across three stages: two-dimensional axial pretraining, three-dimensional anatomical pretraining and report-guided semantic alignment. This training strategy supports slice-level, volume-level and vision-language analysis. Across five downstream task families (segmentation, classification, registration, vision-language understanding and clinical retrieval), FlexiCT matches or exceeds prior task-specific approaches on multiple benchmarks. Its embeddings further organize CT scans along gradients associated with various tumor stages, suggesting that CT foundation models can capture imaging features relevant to disease phenotype characterization. Project page and code are available at: https://ricklisz.github.io/flexict.github.io and https://github.com/ricklisz/FlexiCT.

Natively trained spiking language models must preserve information across time while operating through sparse binary activations, a combination that has produced a persistent quality gap relative to dense Transformers. We present SymbolicLight V1, a spike-gated dual-path language model that couples binary Leaky Integrate-and-Fire (LIF) dynamics with a continuous residual stream. Its Dual-Path SparseTCAM mixer combines a first-order exponential-decay state with windowed local attention on the continuous residual stream, followed by a context-conditioned decoding head. We train four 194M-parameter models from scratch on a 3B-token, 10-domain Chinese-English corpus. On a fixed token-weighted evaluation set the runs reach perplexity (PPL) 8.88-8.93 (mean 8.904, sample standard deviation 0.019). Separation of this set from the training streams has not been verified. Training-time encoder probes have a mean zero-spike fraction of 89.96%; this is not a whole-model sparsity measure. Code tokens are 43.7% of that set; the unweighted mean of the ten domain PPLs is 29.38. Under the same corpus, tokenizer, token budget, and hardware, the token-weighted mean is 7.7% above GPT-2 201M (PPL 8.27). Across five zero-shot benchmarks the two 200M-scale models show no clear accuracy separation. Under sampling with temperature 0.7 and top-k 50, SymbolicLight produces lower 4-gram repetition; an entropy-modulated rule reverses that ranking. On an RTX 2080 Ti, measured generation throughput is 22.8 versus 91.5 token/s; post-generation power readings give rough energy estimates of 2,848 versus 905 mJ/token, without power integration over generation.

Earth observation (EO) foundation models (FMs) are increasingly trained on multisensor data, spanning multispectral imagery (MSI), synthetic aperture radar (SAR), and derived geospatial layers, but hyperspectral imagery (HSI) remains underrepresented. Conversely, existing hyperspectral FMs are trained on HSI alone, leaving joint pretraining and fusion of HSI with co-located EO sensors unexplored. We introduce SpectralEarth-FM, a hierarchical transformer for multisensor EO input with heterogeneous spectral dimensionality. The architecture combines spectral tokenization for hyperspectral inputs, sensor-specific encoders, a cross-sensor fusion module, and a shared hierarchical encoder, enabling joint processing of HSI and lower-channel observations. To pretrain SpectralEarth-FM, we curate SpectralEarth-MM, a dataset that co-locates HSI from three spaceborne sensors (EnMAP, EMIT, DESIS) with Sentinel-2, Landsat-8/9 optical imagery, Landsat land surface temperature (LST), and Sentinel-1 SAR, over common geographic footprints. It comprises approximately 2M globally distributed locations, 25M georeferenced patches, and over 40TB of data. Pretraining uses a Joint-Embedding Predictive Architecture (JEPA)-style objective that matches representations between global views and single-sensor local views from the same location. We evaluate SpectralEarth-FM on hyperspectral downstream tasks and standard EO benchmarks following the PANGAEA protocol, achieving state-of-the-art results across both evaluation settings.

Preconditioned optimizers are central to language model training, but their stochastic update rules are usually treated as direct approximations to population preconditioned descent. We show that this view misses two finite-sample biases. First, the gradient and preconditioner are typically estimated from the same minibatch, introducing gradient--preconditioner coupling bias. Second, even when the preconditioner estimate is unbiased, its inverse or inverse-root is generally biased because inversion is nonlinear. We propose a single-batch bias-correction framework that addresses both effects: cross-fitted preconditioning estimates the numerator and preconditioner from independent microbatch groups, while variance-corrected inversion uses microbatch variability to subtract the leading delta-method bias term. The framework applies to diagonal moment, diagonal curvature, and matrix preconditioning methods, instantiated in AdamW, Sophia, and Shampoo. Bias correction reduces held-out pretraining loss on Qwen2.5-0.5B by $0.15$, $0.07$, and $0.11$ nats, respectively; the effects on mixed-quality pretraining and downstream instruction tuning are consistently neutral-to-positive. Together, these results establish bias correction as a practical mechanism for reducing finite-sample update bias and improving the performance of preconditioned optimizers.