Downstream Fine-Tuning

Binary-choice truth benchmarks ask models to choose between a correct and an incorrect answer, but if the two answers differ systematically in surface-level features, models can exceed chance without performing the intended reasoning. We show that this failure mode is detectable and can be exploited by downstream classifiers. In TruthfulQA, a simple six-feature logistic classifier achieves substantial accuracy in separating correct from incorrect answers. We further show that similar surface-level artifacts are present in additional benchmarks. To counteract this, we developed a general mechanism to clean them by removing the most leakage-reinforcing pairs. We release a version of TruthfulQA with surface-feature leakage reduced close to chance and provide a mechanism, Audit-Prune, so that the datasets can be cleaned before release.

Graph condensation aims to produce a small surrogate graph that preserves the downstream node-classification performance of a much larger original graph. Existing methods rely on Weisfeiler-Lehman neighbourhood aggregation or gradient-based distribution matching, both of which assume that adjacent nodes share the same label, an assumption that breaks down under heterophily. We propose HERALD (High-fidelity Exemplar Retrieval with Adaptive Landmark Distillation), a gradient-free graph condensation framework that adapts the node scoring and feature selection in the condensation pipeline to the graph's measured heterophily. HERALD selects features via a joint Fisher-discriminability and activation-density criterion that down-weights aggregated representations on heterophilic graphs, and scores nodes by a weighted combination of prototype representativeness, decision-boundary proximity, and Local Intrinsic Dimensionality (LID), where the weights are driven by a smooth sigmoid function of the heterophily ratio. Nodes are then assembled into a condensed subgraph through score-ordered BFS expansion, Personalised PageRank pruning, and class rebalancing, all at an identical storage budget to BONSAI, enabling direct comparison. Experiments on eight benchmark datasets spanning homophilic and heterophilic settings show that HERALD matches or outperforms state-of-the-art condensers on heterophilic graphs and remains competitive on homophilic ones across four GNN architectures.

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

Language model tokenizers are typically selected with minimal evaluation, despite the fact that their design choices directly impact model capabilities. This can be partly attributed to a limited understanding of which tokenizer properties affect which aspects of downstream performance. We introduce TokEval, a framework of tokenizer evaluation metrics that goes beyond standard measures like fertility and compression rate to capture linguistically and structurally meaningful properties, e.g., UTF-8 character boundary integrity and digit place-value boundary alignment for mathematics. To validate whether these metrics are predictive of downstream model performance, we conduct controlled language model pretraining experiments, varying solely the tokenizers' training data mixture, pretokenization strategy, and training algorithm. We evaluate the resulting models on bits-per-byte (a tokenizer-agnostic version of perplexity) and several benchmarks, spanning linguistic understanding, mathematical reasoning, and code generation. Our experiments suggest that different intrinsic properties have different impacts on model abilities: information-theoretic metrics predict language modeling abilities (Spearman rho up to 0.80), while structure-sensitive metrics, such as those measuring digit and line-break handling, correlate with task accuracy. We hope TokEval enables more principled tokenizer evaluation, replacing pretraining sweeps with intrinsic measurement wherever the two agree.

Trained transformer models develop privileged bases: coordinate axes whose statistics differ from the rest of the residual stream. But what kind of direction does such a basis select? We investigate the prediction direction, the unembedding direction of the token a model currently predicts, and find that it functions as a content-defined privileged anchor. Measured with respect to this anchor, residual-stream variation is geometrically and behaviorally stratified by proximity to the prediction. The stratification holds in all eighteen models tested (dense and mixture-of-experts, 7B-120B, base and instruction-tuned). A narrow, scale-invariant prediction interface concentrates readout-relevant structure, while the vast prediction-distal complement expands with model scale. Because the prediction direction sits nearly orthogonal to the principal variance axes, variance-based analyses recover this organization only partly, and the shortfall grows with prompt heterogeneity. Anchoring reveals a steep geometric gradient: prediction-proximal regions are highly structured and cluster related prompts, while the complement is flatter and anti-discriminates among prompt groups. The interface is a narrow slice but functionally decisive. Disrupting the variance directions closest to the prediction causes immediate divergence and frequent task-frame shifts; disrupting the next level down delays divergence and preserves framing. The complement is weakly readout-aligned per direction yet causally and temporally load-bearing, and behavior is driven by direction rather than magnitude. These results establish the prediction direction as a privileged anchor distinct from previously described coordinate axes, and give a geometric account of how high-dimensional computation coexists with linear readout.

This work presents DS@GT ARC BioASQ team's work for a biomedical question answering pipeline, integrating multi-source query expansion, neural reranking, retrieval refinement, and OpenBioLLM-assisted answer generation. The system combines PubMed retrieval with fine-tuned MiniLM-based semantic reranking, Reciprocal Rank Fusion (RRF), and feature-based relevance scoring to improve document ranking quality. To address challenging queries with weak retrieval performance, we introduce a conditional weak-question recovery strategy that applies semantic expansion, relationship-aware augmentation, and selective result merging. A post-retrieval pruning stage further removes redundant or low-relevance snippets while preserving evidence coverage for downstream answer generation. Experimental results on BioASQ evaluation batches demonstrate that the proposed recovery and cleanup strategies substantially improve retrieval robustness and MAP@10 performance on difficult question sets. The final system also incorporates output validation and post-processing steps to ensure formatting consistency and submission reliability across BioASQ phases.

Synthetic tabular data is increasingly used in privacy-preserving data sharing, data augmentation, and to mitigate downstream classifier bias. State-of-the-art tabular diffusion models such as TabDDPM and TabSyn achieve excellent distributional fidelity but offer no mechanism for fairness; conversely, fairness-aware tabular generators (DECAF, FairTGAN, FairTabDDPM) impose explicit fairness penalties at training time, yielding modest fairness gains at substantial cost to either sample quality or downstream utility. We introduce FairDiffuseVQVAE, a two-stage architecture that decouples fidelity from fairness: a vector-quantized autoencoder with a row-level discriminator (Stage1, no fairness terms) is followed by a DiffuseVAE-style continuous diffusion refiner that conditions on both the Stage-1 reconstruction and the protected attribute via classifier-free guidance (Stage2). Fairness emerges as a property of the sampling distribution -- uniform sampling of the protected attribute at inference time enforces demographic parity by construction, rather than from competing loss terms. On the Adult, Bank and COMPAS datasets, FairDiffuseVQVAE achieves the highest mean Demographic Parity Ratio ($0.702$, $+47\%$ over FairTabDDPM) and Equalized Odds Ratio ($0.686$, $+100\%$). It also attains the lowest mean pair-wise correlation error ($0.034$) of any published method, while explicitly trading \sim$$15 AUC points for these fairness gains.

Gigapixel Whole-Slide Images (WSIs) present a fundamental computational bottleneck for vision-language models (VLMs) due to extreme sequence lengths. Existing approaches predominantly rely on spatial sampling or training-free pruning, which risk diluting weak but informative signals, leading to the loss of critical diagnostic evidence due to the spatially diffuse nature of pathological cues. We reformulate WSI token pruning as a sequential selection process, enabling the model to autonomously learn an optimal routing strategy rather than relying on static heuristics. We herein propose a decoupled routing framework integrated as an active plugin into the fully pre-trained SlideChat base model, leaving both the slide encoder and large language model frozen. To provide continuous gradients for the non-differentiable pruning operation during training, we introduce PathSelect. PathSelect employs a variance-preserving noise gate to modulate each patch's information flow via a differentiable Soft Top-K operator, paired with a diagonal-attention Denoiser that recovers the perturbed representations without semantic leakage. At inference, the PathSelect module is entirely detached. Relying solely on the trained Scorer, a deterministic Hard Top-K operator executes adaptive, data-dependent trajectory termination, significantly accelerating downstream generative processing with exceptionally low sequential token selection latency. Driven by an empirical average of only 44.86 tokens under a maximum constraint of K = 128, our framework achieves 74.00% overall accuracy on SlideBench (TCGA), representing an approximate 36.6x spatial token reduction relative to the uncompressed baseline average while consistently outperforming sampling-based counterparts.

Federated pre-training offers a way to train foundation models on private or distributed data without centralizing the underlying datasets. However, evaluating federated pre-training remains challenging because differences in client participation and local data availability can make directly comparable evaluation difficult. Moreover, pre-training test perplexity is tied to the pre-training distribution, while downstream benchmarks introduce task-specific adaptation that may not faithfully reflect the test perplexity established during pre-training. In this work, we study which evaluation protocol more reliably reflects federated pre-training quality. Using a controlled set of centralized and federated-trained models of a 16M parameter transformer model trained on identical client data, we assess evaluation protocols by whether they preserve a reference ranking established on the same pre-training testset. We compare downstream fine-tuning on GLUE, including full, head-only, and reduced-data variants, with next-token prediction on GLUE text as an intrinsic evaluation signal. Our results show that downstream fine-tuning does not reliably preserve the pre-training ranking, whereas direct next-token prediction exhibits a strong correspondence with the pre-training test perplexity. These findings suggest that downstream fine-tuning alone can be misleading when comparing federated pre-trained models, and that evaluation signals closer to the original pre-training objective deserve greater attention.

Breast ultrasound is widely used for screening, yet automated analysis remains challenging due to speckle noise, acquisition variability, and weak separation of benign and malignant cases in standard ultrasound imaging. Graph convolutional networks (GCNs) have recently emerged as a promising approach by leveraging relationships among similar patient samples. However, it remains unclear how the choice of image encoder influences graph construction and downstream classification performance. In this work, we systematically evaluate five image encoders spanning convolutional and transformer-based architectures for GCN-based breast ultrasound classification. Image embeddings are used to construct cosine similarity k-nearest-neighbor graphs, which are classified using a single-layer GCN with a linear classification head. Across three patientwise cross-validation folds, higher-capacity encoders consistently improve graph homophily and downstream classification performance, yielding gains in accuracy, AUC, sensitivity, specificity, and F1-score. Moreover, test-set graph homophily exhibits a strong linear correlation with classification accuracy, with higher-capacity encoders consistently occupying the high-homophily, high-accuracy region suggesting that encoder-driven improvements in graph structure are a key mechanism underlying the observed performance gains. These findings establish encoder selection as a critical factor in graph-based breast ultrasound classification and identify graph homophily as a key indicator linking representation quality to downstream classification performance.

In this work, we present Canvas360, a two-stage framework for in-context panoramic generation that combines geometry-aware pretraining with downstream task-specific fine-tuning. To address the lack of large-scale, high-quality training data tailored to in-context panoramic tasks, we propose Canvas360Dataset, a collection of 1M high-quality paired panoramic samples for style transfer, inpainting, outpainting, and editing, enabling effective supervision across diverse in-context generation scenarios. On the modeling side, Canvas360 enhances text-to-panorama generation through parallel depth generation, velocity circular padding, and similarity loss regularization, enabling the model to learn geometry-aware representations, capture object distortion details, and improve geometric consistency and global coherence. Furthermore, empowered by strong panoramic priors, Canvas360 enables a unified in-context panoramic generation framework that supports diverse downstream tasks via token-level concatenation, surpassing prior methods in both task coverage and modeling flexibility. Extensive experiments show that Canvas360 improves panoramic image fidelity, achieving particularly strong performance on the panorama-specific FAED metric and competitive or leading results across the reported quantitative evaluations. More information can be found on our project page: https://zry000.github.io/Canvas360/

Mixed-integer linear programming (MILP) instances used for solver development are hard to obtain when models come from private or application-specific pipelines. A generator must keep the structure that solvers and learned policies rely on. Existing general generators usually choose their generation unit from a formulation template, summary statistics, local graph edits, or blocks found after recombination. These units do not explicitly record how a local part of the MILP is coupled to the rest of the instance. We propose GraphBU, a graph-native generator whose basic unit is a local subproblem plus its interface. The method promotes coupling nodes into master constraints or boundary variables and uses the resulting block units for compatibility-checked replacement. The analysis focuses on the properties needed by this construction: promotion separates interfaces, replacement can preserve feasibility under an interface-slack condition, and the graph construction is invariant to row-column permutations. On MILP instances generation, this unit keeps graph statistics close to the source family, preserves feasibility on most datasets, and improves downstream Predict-and-Search training. Genrated by GraphBU, The average graph-statistical similarity was approximately 0.934, the average feasibility was approximately 96.7%, and the average increase in the main index of downstream PS was approximately 8.0%.

Data-generating priors are a central component of tabular foundation models because they define the task distribution used during pretraining. However, priors are rarely evaluated as independent components, making it difficult to understand how much they affect downstream model behavior. This raises a methodological question: how can priors from different tabular foundation models be compared independently of the architectures and training protocols they were introduced with? To study this question, we implement a unified interface for publicly available priors from recent tabular foundation models and priors constructed from real datasets. We generate training tasks from each prior, train the same model architecture under a fixed training protocol, and evaluate the resulting models on shared downstream classification tasks. We compare priors through both generated-task statistics and downstream predictive performance. Our results show that different priors favor different downstream behaviors, with some achieving stronger absolute performance and others exhibiting more consistent relative rankings across datasets. We further find that data-level similarity only partially explains downstream behavior. Our code is available at https://github.com/automl/TFM-Playground/tree/prior-dev.

Advanced semiconductor nodes drastically increased demand for Transmission Electron Microscopy (TEM), yet destructive sample preparation, slow imaging and high costs severely limit the availability of diverse datasets needed for downstream machine learning (ML). Synthetic data generation is becoming essential, but current generative models often miss TEM-specific noise, structural detail, and stochastic variability crucial for evaluation. We present a Denoising Diffusion Probabilistic Model (DDPM) framework for synthetic TEM image generation under extreme data scarcity. A progressive patch-based training strategy scales from low-resolution patches to full images, enabling from-scratch training with only 15 samples. We integrate a custom TrivialAugment adaptation, cross-process domain transfer, classifier guidance, and RePaint-style inpainting, culminating in full-image generation that preserves global structural and spatial relationships in compliance with FAB metrology requirements. Beyond synthesis, we repurpose DDPM feature representations for segmentation, partitioning encoder feature maps to obtain coherent region masks. Our synthetic images achieve up to MS-SSIM > 0.98 and qualitative expert assessment consistent with structural similarity results, facilitating downstream ML training for defect detection, segmentation, and metrology while preserving statistical and physical realism.

Efficient optimization is essential for training large language models. Although intra-layer selective updates have been explored, a general mechanism that enables fine-grained control while ensuring convergence guarantees is still lacking. To bridge this gap, we propose \textbf{MGUP}, a novel mechanism for selective updates. \textbf{MGUP} augments standard momentum-based optimizers by applying larger step-sizes to a selected fixed proportion of parameters in each iteration, while applying smaller, non-zero step-sizes to the rest. As a nearly {plug-and-play} module, \textbf{MGUP} seamlessly integrates with optimizers such as AdamW, Lion, and Muon. This yields powerful variants such as \textbf{MGUP-AdamW}, \textbf{MGUP-Lion}, and \textbf{MGUP-Muon}. Under standard assumptions, we provide theoretical convergence guarantees for \textbf{MGUP-AdamW} (without weight decay) in stochastic optimization. Extensive experiments across diverse tasks, including MAE pretraining, LLM pretraining, and downstream fine-tuning, demonstrate that our \textbf{MGUP}-enhanced optimizers achieve superior or more stable performance compared to their original base optimizers. We offer a principled, versatile, and theoretically grounded strategy for efficient intra-layer selective updates, accelerating and stabilizing the training of large-scale models. The code is publicly available at https://github.com/MaeChd/MGUP.

Binary reversing is fundamental to software understanding, vulnerability discovery, malware investigation, and firmware auditing. However, it remains inherently challenging due to the irreversible loss of semantic information during compilation. Recent advances in machine learning, large language models (LLMs), and agentic AI systems have accelerated the adoption of AI-augmented binary reversing. Yet, the resulting body of work has become increasingly fragmented across reversing domains, artifact representations, learning approaches, and evaluation practices. This paper presents the first comprehensive systematization of knowledge on AI-augmented binary reversing. We analyze 144 research papers published since 2015, and organize them into 22 binary reversing domains according to the inference tasks. We further introduce a unified taxonomy spanning conventional and AI-augmented reversing pipelines. Our taxonomy connects traditional analysis techniques, binary-derived artifacts, representation strategies, learning paradigms, and downstream inference tasks, while clarifying the emerging roles of LLMs and agentic AI systems. By establishing a common vocabulary and structured framework, we provide a holistic view of the field's evolution over the past decade. Our study reveals common structures underlying seemingly disparate approaches, highlights persistent technical challenges and evaluation gaps, and identifies promising opportunities for future research. Collectively, these insights clarify the current state of the field and provide a foundation for the next generation of reliable and scalable AI-augmented binary reversing systems.

Foundation models (FMs) trained on large datasets and fine-tuned on downstream tasks have emerged as a powerful paradigm in AI for science. Industrial FMs are typically trained using self-supervision with masking due to the lack of labels. In many scientific domains, accurate simulations are plentiful and facilitate large, labeled datasets. This opens up new possibilities for pre-training. We present a systematic comparison of pre-training methods using the OmniLearned High Energy Physics FM framework. We test supervised classification, flow-matching generation, and self-supervised masked particle modeling. All models are pre-trained on the JetClass dataset and fine-tuned on two representative downstream tasks, top jet classification and JetNet conditional generation. Among other observations, for classification tasks, we find that pure classifier pre-training is optimal when downstream labels and model capacity are plentiful, but combining it with self-supervised masked particle modeling (MPM) is uniquely powerful in the low-finetuning label regime. Flow matching-based generative pre-training seems to provide little benefit for downstream classification, and interestingly, for downstream generation, we find that flow matching must be in the pre-training objective to see a significant finetuning advantage, hinting at the orthogonality of classification and generation tasks. That is, for a model to transfer to both generative and classification downstream tasks, it must be pre-trained on both. This study provides a template for controlled scaling analysis of pre-training objectives for foundation models in simulation-based sciences.

Self-supervised geospatial foundation models (GeoFMs) learn transferable representations from remote sensing data, but their downstream behavior is difficult to characterize. We study six representative GeoFMs spanning joint-embedding, reconstruction, and multimodal pretraining families, and evaluate transfer across classification, regression, and segmentation benchmarks under different label availability and downstream pipelines. We find that model rankings change across tasks and adaptation settings. Layerwise probing shows that, in most cases, task-relevant information is more accessible in intermediate transformer blocks compared to final-layer embeddings, and that GeoFMs exhibit distinct depthwise profiles. In segmentation case studies on PASTIS and Sen1Floods11, downstream adaptation settings such as decoder design and fine-tuning can be as impactful as the choice of GeoFM, and standard dense-prediction heads may be poorly aligned with how GeoFMs organize information over depth. Finally, CKA analysis on case studies shows that fine-tuning does not rewrite GeoFMs uniformly across depth, and the strongest changes are localized to the first linear layer of the MLP in ViT blocks. These results help explain why GeoFM rankings shift across benchmarks and motivate more representation-aware evaluation and adaptation strategies.

Offsite conversion rate (OCVR) prediction is an important ranking problem in computational recommendation systems. This task presents a modeling challenge: click signals are abundant and exhibit short temporal horizons, whereas conversion signals are inherently sparse, long-delayed, and frequently unattributed. Despite these statistical disparities, both signal types must inform models that operate within strict serving-latency constraints. Prior pre-training approaches address this heterogeneity with a single, undifferentiated encoder applied uniformly across both data streams. We propose DUET (Dual User Embedding Transformers), a framework that explicitly partitions user behavioral data into two domain-coherent streams -- clicks and conversions -- and pre-trains dedicated transformer encoders with architectures tailored to each stream's statistical characteristics: multi-layer self-attention for the dense click stream and interleaved cross- and self-attention for the sparse conversion stream. The resulting complementary embeddings are jointly consumed by a downstream ranker without exceeding serving-latency budgets. Evaluation demonstrates up to 0.38% normalized entropy (NE) reduction relative to the strongest baseline, and A/B test shows consistent improvements in OCVR prediction accuracy.

Scientific generative modeling often requires size transfer, where models trained on small systems are evaluated on larger ones. While translation-invariant architectures enable this evaluation, we show that architectural locality alone does not guarantee stable size extrapolation. Instead, stable extrapolation is governed by the quasi-locality of the Gaussian-smoothed score. Through Tweedie's formula, far-away perturbations can influence local score components via posterior covariance, meaning a local model succeeds only if its receptive field covers the smoothed score's response range. We formalize this mechanism, proving a size-uniform comparison theorem for local marginals under reverse diffusion. We also introduce Finite-Depth Local Flow (FDLF), a white-box diagnostic benchmark with exact scores, densities, and controllable response ranges. Empirically, we validate the interplay between spatial mixing, smoothed-score quasi-locality, and model receptive fields. Under spatial mixing, the smoothed score remains quasi-local relative to the receptive field, enabling stable extrapolation. Conversely, when spatial mixing weakens, the score's locality rapidly degrades, causing size transfer to fail.

While neural collapse (NC) predicts that a $K$-class-balanced classifier should organize terminal representations as a $(K-1)$-dimensional simplex equiangular tight frame (ETF), modular addition consistently enters a different regime: networks compress to a two-dimensional cyclic geometry in which both classifier weights and token embeddings lie on circles. We refine the explanation of this phenomenon in three directions. First, we formalize a layerwise non-uniform training mechanism: downstream classifier weights are driven by dense cross-entropy gradients into a rank-2 equiangular configuration before upstream embeddings fully reorganize, and once this classifier plane forms, backpropagated feature gradients constrain embedding motion to the same plane while weight decay suppresses orthogonal components. Second, after this subspace locking, the induced in-plane dynamics admit an entropy-regularized transport interpretation on $S^1$; combined with modular-addition labels, this reduces embedding formation to phase alignment, whose minimizers are single-frequency characters of $\mathbb{Z}/P\mathbb{Z}$ and hence equal-angle points on a circle. Third, we quantify why this solution prevails over NC: a simplex ETF gains only an $O(1)$ advantage in cross-entropy, whereas the cyclic rank-2 solution enjoys a $Θ(K)$ advantage under Schatten or weight-decay surrogates, yielding a critical threshold $λ_{\mathrm{crit}} = Θ(1/K)$. Our results explain both why classifier weights move first and why embeddings subsequently align with them, showing that grokking on modular arithmetic is governed not by maximal separation alone but by a task-structured trade-off between separation, symmetry, and complexity.

Discrete speech tokens obtained from self-supervised learning (SSL) models provide efficient data compression while maintaining strong performance, and have been widely used as intermediate representations in various tasks. However, discretization inevitably causes information loss, leading to degraded performance compared with continuous SSL features. In this work, we propose to apply soft token assignment only during downstream inference. This approach preserves the efficiency of hard discretization during training while enhancing the expressiveness of the tokens at inference. The proposed method outperforms conventional hard assignment on both ASR and speech synthesis tasks, and exhibits particularly strong generalizability to out-of-domain data. For ASR of non-native speech, it even surpasses models using continuous SSL features. Moreover, analysis of the resulting representations shows they align more accurately with phonemes compared with conventional hard assignment.

Many modern applications of deep learning involve training a neural network via a one-step prediction loss (e.g., $L^2$ regression, cross-entropy), but deploy the network by rolling out along its own predictions. Key examples include autoregressive language modeling, flow-based generative modeling, and robot policy learning. It is well-documented that these settings induce a phenomenon we call test-time feedback (TTF): the mismatch between the training/validation loss and downstream metrics of interest, such as task success rate and generation quality, which grows with task length. While data curation, architecture, and objective design have been proposed to combat train-test shift in TTF settings, this paper proposes optimization as a new design axis to mitigate error accumulation. Specifically, we introduce a new optimization paradigm called double-preconditioning (DoPr) uniquely tailored to the challenges of TTF. DoPr combines gradient-wise preconditioning, as in Adam and Muon, with activation-wise preconditioning (AP), such as in KFAC. We show that the addition of AP yields a drop-in intervention for increasing downstream model performance across a range of TTF settings. Interestingly, these gains in test-time performance do not consistently accompany improvements in validation loss, opening new questions about how to properly evaluate models trained with one-step supervised objectives.

RoPE-trained transformers distinguish absolute position in their attention patterns, even though RoPE encodes only relative offsets in the inner product. We trace this leakage to two architectural components, The causal mask is responsible for the first: its per-query softmax denominator depends on the absolute query position by construction. The residual stream supplies the second. Under causal attention the activation at position $0$ attends only to itself and runs as a closed dynamical system from the embedding of the token at that position; downstream attention reads this trajectory through sink-reading heads. Both components appear in all three architectures we study, in architecturally specific balance: NTK scaling suppresses the residual-stream component, sliding-window attention allows it to accumulate with depth, and standard RoPE sits between. Replacing the \texttt{BOS} embedding before the forward pass removes $40\%$ of the residual-stream component at early queries. Attention sinks are token-anchored stabilizers that pass forward a deterministic fingerprint of the token at position $0$, constant across inputs when that token is the auto-prepended \texttt{BOS} and varying with it otherwise.

Post-training compression of Large Language Models (LLMs) removes entire architectural components, either deleting them or replacing them with fitted modules. Existing replacement-based methods share two design constraints: full-layer granularity and contiguous selection. We argue that this is overly restrictive: in fact, redundancy in pretrained transformers is not confined to contiguous regions, nor does it evenly distribute between Attention and FeedForward outputs, implying that different strategies best approximate different submodule types and that removable components need not cluster within contiguous depth ranges. Based on this intuition, we introduce SubFit (Submodule-level Fitted residual replacement), which compresses LLMs at the submodule level: Attention and FeedForward submodules are selected non-contiguously, and each receives its own lightweight fitted residual bypass. SubFit operates post-training and requires only calibration data. Across ten LLMs (five base, five instruction-tuned), five sparsity levels from 12.5% to 37.5%, and four replacement-based baselines, SubFit achieves the best aggregate perplexity-accuracy trade-off across the evaluated sparsity levels, with larger gains under aggressive compression. At 25% sparsity, it retains 84.6% of dense downstream accuracy and incurs 2.42x perplexity degradation, against 81.6% and 4.34x for the strongest baselines, while delivering measurable inference speedup and KV-cache savings. Code is available at https://github.com/eliacunegatti/SubFit.

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.

AI agents are increasingly being tasked with automating AI research itself, particularly the critical post-training phase that transforms base LLMs into aligned assistants. However, recent evaluations reveal that even frontier agents struggle to perform this task. While the success of post-training fundamentally relies on acquiring high-quality data, relying on agents to autonomously curate targeted training datasets from the open web introduces severe challenges. Executing the long-horizon tasks of searching, filtering, and balancing data within noisy web environments frequently overwhelms an agent's limited context, ultimately leading to degraded dataset quality and suboptimal downstream training performance. To bridge this gap, we introduce Andes (Agent Native Data Evolving Synthesis), a framework that reimagines data generation as a plug-and-play \emph{agent skill}. Rather than forcing agents to devise complex data-gathering strategies from scratch, \textsc{Andes} provides an intelligent abstraction layer. By leveraging a self-evolving World Tree routing mechanism and actionable diagnostic reports, it allows trainer agents to dynamically steer data synthesis through an interactive, closed-loop interface. We demonstrate that under strict compute constraints, equipping foundationally weaker agents with Andes improves automated alignment, securing state-of-the-art performance on PostTrainBench and robust cross-task generalization. Our project is available at https://github.com/zzy1127/ANDES.

Safety detection models require examples of HHH (Helpful, Harmless, Honest)-violating outputs for robust generalization, however such examples are scarce. Activation Steering (AS) has emerged as a data-efficient method for generating target-concept-aligned responses. We investigate whether AS can generate high-quality training datasets for downstream classifiers, a question that remains untested. We present a two-fold study with intrinsic and extrinsic evaluation across $4$ concepts $\times\,2$ models $\times\,4$ steering methods. Intrinsically, beyond the field-standard rubric of steering success (concept alignment) and coherence, we introduce sample- and set-level diversity as a quality axis previously absent from the literature, and find that increasing steering strength reduces response diversity. Extrinsically, we replace HHH-violating examples in the available training data with steered generations and fine-tune detection classifiers. AS-generated data results in a better classifier than the prompting-generated data on $3$ of $4$ concepts. However, only $41$ of $136$ AS configurations outperform prompting, indicating that downstream utility lies in a narrow regime that jointly satisfies success, coherence, and diversity. The harmonic mean of these three axes correlates with downstream AUROC more consistently across concepts than success and coherence alone, providing a practical heuristic target for practitioners tuning AS hyperparameters. Together, our results highlight the potential of AS in synthetic data generation for improving safety detection and identify diversity as a critical, previously overlooked axis for tuning AS.

Scaling laws guide large language model training by relating compute to cross-entropy loss, and recent work further extends them to predict downstream benchmark performance. However, prior approaches face generalization limitations from two aspects: focusing on benchmark-level performance introduces scenario-specific artifacts, while relying on IID validation loss fails to track capability improvements when training distributions vary. In this work, we argue that downstream scaling should be studied at the capability level, which captures shared skill factors across related tasks while abstracting away benchmark-specific noise. We propose SuperValid, a framework that synthesizes OOD (out-of-distribution), capability-aligned validation data by distilling core concepts from benchmarks within a capability domain and expanding them into diverse, knowledge-rich texts. Extensive experiments spanning 16 benchmarks grouped into 6 capability domains show that SuperValid loss exhibits strong and stable correlation with downstream performance across models of different architectures, scales, and training data distributions. As a training-free metric computable during training without benchmark evaluation, SuperValid enables effective model selection, early stopping, and scaling decisions.

Multimodal Large Language Models have advanced visual reasoning, yet a purely textual chain of thought remains a bottleneck for questions that require fine-grained focus or view transformations. The ''think with images'' paradigm narrows this gap, but existing approaches are either constrained by fixed predefined toolkits or produce noisy intermediate images from unified multimodal methods. We pursue a third option: using a dedicated image editing model and decouple it with an understanding model. However, off-the-shelf image editors fail as reasoning assistants with two complementary gaps: a language-side gap, where editors trained as passive instruction-followers cannot map an abstract question to an appropriate visual transformation, and a generation-side gap, where edit correctness degrades as reasoning depth grows. Guided by this analysis, we introduce ETCHR (Editing To Clarify and Harness Reasoning), a question-conditioned, reasoning-aware image editor decoupled from the downstream understanding model and trained with a two-stage recipe targeted at the two gaps: Reasoning Imitation via supervised fine-tuning on edit trajectories, followed by Reasoning Enhancement with VLM-derived rewards for edit correctness and downstream reasoning accuracy. Since the editor is decoupled, ETCHR plugs into different open- and closed-source MLLMs in a training-free manner. Across five task families (fine-grained perception, chart understanding, logic reasoning, jigsaw restoration, and 3D understanding), ETCHR raises average Pass@1 from 55.95 to 60.77 (+4.82) with Qwen3-VL-8B, from 65.08 to 70.55 (+5.47) with Gemini-3.1-Flash-Lite, and from 76.55 to 81.16 (+4.61) with the 1T-parameter MoE model Kimi K2.5.

Motion capture is the gold standard for measuring human movement, but clinical use remains limited by cost, technical complexity, and privacy concerns. AIGaitor is a privacy-preserving, cloud-free motion analysis system that runs markerless monocular motion-capture pipelines and downstream deep-learning analysis entirely on a consumer smartphone using on-device neural accelerators. To motivate its design, we surveyed 74 rehabilitation clinicians: 92 percent said they would adopt an accurate, cost-effective, easy-to-use AI gait analysis tool, while 79.7 percent cited operating cost, 68.9 percent insufficient training, and 64.9 percent privacy concerns as leading barriers. We then optimized and benchmarked mobile iOS implementations of current monocular pipeline components, including 2D and 3D pose estimation, pose optimization, skeleton-based deep-learning analysis, and a vision-language model. A Time-Priority end-to-end on-device pipeline processes a 10 s 4K 60 fps video clip in 77 s on an iPhone 14, matching or beating the same pipeline on a high-end NVIDIA H200 cloud server when network transfer is included: 94 s at global mobile-average uplink and 66 s at developed-world Wi-Fi. Lightweight models such as ViTPose-s achieve real-time keypoint extraction, and skeleton-based action-recognition models provide sub-millisecond gait classification on the same clip. To our knowledge, AIGaitor is the first monocular system to demonstrate end-to-end on-device motion capture and downstream deep-learning analysis, supporting clinically applicable movement analysis that is low-cost, private, and accessible to smartphone users.

Many real-world combinatorial problems involve uncertain parameters, which can be predicted given contextual features and historical data. These predict-then-optimize' or contextual optimization' problems have gained significant attention: end-to-end training methods can now minimize the downstream task cost rather than the predictive error. However, despite their effectiveness, these decision-focused learning (DFL) approaches often rely on repeated solving of the underlying combinatorial optimization problem during training, making them computationally expensive and difficult to scale. We reframe the learning problem as a cost-sensitive multi-output regression problem: multi-output due to the combinatorial problem having multiple uncertain parameters, and cost-sensitive due to the downstream task cost being the real target. Our technical contribution is the formalization of multiple loss function components that follow from this reframing: cost-insensitive normalization, decision-aware asymmetric penalization of over- and underpredictions, and instance-based costs that mimic the true downstream task-based loss locally. These components require zero or one solve per training data instance, while requiring no further solves during training. Experiments show that the combination of loss components achieves comparable downstream task quality to the state of the art, while being significantly more efficient, enabling scaling to problem sizes that have not been tackled before with DFL.

Mixture-of-Experts (MoE) models rely on balanced expert utilization to fully realize their scalability. However, existing load-balancing methods are largely heuristic and operate on noisy mini-batch assignment statistics, introducing bias relative to population-level objectives. We propose $φ$-balancing, a principled framework that directly targets population-level expert balance by minimizing a strictly convex, symmetric, and differentiable potential of the expected routing distribution. Using convex duality, we derive an equivalent min-max formulation and obtain a simple online algorithm via mirror descent, yielding an efficient EMA-based routing adjustment with negligible overhead. Across large-scale pretraining and downstream fine-tuning, $φ$-balancing consistently outperforms prior Switch-style and loss-free baselines, demonstrating more stable and effective expert utilization.

Validating training data for reasoning models typically requires expensive trial-and-error fine-tuning cycles. In this work, we investigate whether the utility of a reasoning dataset can be reliably predicted prior to training using intrinsic data metrics. We propose a suite of quantitative measures and evaluate their predictive power by fine-tuning 8B and 11B models on semantically distinct variants of a Polish reasoning dataset. Our analysis reveals that these intrinsic metrics demonstrate strong and significant correlations with downstream model performance. Crucially, we find that the predictors of utility are scale-dependent: smaller models rely on alignment-focused metrics to ensure precision, whereas larger models benefit from high redundancy, utilizing verbose traces to solve complex tasks. These findings establish a scale-aware framework for validating reasoning data, enabling practitioners to select effective training sets without the need for exhaustive empirical testing.

Artificial intelligence systems typically rely on large, centrally collected datasets, a premise that does not hold in many real-world domains such as healthcare and public institutions. In these settings, data sharing is often constrained by storage, privacy, or resource limitations. For example, small wearable devices may lack the bandwidth or energy capacity needed to store and transmit high-resolution data, leading to aggregation during data collection and thus a loss of information. As a result, datasets collected from different sources may consist of a mixture of high- and low-resolution samples. Despite the prevalence of this setting, it remains unclear how informative low-resolution data is when models are ultimately evaluated on high-resolution inputs. We provide a theoretical analysis based on the Kullback-Leibler divergence that characterises how the influence of a datapoint changes with resolution, and derive bounds that relate the relative contribution of high- and low-resolution observations to the information lost under downsampling. To support this analysis, we empirically demonstrate, using both a vision transformer and a convolutional neural network, that adding low-resolution data to the training set consistently improves performance when high-resolution data is scarce.

In many real-world computer vision applications, including medical imaging and industrial inspection, binary classification tasks are characterized by a severe scarcity of positive samples. A widely adopted solution is to generate synthetic positive data using image-to-image transformations applied to negative samples. However, a fundamental challenge remains: how can we reliably assess whether such synthetic data will improve downstream model performance? In this work, we propose a geometry-driven metric that predicts the utility of synthetic data without requiring model training. Our approach operates in the embedding space of a pre-trained foundation model and represents the dataset through difference vectors between samples. We evaluate whether the weight vector of a linear classifier can be expressed within the subspace spanned by these variations by measuring the relative projection error. Intuitively, if the variations induced by synthetic data capture task-relevant directions, their span can approximate the classifier, resulting in low projection error. Conversely, poor synthetic data fails to span these directions, leading to higher error. Across multiple datasets and architectures, we show that this metric exhibits strong correlation with downstream classification performance of CNNs trained on mixtures of real negative and synthetic positive data. These findings suggest that the proposed metric serves as a practical and informative tool for evaluating synthetic data quality in data-scarce settings.

TabPFN has recently gained attention as a foundation model for tabular datasets, achieving strong performance by leveraging in-context learning on synthetic data. However, we find that TabPFN is vulnerable to label shift, often overfitting to the majority class in the training dataset. To address this limitation, we propose DistPFN, the first test-time posterior adjustment method designed for tabular foundation models. DistPFN rescales predicted class probabilities by downweighting the influence of the training prior (i.e., the class distribution of the context) and emphasizing the contribution of the model's predicted posterior, without architectural modification or additional training. We further introduce DistPFN-T, which incorporates temperature scaling to adaptively control the adjustment strength based on the discrepancy between prior and posterior. We evaluate our methods on over 250 OpenML datasets, demonstrating substantial improvements for various TabPFN-based models in classification tasks under label shift, while maintaining strong performance in standard settings without label shift. Code is available at this repository: https://github.com/seunghan96/DistPFN.

Long-tailed data bias decision boundaries toward head classes and degrade tail class accuracy. Diffusion-based generative augmentation address this problem by generating additional data, while head-to-tail transfer further mitigate the generator bias inherit from long-tailed dataset. However, we show that while head-to-tail transfer helps balance the decision space of the classifier, it also induces latent non-local feature mixing that entangles inter-class features, causing decision boundary overlap and tail class distribution shift. To address this, we first identify the problem of boundary ambiguity and then propose Decision Boundary-aware Generation (DBG) framework, which promotes near-boundary representation learning by generating informative near-boundary samples. Overall, DBG rebalances the long-tailed dataset while yielding more separable decision space for long-tailed learning. Across standard long-tailed benchmarks, DBG consistently improves tail class and overall accuracy with less inter-class overlap. The code of DBG is available at https://github.com/keepdigitalabc-svg/DBG.

Long-term conversational agents must decide which turns to store in external memory, yet recent systems rely on autoregressive LLM generation at every turn to make that decision. We present MemRouter, a write-side memory router that decouples memory admission from the downstream answer backbone and replaces per-turn memory-management decoding with an embedding-based routing policy. MemRouter encodes each turn together with recent context, projects the resulting embeddings through a frozen LLM backbone, and predicts whether the turn should be stored using lightweight classification heads while training only 12M parameters. Under a controlled matched-harness comparison on LoCoMo, where the retrieval pipeline, answer prompts, and QA backbone (Qwen2.5-7B) are held identical, MemRouter outperforms an LLM-based memory manager on every question category (overall F1 52.0 vs 45.6, non-overlapping 95% CIs) while reducing memory-management p50 latency from 970ms to 58ms. Descriptive factorial averaging further shows that learned admission improves mean F1 by +10.3 over random storage, category-specific prompting adds +5.2 over a generic prompt, and retrieval contributes +0.7. These results suggest that write-side memory admission can be learned by a small supervised router, while answer generation remains a separate downstream component in long-horizon conversational QA.

Internal ice layers imaged by radar provide key evidence of snow accumulation and ice dynamics, but radar-derived layer boundary observations are often incomplete, with discontinuous traces and sometimes entirely missing layers, due to limited resolution, sensor noise, and signal loss. Existing graph-based models for ice stratigraphy generally assume sufficiently complete layer profiles and focus on predicting deeper-layer thickness from reliably traced shallow layers. In this work, we address the layer-completion problem itself by synthesizing complete ice-layer thickness annotations from incomplete radar-derived layer traces by conditioning on colocated physical features synchronized from physical climate models. The proposed network combines geometric learning to aggregate within-layer spatial context with a transformer-based temporal module that propagates information across layers to encourage coherent stratigraphy and consistent thickness evolution. To learn from incomplete supervision, we optimize a mask-aware robust regression objective that evaluates errors only at observed thickness values and normalizes by the number of valid entries, enabling stable training under varying sparsity without imputation and steering completions toward physically plausible values. The model preserves observed thickness where available and infers only missing regions, recovering fragmented segments and even fully absent layers while remaining consistent with measured traces. As an additional benefit, the synthesized thickness stacks provide effective pretraining supervision for a downstream deep-layer predictor, improving fine-tuned accuracy over training from scratch on the same fully traced data.

Robot learning increasingly depends on large and diverse data, yet robot data collection remains expensive and difficult to scale. Egocentric human data offer a promising alternative by capturing rich manipulation behavior across everyday environments. However, existing human datasets are often limited in scope, difficult to extend, and fragmented across institutions. We introduce EgoVerse, a collaborative platform for human data-driven robot learning that unifies data collection, processing, and access under a shared framework, enabling contributions from individual researchers, academic labs, and industry partners. The current release includes 1,362 hours (80k episodes) of human demonstrations spanning 1,965 tasks, 240 scenes, and 2,087 unique demonstrators, with standardized formats, manipulation-relevant annotations, and tooling for downstream learning. Beyond the dataset, we conduct a large-scale study of human-to-robot transfer with experiments replicated across multiple labs, tasks, and robot embodiments under shared protocols. We find that policy performance generally improves with increased human data, but that effective scaling depends on alignment between human data and robot learning objectives. Together, the dataset, platform, and study establish a foundation for reproducible progress in human data-driven robot learning. Videos and additional information can be found at https://egoverse.ai/

Self-supervised pre-training, where large corpora of unlabeled data are used to learn representations for downstream fine-tuning, has become a cornerstone of modern machine learning. While a growing body of theoretical work has begun to analyze this paradigm, existing bounds leave open the question of how sharp the current rates are, and whether they accurately capture the complex interaction between pre-training and fine-tuning. In this paper, we address this gap by developing an asymptotic theory of pre-training via two-stage M-estimation. A key challenge is that the pre-training estimator is often identifiable only up to a group symmetry, a feature common in representation learning that requires careful treatment. We address this issue using tools from Riemannian geometry to study the intrinsic parameters of the pre-training representation, which we link with the downstream predictor through a notion of orbit-invariance, precisely characterizing the limiting distribution of the downstream test risk. We apply our main result to several case studies, including spectral pre-training, factor models, and Gaussian mixture models, and obtain substantial improvements in problem-specific factors over prior art when applicable.

Large language models are commonly adapted to downstream tasks through supervised fine-tuning (SFT), but substantial distribution mismatch between downstream supervision and a model's generation distribution can intensify catastrophic forgetting. Data rewriting offers a data-centric way to narrow this mismatch before SFT. Existing methods, however, typically sample rewrites from a prompt-induced conditional distribution, which need not align with the backbone's natural question-answering generation distribution, and fixed templates can reduce output diversity. We formulate data rewriting as a policy-learning problem and train a lightweight LoRA rewriting policy with reinforcement learning. The policy optimizes question-answering-style distributional alignment and semantic diversity under a hard task-consistency gate, producing verified supervision for downstream SFT. Across three instruction-tuned backbones, the resulting models attain downstream gains broadly comparable to standard SFT while reducing degradation on non-downstream benchmarks in every evaluated setting. Additional experiments on logical reasoning and medical question answering provide preliminary evidence that a rewriting policy can be reused across domains for the same backbone.

Supervised fine-tuning (SFT) datasets are critical to the downstream performance of large language models, yet they often contain low-quality or harmful question-response pairs. To improve SFT data quality, we develop a unified bilevel framework that combines offline data selection with the online self-refining generation. In the offline setting, bilevel data selection (BDS) selects question-response pairs from the offline SFT dataset to maximize the validation performance. We theoretically show that the optimal model given by BDS outperforms direct data mixing approach in useful data coverage. Moreover, we provide a global convergence analysis for gradient-based BDS approach for one-layer Transformer, showing that the epsilon-global optimum of offline BDS is achievable in finite time. Although efficient, offline BDS discards potentially harmful questions together with responses, thereby reducing question diversity. We address this limitation by refining the responses to selected questions using online self-refining generation framework. However, BDS is inefficient to update the response weights when responses are regenerated online. To address this issue, we introduce bilevel multi-objective optimization (BMO) for response-level weighting. We show that BMO recovers the same validation-aligned solution as BDS, but admits a closed-form importance-ratio weight that adapts to regenerated responses. Experiments on LLM quality enhancement and safety-aware fine-tuning demonstrate that the proposed framework consistently improves both data quality and downstream fine-tuning performance.