Causal Foundation Models

Causal inference is the practice of estimating the effect of a treatment or intervention from data. It traditionally requires a bespoke pipeline for every new problem: first proposing a causal mechanism, selecting a compatible estimator, and finally training it. Meanwhile, across diverse settings and modalities, much of machine learning has shifted to the paradigm of foundation models: networks pretrained once at scale and applied to new tasks without fine-tuning. Causal foundation models (CFMs) bring this paradigm to causal inference. CFMs are pretrained neural networks that estimate causal quantities, such as the average treatment effect, on entirely new datasets using in-context learning without requiring model updates. This work provides a practical introduction to this emerging area. We summarize the necessary background in causal inference and machine learning before discussing CFMs. Throughout, we include example code and Jupyter notebooks.

Graph Machine: Towards Better Pretraining via Edges

We introduce the Graph Machine (GM), an architecture that maintains an $O(n)$-sized state and accesses it through sparse, dynamic routing. Unlike methods with fixed-size states or sparse but static routing, GM preserves $O(n)$ complexity in its sparse layers without restricting the potentially accessible state size to $O(1)$. Instead, GM uses edges - pointer-like objects updated differentiably by a referral mechanism resembling pointer chasing. We replace 75% of the dense Transformer layers in Qwen3-0.6B with GM sparse layers and pretrain from scratch on 15.7B tokens. With only 2 of 4,096 tokens retrieved per KV head in each sparse layer, loss degrades only slightly; with 4, the best model marginally improves loss.

Instella-MoE Technical Report

In this work, we introduce Instella-MoE, a fully open Mixture-of-Experts (MoE) language model with 16 billion total parameters and 2.8 billion active parameters per token, trained entirely from scratch on AMD Instinct MI300X and MI325X GPUs. Instella-MoE combines a sparsely activated MoE design with architectural and system-level innovations, including Gated Multi-head Latent Attention (Gated MLA) and FarSkip-Collective connectivity, enabling efficient large-scale training and inference. The model is developed through a multi-stage pipeline comprising pre-training, mid-training, long-context extension, supervised fine-tuning with feedback-driven data curation, direct preference optimization, and reinforcement learning with Multi-Teacher On-Policy Distillation. Instella-MoE achieves an average score of 76.7 across standard pre-training benchmarks, outperforming prior fully open models including OLMo-3-7B, SmolLM3-3B, and OLMoE-1B-7B, while remaining competitive with open-weight MoE and dense baselines at comparable active-parameter scales, including Moonlight-16B-A3B and Qwen3.5-4B. After post-training, our final Think checkpoint achieves an average score of 73.2 across instruction-following, reasoning, math, coding, and chat benchmarks, outperforming both fully open and open-weight models with comparable or larger active parameter counts in our evaluation. To support transparent and reproducible research, we release the complete Instella-MoE model flow, including model weights, training configurations, data mixtures, and training code. Together, these contributions establish Instella-MoE a strong, fully open foundation for efficient, high-performing MoE models and reproducible research.

TUTTI: Toward generalizable audio-to-score transcription via fully synthesized data

Generalizable Audio-to-Score (A2S) transcription is fundamentally constrained by the severe scarcity of high-quality, real-world paired data. Relying solely on existing human-annotated datasets often restricts the generalization of A2S models, limiting their efficacy primarily to single-instrumentation domains. To break this dependency on scarce real-world data, we introduce TUTTI (Transformer for Unified audio-To-score Transcription trained on Synthetic multi-Instrumentation Data), a pre-training paradigm driven by a purely synthetic, large-scale dataset. Rather than using human-composed scores, we leverage a symbolic music generation model to generate a massive, highly scalable multi-instrumentation corpus and create audio-score pairs with expressive acoustic characteristics. Capitalizing on the generated data, we employ a standard Transformer encoder-decoder architecture. We empirically demonstrate that pre-training a unified attention-based model on generated, multi-instrumentation data yields a consistently stronger foundational representation than single-instrumentation training. When fine-tuned with downstream real-world datasets, TUTTI outperforms previous approaches, establishing new overall state-of-the-art results across various A2S baselines. Notably, TUTTI shows remarkable cross-instrument transferability, effectively adapting to unseen instruments with highly competitive performance. The source code and the TuttiCorpus dataset will be made publicly available at https://github.com/a-musiclover/TUTTI.

EEG-VID: Task-Guided Latent Predictive Pretraining for EEG Decoding and Assistive Target Selection

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

Pretrained, Curriculum-Tuned, and Ensembled: A Tracer-Aware Interactive Segmentation Pipeline for AutoPET V

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

A Composition-Aware Pretraining Framework for Geospatial Foundation Models

Geospatial foundation models have emerged as state-of-the-art methods for downstream Earth observation tasks. However, existing pretraining methodologies process imagery through a single-concept lens, failing to capture the highly compositional nature of complex satellite scenes. We propose a composition-aware pretraining framework that explicitly encodes fractional land-cover mixtures. Each satellite image cell is mapped to a histogram representing its fractional land-cover distribution, which we term the "composition target". These targets serve as the primary prediction objective and are distilled into the backbone using Earth Mover's Distance. Experimental evaluation shows that composition-aware pretraining yields substantial gains on region-level understanding tasks requiring semantic similarity judgment, including zero-shot image retrieval and scene classification, while remaining competitive on tasks requiring fine-grained spatial precision, such as segmentation and object detection. With a 36.8M-parameter backbone, our framework outperforms SatMAE and Prithvi-EO-2.0, which contain 303M and 600M parameters, respectively, in most retrieval and scene classification settings. On the fine-grained ForestNet-12 dataset, a rigorous testbed for compositional discrimination, our method boosts baseline mAP@10 from 0.279 to 0.434, a 55.6% relative improvement, providing direct evidence for the effectiveness of explicit composition modeling. The code implementation can be found at https://github.com/05kashyap/GFM_Composition_Pretraining

Beat-Synchronous Tokenization for ECG Transformers

Transformer-based electrocardiogram (ECG) models commonly tokenize waveforms into fixed temporal patches. Though convenient, fixed patching can split heartbeat structures across token boundaries. We study beat-synchronous tokenization as a physiologically grounded alternative, comparing fixed patches with three beat-aligned strategies: resampled beats, adaptive pooled beats, and resampled beats augmented with R--R interval information. Experiments span two settings: 10-second 12-lead diagnostic classification on PTB-XL after MIMIC-IV-ECG masked pretraining, and 60-second single-lead rhythm classification on Icentia11k after patient-level contrastive pretraining. On PTB-XL, resampled beat tokens achieve the highest mean macro Area Under the ROC Curve (AUROC; 0.8945) and nearly match the best fixed-patch macro Area Under the Precision-Recall Curve (AUPRC; 0.7414), reducing average sequence length from 100 to 11.2 tokens. On Icentia11k, beat-synchronous tokenizers obtain comparable AUPRC to fixed patching with better stability across runs. These results suggest morphology-preserving beat tokenization is a compact, competitive alternative to fixed temporal patching.

Deriving Scaling Laws for OpenEuroLLM Models: Learning Rate, Batch Size and Loss

We study the scaling behavior of learning rate and batch size in pretraining dense large language models on English-prevalent corpora. Beyond scaling jointly optimal learning rates and batch sizes, we investigate their marginal evolution with model capacity and data scale and develop a model that captures these relationships. As we employ a Warmup-Stable-Decay learning rate schedule, we further investigate the gains from learning rate annealing over a broad range of hyperparameters settings, models and data budgets, and whether the optimal learning rate and batch size transfer between the stable and decay phases. Finally, we characterize the dependence of loss on model capacity and dataset size, evaluating recently proposed scaling forms that explicitly model their interaction. We find these approaches particularly effective at capturing both undertraining and overtraining regimes across our experiments. This study establishes a first baseline and scaling procedure for the development of future OpenEuroLLM models. We open-source the complete collection of pretraining runs used in this study.

GAAT: Geometry-Aware Alignment Transformer for Multimodal UAV Perception

Unmanned aerial vehicle (UAV) multimodal perception integrates visible (RGB), infrared (IR), synthetic aperture radar (SAR), and depth sensors for scene understanding under diverse conditions. However, differences in optics, resolution, and mounting often limit practical systems to global or image-center alignment. After tokenization, parallax, platform motion, and lens distortion can shift corresponding patch centers across modalities, weakening the spatial correspondence assumed by dense contrastive learning and cross-modal fusion. We propose GAAT (Geometry-Aware Alignment Transformer), an alignment-first pretrained model that estimates local correspondence reliability before cross-modal interaction. GAAT introduces syncPATC, which learns patch-center consistency under synchronized view transformations without correspondence annotations. It emits geometric priors, including token and query confidence, query centers, and sub-token offsets, that identify reliable local anchors across residual misalignment. Guided by these priors, MG-Sparse-MMA performs query-mediated sparse fusion over top-K_s reliable regions, replacing dense all-patch interaction with geometry-calibrated local updates. RA-QCGCL aligns pretraining supervision with this sparse query bottleneck through reliable patch-to-patch, patch-to-query, and query-to-query contrastive branches. We introduce UAVMeta and StateBench, which provide four acquisition-state scores derived from platform telemetry and image statistics: camera reliability, observation scale, viewpoint stability, and flight maneuver complexity. Extensive experiments across six downstream tasks demonstrate consistently superior transfer performance, establishing GAAT as a state-of-the-art multimodal foundation model for UAV perception. StateBench further enables a systematic diagnosis of real-world acquisition conditions.

Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090

Language model pretraining has become almost synonymous with prohibitive cost, placing it out of reach for much of the academic and open-source communities. Although strong open-source efforts already exist, including open-weight models and open-source training recipes, a cost-efficient, hardware-accessible, and open-source pretraining recipe has long been missing. Even at a small scale, training Llama-3.2-3B costs over \$1.5M, and reproducing SmolLM3-3B needs over \$700K. In this report, we present an open pretraining recipe designed to lower this barrier. Using this recipe, we train a collection of Puro-2B models from scratch on up to 1.4 trillion tokens with FP8 precision on consumer-grade RTX 5090 GPUs. The models in the collection differ in token budgets and selected recipe variants. Our best model is trained at a compute cost of less than \$6.9K and approaches Qwen2.5-1.5B performance under our evaluation protocol. This cost efficiency is enabled by a combination of approaches, including hardware selection, low-precision training, hyperball optimization, curriculum model averaging, and the data recipe. Beyond the recipe itself, we provide two additional results. First, across the Puro-2B collection, we derive a Puro Cost Scaling Law that relates training cost to average model performance; the fitted law suggests that about \$4.4K, less than \$5,090, is sufficient to reach the performance of Qwen2-1.5B. Second, as an end-to-end case study, we examine how pretraining data curricula shape downstream performance after post-training. Such controlled studies are enabled by having access to the full pretraining pipeline rather than model weights alone. We release the full training recipe for Puro-2B, including data, code, and model weights under Apache 2.0 at https://huggingface.co/collections/thu-pacman/puro-2b.

Robust Neural Stimulation Response Modeling Through Meta-Learning and Pretraining

Objective: Model-based closed-loop neural stimulation holds promise for therapeutic applications ranging from Parkinson's disease to sensory restoration, but deployment has been limited by two obstacles: 1) forecasting models for predicting the consequences of stimulation fail catastrophically on a meaningful fraction of sessions, and 2) per-session calibration requirements are often incompatible with clinical constraints. We address both by demonstrating, for the first time, that meta-learning and pretraining can be applied to neural stimulation response modeling. Methods: Temporal basis function models (TBFMs) forecast state-dependent neural responses to stimulation. We extend TBFMs with cross-session pretraining using a novel architecture and algorithm based on model-agnostic meta-learning (MAML), evaluating them on 40 sessions of optogenetic stimulation in primary sensorimotor cortex of two non-human primates. Results: Meta-learning substantially reduces catastrophic forecast failure: for a 1k calibration set size, sessions with test R-squared < 0.05 drop from 16 of 40 (single-session training) to 1 (MAML-pretrained), and prediction intervals become significantly narrower (p < 0.05). Calibration requirements are reduced by 50-90% at matched accuracy, enabling experiments otherwise infeasible within clinical session-time constraints. Conclusion: Our results demonstrate that cross-session structure in stimulation responses is consistent enough to support pretraining, providing the first empirical evidence that meta-learning approaches are viable for neural stimulation. Significance: The robustness and sample efficiency gains directly address known obstacles to deploying model-based stimulation controllers. Our results motivate community efforts to assemble standardized multi-site stimulation datasets and to further explore meta-learning for robust closed-loop stimulation.

MoganBert-TR: A Turkish Encoder Foundation Model Trained from Scratch with a CLM-to-MLM Curriculum

Turkish encoder models have adopted modern architectures while leaving the pretraining objective fixed at masked language modelling. This paper introduces MoganBert-TR, a 149M-parameter Turkish encoder foundation model trained from scratch on a language-specifically filtered corpus, together with an embedding model derived from it (MoganBert-Embed). MoganBert-TR is trained over 237.3B tokens with a two-stage CLM-to-MLM curriculum: causal language modelling first, masked language modelling for the remainder, with the transition made inside the stable phase of a WSD schedule. In a controlled ablation under an equal step budget, this design outperforms pure MLM by 2.7-3.7x on Turkish MS MARCO retrieval; the measured mechanism is embedding geometry, where a single direction absorbs 28.1% of the variance under pure MLM against 11.9% under the curriculum. Long-context extension and learning-rate decay are then split into two branches after a shared prefix: running the final portion of decay at 1024 context improves the TrGLUE average by 0.49 +/- 0.26 points across five paired seeds (p = 0.013) and beats a model-soup alternative by 0.75 points at ~4.3% additional cost. MoganBert-TR attains 78.41 on TrGLUE, the best among the Turkish ModernBERT models compared, and 77.73 on TabiBench, where it leads two of the eight categories with the largest margin on code retrieval (+3.62 points over TabiBERT). MoganBert-Embed, produced through teacher distillation and multi-signal contrastive fine-tuning, ranks first among student models on the MTEB(Turkish) overall average with 68.30 and reaches 99.5% of its 7.57B-parameter teacher's score with a 51x smaller backbone. The accompanying 50,048-token tokenizer outperforms all compared Turkish tokenizers on compression and fertility across two independent test sets. Weights, tokenizer, embedding model and evaluation code: https://huggingface.co/moganai

Effective Learning Rate Governs Loss Dynamics in Language Model Pretraining

We uncover ELR collapse in language model pretraining: learning rate (LR) and parameter norm govern loss dynamics primarily through their ratio, the effective learning rate (ELR). When ELR is matched across runs, their loss trajectories collapse throughout training despite substantially different LRs and parameter norms. Across optimizers, architectures, datasets, and model scales, mean collapse errors are typically a few x 10^-3, below the seed-to-seed variation measured in a representative configuration. Systematic ablations identify normalization design and the timescale of LR-norm variation as key determinants of collapse precision. Controlled interventions further show that weight decay and Hyperball shape loss dynamics primarily through the ELR schedules they induce. Replacing LR with ELR enables a fitted functional scaling law (FSL) to transfer across norm-control methods. The resulting ELR-based FSL also explains delayed acceleration, a recurring effect of norm control. Together, these results establish ELR as a common coordinate linking LR scheduling, norm control, and loss dynamics.

B-MIM: Biased Masked Image Modeling for Generalizable Segmentation of Fine-Grained Anatomical Structures

Self-supervised pretraining enables transferable representations for medical imaging, yet most CT encoders remain biased toward coarse semantic understanding, limiting their sensitivity to fine-grained anatomical structures such as vessels or small tumors. In this paper, we introduce Biased Masked Image Modeling (B-MIM), a modification of the iBOT objective that stochastically reduces global semantic alignment to prioritize local patch reconstruction. This bias encourages the encoder to capture high-frequency morphological details and structural continuity. We curate a multi-institutional CT abdominal dataset of 9,955 filtered studies from 17 public sources and pretrain a 3D Swin Transformer backbone using B-MIM. Across inter-dataset experiments on liver vessel segmentation, the proposed encoder improves topological fidelity (clDice) and achieves competitive Dice scores in tumor segmentation, compared to fully fine-tuned baselines, despite updating only a fraction of the parameters. Our results suggest that reducing global semantic pressure during pretraining enhances generalization to intricate anatomical structures.

Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records

Predictive models over structured electronic health records (EHRs) remain central to machine learning for healthcare, but few have jointly emphasized quantitative laboratory information and interpretability with respect to input medical events. We present BERT-LER, a BERT-style model for coded EHR timelines pretrained and fine-tuned from a de-identified EHR dataset of 75 million patients, that encodes laboratory test results as discrete tokens while retaining graded information through percentile-based binning, paired with Integrated Gradients for token-level attributions grounded in the input EHR sequence. We benchmark our approach on the public EHRShot benchmark suite and on an asthma severity progression study based on real-world data. This addresses a methodological gap in EHR foundation-style modeling by unifying laboratory value representation and explainability in a single framework, while assessing whether both predictive performance and explanations generalize beyond standard clinical prediction tasks. Across EHRShot and asthma tasks, BERT-LER achieves predictive performance that is competitive with, and on laboratory-related tasks often exceeds, publicly available benchmark models, and provides attributions that align with clinically known risk factors. Our architecture and explainability approach can be applied to many therapeutic areas and prediction tasks using language models trained on structured EHRs.

Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts

Mixture-of-Experts (MoE) architectures significantly expand model capacity without a proportional increase in computational cost. However, optimizing their hyperparameters---particularly the learning rate---at extreme scales of both model size and token budget via sweeping remains computationally prohibitive. In this paper, we propose a compute-efficient, two-step hyperparameter transfer framework that estimates optimal learning rates for training large MoE models by transferring them across scaling model widths, and subsequently extrapolating to trillion-token horizons. First, we formulate a Maximal Update Parameterization ($μ$P) adaptation for MoE architectures utilizing Multi-head Latent Attention (MLA) and the Muon optimizer, demonstrating that optimal learning rates transfer consistently across width-scaled models. Second, we extend this transferability along the token dimension by establishing a predictive scaling law. By applying linear regression to the optimal values derived from small proxy models on limited budgets, we successfully extrapolate the ideal learning rate to massive training horizons (e.g., 10 trillion tokens) with high fidelity ($R^2=0.95$). Consequently, this indicates that proxy training on small models is sufficient to determine the optimal learning rate for the extensive training of large-scale MoEs. We apply the proposed methodology to pretrain our foundation model (155B total, 17B active parameters) from scratch, and the stable training and evaluation results validate that optimal configurations for full-scale target models can be accurately predicted with minimal ablation costs.

Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking

Pretraining time series foundation models across heterogeneous datasets necessitates effective handling of varying sampling frequencies. Current methods either employ dataset-specific patch sizes and separate FFNs, leading to fragmented representations, or enforce a fixed patch size that neglects inherent temporal variations. To address this, we propose SATS, featuring a scale-aware token alignment mechanism that treats patch size as an explicit notion of scale. By incorporating a contrastive-inspired alignment regularizer, SATS aligns representation spaces across scales while preserving distinct modeling capacities. Furthermore, a hybrid masking strategy combining random and contiguous masking is introduced to capture multi-scale temporal structures. Experimental results on LSTF benchmarks demonstrate that SATS achieves a 9.2% improvement in MSE and an 8.3% gain in GIFT-Eval MASE compared to competitive baselines. Notably, SATS consistently delivers SOTA performance while achieving a 65.6% increase in model efficiency over advanced baselines, highlighting its effectiveness and scalability in time series pretraining.

Activation-Keyed Momentum: An Anisotropic Momentum Update via the Delta Rule

Most modern optimizers form their momentum as an exponential moving average (EMA) of past gradients, forgetting every direction at one fixed rate. However, the inputs a deep network sees during training can be highly anisotropic, with a few directions queried frequently while most are seen rarely. Preconditioning methods address this anisotropy by wrapping extra processing around this buffer and leave the momentum update itself unchanged. We propose Activation-Keyed Momentum (AK-Momentum), which builds direction-awareness into the momentum update rule. The gradient of a linear layer splits into an input activation that acts as a key and an output-side error that acts as a value. Keying on that activation, AK-Momentum updates the momentum buffer by the canonical delta rule, so each direction is forgotten at a rate set by how often it appears. We prove that it is a valid momentum, that it applies the input-side curvature correction without matrix inversion, and that it clears stale directions faster than EMA under both a fixed and a drifting optimum. It is a drop-in replacement for the momentum buffer of any optimizer, its coefficient transfers across widths under $μ$P, and its extra compute stays between $22.2\%$ and $25.0\%$ of a gated-MLP block's linear cost with no persistent memory. In FineWeb-Edu pretraining, AdamW with AK-Momentum (AK-AdamW) reaches AdamW's validation loss in up to $46.39 \pm 4.32\%$ fewer steps at 67M and $22.12 \pm 0.80\%$ at 370M over three seeds, and the gain persists at 1B on a Chinchilla-optimal budget. A Muon baseline tuned under the same protocol sits above AK-AdamW at both language-model scales, and the gain holds for SGD, ResNet-18, and ViT-Tiny on CIFAR-10. Training-time diagnostics confirm the predicted mechanism, better gradient tracking and healthier input directions.

Pretraining Reusable Inference Across Views with Synthetic Task Priors

Modern pretrained encoders make representations from heterogeneous views increasingly reusable, but the procedure that determines view utility and combines evidence is still relearned for each downstream task. Consequently, knowledge about view relevance, complementarity, reliability, and missingness is repeatedly discarded rather than transferred across tasks. We therefore reformulate multi-view learning as learning a reusable, task-conditioned inference procedure rather than a fixed fusion function. Based on this perspective, we propose SIMPLE, a prior-fitted multi-view in-context learner that predicts query labels by conditioning on a small labeled support set. Since existing real-world datasets cover only a limited range of view configurations and task structures, we construct a controllable synthetic task prior in embedding space. It generates diverse support-query episodes with varying class structures, shared and view-specific factors, representation geometries, cross-view dependencies, reliability levels, missingness patterns, and distribution shifts. A hierarchical inference architecture then performs reasoning within views, across views, and across support and query samples. Experiments on multi-view and multi-omics benchmarks demonstrate that the frozen variant of SIMPLE achieves competitive performance without updating the inference backbone, while lightweight adapter calibration attains leading performance on most evaluated datasets. Together, the results under frozen, one-shot, and missing-view settings support the central hypothesis that multi-view reasoning itself can be pretrained and reused, while lightweight adapter calibration provides task-specific alignment when needed.

WhiteMatter: All-to-All Cross-Layer Connections via KV Mixing

In a Transformer, each layer attends to past tokens only through KV produced at its own depth, despite the presence of deeper representations during autoregressive decoding. Feedback architectures allow shallow consumer layers to attend to KV produced by deeper past-token representations, but give all consumer layers the same fixed connection patterns to source layers. We propose WhiteMatter, which connects every attention layer to the representations from all layers of each past token, with connection weights that can vary across consumer layers and adapt to the source token. For each token, a router implements these connections by mixing its $L$ layer states into $k$ KV channels that are cached for subsequent tokens; each consumer layer attends to one of the channels. The number of channels $k$ controls the KV-cache size. Setting $k<L$ reduces the cache's memory footprint. In our pretraining experiments, WhiteMatter outperforms a vanilla Transformer with 50% more layers and retains most of this gain with a 50% KV-cache compression.

Beyond Natural-Image Foundation Models: Benchmarking Satellite Pretraining for Ophthalmic Image Analysis

Vision Foundation Models (VFMs) have emerged as a promising approach in medical imaging, producing broadly applicable systems that can be efficiently adapted across diverse imaging modalities, anatomical regions, and clinical tasks. However, VFMs require extensive training data, and their progress in medical image analysis is constrained by limited data availability, privacy concerns, and high development costs. To alleviate these constraints, medical VFMs (MedVFMs) are often built upon weights from generalist models pretrained on vast amounts of publicly available natural images, introducing a substantial distribution shift for medical task adaptation. To address this, we propose satellite imagery as a novel pretraining domain for MedVFM development and benchmarking, motivated by its closer visual alignment with medical data and its freedom from the privacy constraints that limit medical datasets. Across multiple ophthalmic imaging modalities, we compare DINOv3-SAT493m pretrained on 493 million satellite images against DINOv3-LVD1689m pretrained on 1.7 billion natural images, together with two medical specialist baselines: DINOv3-RETFound and MAE-RETFound. Our experiments show that satellite imagery is a stronger pretraining source than natural images for ophthalmic tasks, particularly on en face vascular-rich modalities. On several tasks, satellite pretraining matches or exceeds the medical specialists on high-resolution en face inputs, despite using no medical data.

Scaling Domain Data Repetition in LLM Pretraining

As large language models scale, their training-token budgets must also increase to maintain an appropriate tokens-per-parameter ratio (\(\mathrm{TPP}\)). However, high-quality domain data is much harder to scale than general web data. As model size and the training-token budget increase, its fraction in the training mixture tends to decrease. Repeating the available high-quality data provides an effective way to counteract this dilution, but excessive repetition may lead to overfitting. We study this trade-off under practical LLM scaling, where the training-token budget grows proportionally with model size. For a fixed domain, we first find that, surprisingly at a fixed \(\mathrm{TPP}\), the optimal repetition count mildly increases with model size. Across different domains, we find that the optimal repetition count is strongly negatively correlated with the final validation loss of a domain: domains with lower loss can generally benefit from more repetitions. In contrast, the amount of unique domain data is only weakly related to the optimal repetition count. These findings suggest that repetition counts tuned on smaller proxy models with the same \(\mathrm{TPP}\) can provide a practical estimate for larger models.

Measuring Task-Agnostic Training Data Influence Across Language Model Pretraining

Measuring training data influence consistently across language model pretraining is challenging. It is difficult to select downstream tasks or validation sets representative of a model's general capabilities, and reliance on task performance at intermediate checkpoints complicates comparisons across training. We propose a measure of training data influence that does not require selecting a downstream task or validation set as the attribution target. Specifically, we define an example's influence by how much its gradient update reduces the squared distance to the final parameters of a given pretraining run, and estimate this quantity from intermediate checkpoints without retraining. Applying the method to 18 configurations from the Pythia and PolyPythia suites, we find systematic temporal changes in influential data. Early in training, literature-related data are more strongly aligned with the trajectory toward the final parameters, whereas STEM data become more strongly aligned in later stages. This qualitative crossover is broadly consistent across model configurations. Our results provide a tractable trajectory-level view of how influential data change throughout pretraining, complementing influence analyses defined with respect to specific downstream tasks or validation sets.

Synthetic Persona Pretraining: Alignment from Token Zero

As language-model-based AI is increasingly deployed in autonomous settings, aligning its goals and values with those of humans becomes critical. Today, alignment, and the assistant identity itself, are typically introduced only after pretraining, once behavioral priors are already established. This can make values a thin overlay, rather than deeply rooted, and facilitate subsequent misalignment. Pursuing a different paradigm, we introduce Synthetic Persona Pretraining (SPP), which installs the desired assistant persona from token zero in pretraining. First, we annotate pretraining documents with value-aligned first-person reflections derived from a normative value constitution. Second, we pretrain via the standard cross-entropy loss on standard pretraining documents as well as their reflections, which installs the desired persona among a multitude of other personas. Finally, we post-train on user-assistant dialogue data, which binds this desired persona to the assistant identity, a process we call persona binding. By pretraining models up to 3B parameters on 500B tokens, we show that SPP improves constitution following and jailbreak robustness, and reduces the misalignment rate in out-of-distribution moral dilemmas, while preserving capabilities. Early intervention matters: compared with alignment from token zero, introducing SPP only at the end of pretraining yields weaker constitution adherence, does not shift value priorities, and leads to less aligned choices in dilemmas. This advantage depends on persona binding and, importantly, increases with pretraining budget. Overall, our results show that shaping values early is critical for alignment and establish pretraining-time persona interventions as an effective approach to do so.

Mixture of Training: Recombining Small-Scale Scaffolded Pretraining Runs into a Larger Language Model

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

Is Convergence Inevitable? Tracing Output Homogeneity Back to Base Models

The lack of diversity in LM content is widely attributed to the alignment process, but how and where exactly in the pipeline this collapse begins is unknown. We argue that output homogeneity is likely learned during the pretraining phase, and only revealed or magnified during the alignment process. Specifically, we find that semantic convergence is observed from the first alignment stage--the instruction-tuning phase (SFT)--suggesting that homogeneity might already exist in the pre-alignment model. To investigate this, we conduct controlled SFT experiments examining how training data influences output convergence on specific input/output pairs. We find that convergence can be revealed and amplified, but not introduced by the SFT data, supporting its role as a catalyst rather than a cause. To further test whether homogeneity originates before alignment, we measure convergence in base models. We find that instruct-like collapse can be induced through prompting alone, even without alignment. Taken together, our results suggest that semantic convergence may arise naturally from the objectives underlying LM training, making it difficult to mitigate through post-alignment interventions alone.

DegradeQuery: Counterfactual Tuple Pretraining for Context-Aware PROTAC Degradation Prediction

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

Rethinking Data Efficiency in Industrial Dense Prediction: Pretraining Coherence, Not Inductive Bias, Determines ViTs Low-Data Advantage

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

Cracks in the Foundation: Seemingly Minor Architectural Choices Impact Long Context Extension

One might imagine that architectural variations within the dense transformer paradigm have a limited effect on accuracy. However, we demonstrate that this is not the case in the long context setting. Specifically, we show that a set of four minor architectural decisions --- all made by at least one of the Olmo, Llama, and Qwen dense model families --- have a compoundingly negative effect on long context extensibility. Any one of these choices alone has a minor impact on long context performance, but combining three or more can drop the performance downstream by up to 47%. Furthermore, these differences are not detectable from short-context loss or validation datasets. We show that much of the variation in long context ability across model families is driven by these architectural features and detectable from applying context extension early in pretraining. We demonstrate this with controlled ablations that hold data, tokenizer, and extension recipe fixed while varying normalization, GQA, pretraining context length, and sliding window attention. After over 170,000 GPU hours of training, we release the resulting set of models as OlmPool, a set of 26 comparable 7B models with checkpoints before and after long-context extension. This pool includes several architectures that outperform the Llama 3 architecture on long context extensibility. In an analysis of our ablation models, we identify patterns in attention sink behavior and attention distributions across context that are attributable to specific architectural differences.