Attention is the dominant source of latency during long-context LLM inference, an increasingly popular workload with reasoning models and RAG. We propose Kascade, a training-free sparse attention method that leverages known observations such as 1) post-softmax attention is intrinsically sparse, and 2) the identity of high-weight keys is stable across nearby layers. Kascade computes exact Top-k indices in a small set of anchor layers, then reuses those indices in intermediate reuse layers. The anchor layers are selected algorithmically, via a dynamic-programming objective that maximizes cross-layer similarity over a development set, allowing easy deployment across models. The method incorporates efficient implementation constraints (e.g. tile-level operations), across both prefill and decode attention. The Top-k selection and reuse in Kascade is head-aware and we show in our experiments that this is critical for high accuracy. Kascade achieves up to 4.1x speedup in decode attention and 2.2x speedup in prefill attention over FlashAttention-3 baseline on H100 GPUs while closely matching dense attention accuracy on long-context benchmarks such as LongBench and AIME-24.
Dhruv Deshmukh, Saurabh Goyal, Nipun Kwatra et al.· 0 citations
We study how explicit world-modeling objectives affect the internal representations and downstream capability of Transformers, using Rubik's Cubes as our training domain. We ask: (1) how does explicitly pretraining a world model affect a model's latent representations, (2) how does world-model quality affect post-training performance, and (3) how should a finite data budget be split between pretraining and task-specific fine-tuning? We compare standard action-prediction fine-tuning with two strategies that add explicit state-prediction supervision: state pretraining followed by fine-tuning, and joint action and state training. We measure task accuracy after Group Relative Policy Optimization (GRPO). We further find that explicit world-modeling yields better representations in terms of higher probing accuracy and steerability of the model, and that better representations yield larger gains from GRPO, especially on harder cube states. Finally, when the fine-tuning budget is held fixed, probe accuracy strongly predicts the GRPO improvement. Under a fixed total data budget, accuracy is maximized by allocating only a small fraction to pretraining.
Prakhar Gupta, Henry Conklin, Sarah-Jane Leslie et al.· 0 citations
Training process reward models (PRMs) requires step-level correctness labels, obtained either through expensive human annotation or by relying on ground-truth answers, limiting the ability to scale process-level supervision. We propose ScalePRM, which scales verification compute as an alternative: given a problem and a candidate solution, we generate multiple independent verifications of each reasoning step and aggregate their judgments to produce synthetic step-level labels without ground truth. We explore two representative inference-time scaling strategies, parallel scaling through self-consistency and sequential scaling through meta-critique, and train generative PRMs on the resulting synthetic data. On ProcessBench, a benchmark for identifying erroneous steps in mathematical reasoning, PRMs trained on step-level self-consistency data achieve 67.5 F1, surpassing reference-guided training with ground-truth access (66.4 F1) and GPT-4o as a critic (61.9 F1). When deployed as reward signals in RL training with Qwen2.5-Math-7B, our best PRM achieves 47.4% average accuracy across six mathematical reasoning benchmarks, outperforming ground-truth-based RLVR (43.9%). We also identify and address reward exploitation patterns unique to generative PRM-based RL. Our results demonstrate that scaling verification compute is a viable alternative to ground-truth supervision for training process reward models.
Salman Rahman, Sruthi Gorantla, Arpit Gupta et al.· 0 citations
Deploying deep neural networks on resource-constrained devices faces two critical challenges: maintaining accuracy under aggressive quantization while ensuring predictable inference latency. We present a curiosity-driven quantized Mixture-of-Experts framework that addresses both through Bayesian epistemic uncertainty-based routing across heterogeneous experts (BitNet ternary, 1-16 bit BitLinear, post-training quantization). Evaluated on audio classification benchmarks (ESC-50, Quinn, UrbanSound8K), our 4-bit quantization maintains 99.9 percent of full-precision F1 (0.858 vs 0.859) with 4x compression and 31 percent energy savings versus 8-bit, while both achieve statistical parity with full precision (p > 0.05).
Crucially, curiosity-driven routing simultaneously improves accuracy and stability: on Quinn, F1 increases from 0.802 to 0.809 while cross-fold variance drops by 85 percent (p < 0.001, Levene's test), with reductions of 50 to 94 percent across datasets. The routing is self-organizing, with the high-precision 8-bit expert automatically receiving the most uncertain samples (20 percent lower confidence, p < 0.001), while lightweight experts handle easier inputs. Datasets with already low baseline variance show no artificial stability gain, confirming the mechanism targets genuine epistemic uncertainty rather than overfitting routing decisions.
At 1.2M parameters, the framework provides interpretable, precision-aware routing suitable for safety-sensitive edge deployments where both accuracy and predictability are critical.
Sebasti\'an Andr\'es Cajas Ord\'o\~nez, Luis Fernando Torres Torres, Mackenzie J. Meni et al.· 0 citations
Reach audiences
Advertise in front of researchers, engineers, and readers.
Personalized treatment outcome prediction based on trial data for small-sample and rare patient groups is a critical task in precision medicine. However, the high cost and scarcity of trial data limit the prediction performance. To address this issue, we propose a cross-fidelity knowledge distillation and adaptive fusion network (CFKD-AFN), which leverages abundant but low-fidelity simulation data to enhance the prediction on scarce but high-fidelity trial data. CFKD-AFN incorporates a dual-channel knowledge distillation module to extract complementary knowledge from the low-fidelity model, along with an attention-guided fusion module to adaptively integrate multi-source information. Experiments on chronic obstructive pulmonary disease show that CFKD-AFN reduces the mean squared error by 6.67% ~ 74.55% and the mean absolute percentage error by 1.43% ~ 51.54% compared to the evaluated competing methods and remains robust to varying high-fidelity dataset sizes. Furthermore, we extend the CFKD-AFN framework to an interpretable variant for exploratory analysis of feature-attribution patterns associated with treatment outcomes.
Wenjie Chen, Li Zhuang, Ziying Luo et al.· 0 citations
Recent advances in self-supervised learning for EEG representation have largely relied on masked reconstruction, where models are trained to recover randomly masked signal segments. While effective at modeling local dependencies, the training objective of masked reconstruction does not compel the model to capture global generative constraints essential for characterizing neural activity. To address this limitation, we propose EEGDM, a novel self-supervised framework that leverages latent diffusion models to generate EEG signals as an objective. Unlike masked reconstruction, diffusion-based generation progressively denoises signals from noise to realism, compelling the model to capture holistic temporal patterns and cross-channel relationships. Specifically, EEGDM incorporates an EEG encoder that distills raw signals and their channel augmentations into a compact representation, which serves as conditional information to guide the diffusion denoising process, thereby enabling the encoder and diffusion model to be jointly optimized through the generative objective. This design endows EEGDM with a compact latent space, which not only offers ample control over the generative process but also can be leveraged for downstream tasks. Experimental results show that EEGDM (1) reconstructs high-quality EEG signals, (2) learns robust representations, and (3) achieves competitive performance across diverse downstream tasks, thus exploring a new direction for self-supervised EEG representation learning.
Shaocong Wang, Tong Liu, Yihan Li et al.· 0 citations
Tabular data derives its value from inter-feature dependencies, yet preserving them during synthesis is fragile when samples are scarce. Existing approaches either learn dependencies implicitly through distribution fitting, rely on statistical graph learning that becomes unstable with few samples, or encode structure through flat text serialization. Recent graph-aware methods use dependency graphs as attention biases or as backbones for non-LLM samplers, but they do not use the graph as a prompt-level plan for organizing the LLM's own generation process. We introduce StructSynth, a framework that treats a dependency graph as a generation plan---determining the generation order, conditioning context, and scope of each black-box LLM call. In Evidence-Grounded Graph Induction, LLM reasoning and statistical association cues jointly construct a Directed Acyclic Graph (DAG) from limited samples. In Graph-Planned Conditional Synthesis, this DAG drives autoregressive synthesis in topological order, conditioning each feature on previously generated values with graph-specified parent structure guiding each step, making the conditioning schedule explicit throughout synthesis. Experiments show that StructSynth achieves state-of-the-art downstream utility and the best privacy-risk ranking among fourteen compared generators in low-data settings.
Siyi Liu, Yujia Zheng, Haoyang Li et al.· 0 citations
Learning from multi-variate time-series with heterogeneous channel configurations remains a fundamental challenge for deep neural networks, particularly in clinical domains such as intracranial electroencephalography (iEEG), where channel setups vary widely across subjects. In this work, we introduce multi-variate parallel attention (MVPA), a novel self-attention mechanism that disentangles content, temporal, and spatial attention, enabling flexible, generalizable, and efficient modeling of time-series data with varying channel counts and configurations. We use MVPA to build MVPFormer, a generative foundation model for human electrophysiology, trained to predict the evolution of iEEG signals across subjects. To support this and future efforts by the community, we release the SWEC iEEG dataset, the largest publicly available iEEG dataset to date, comprising nearly 10,000 hours of recordings from heterogeneous clinical sources. MVPFormer leverages MVPA to achieve strong generalization across subjects, demonstrating expert-level performance in several iEEG tasks. MVPFormer surpasses state-of-the-art (SOTA) Transformer baselines in seizure detection across the SWEC, the MAYO, and the FNUSA datasets, while also achieving SOTA performance on four Brain TreeBank iEEG decoding tasks (volume, pitch, onset, and speech). We further validate MVPA on standard time-series forecasting and classification tasks, where it matches or exceeds the performance of existing attention-based models. Together, our contributions establish MVPA as a general-purpose attention mechanism for heterogeneous time-series and MVPFormer as the first open-source, open-weights, and open-data iEEG foundation model with SOTA clinical performance. The code is available at https://github.com/IBM/multi-variate-parallel-transformer. The SWEC iEEG dataset is available at https://huggingface.co/datasets/NeuroTec/SWEC_iEEG_Dataset.
Francesco Carzaniga, Michael Hersche, Abu Sebastian et al.· 0 citations
Although large language models (LLMs) have achieved remarkable performance, the inherent stochasticity of their reasoning processes and varying conclusions present significant challenges. Majority voting or Best-of-N with external verifiers has been explored to mitigate this, but these approaches are limited in applicability or require additional training. To address this problem, we propose a novel framework that Recycles Few-shot examples to verify LLM outputs (ReFeri). Our key idea is to utilize the given few-shot examples not only to generate outputs, but also to evaluate the candidate outputs. Specifically, ReFeri combines a forward confidence score with a backward reconstruction penalty to select candidates that follow few-shot guidance while avoiding demonstration-specific overfitting. Experiments with three different LLMs across seven diverse tasks demonstrate that our framework significantly improves the accuracy of LLMs---achieving an average relative gain of 8.2%---through effective response selection.
Dongseok Lee, Jimyung Hong, Dongyoung Kim et al.· 0 citations
$\textbf{Objective:}$ Brain-predicted age difference (BrainAGE) is a neuroimaging biomarker reflecting brain health. However, training robust BrainAGE models requires large datasets, often restricted by privacy concerns. This study evaluates the performance of federated learning (FL) for BrainAGE estimation in ischemic stroke patients treated with mechanical thrombectomy, and investigates its association with clinical phenotypes and functional outcomes.
$\textbf{Methods:}$ We used FLAIR brain images from 1674 stroke patients across 16 hospital centers. We implemented standard machine learning and deep learning models for BrainAGE estimates under three data management strategies: centralized learning (pooled data), FL (local training at each site), and single-site learning. We reported prediction errors and examined associations between BrainAGE and vascular risk factors (e.g., diabetes mellitus, hypertension, smoking), as well as functional outcomes at three months post-stroke. Logistic regression evaluated BrainAGE's predictive value for these outcomes, adjusting for age, sex, vascular risk factors, stroke severity, time between MRI and arterial puncture, prior intravenous thrombolysis, and recanalisation outcome.
$\textbf{Results:}$ While centralized learning yielded the most accurate predictions, FL consistently outperformed single-site models. BrainAGE was significantly higher in patients with diabetes mellitus across all models. Comparisons between patients with good and poor functional outcomes, and multivariate predictions of these outcomes showed the significance of the association between BrainAGE and post-stroke recovery.
$\textbf{Conclusion:}$ FL enables accurate age predictions without data centralization. The strong association between BrainAGE, vascular risk factors, and post-stroke recovery highlights its potential for prognostic modeling in stroke care.
Vincent Roca, Marc Tommasi, Paul Andrey et al.· 0 citations
Diffusion models represent the state-of-the-art for solving inverse problems such as image restoration tasks. Diffusion-based inverse solvers incorporate a likelihood term to guide prior sampling, generating data consistent with the posterior distribution. However, due to the intractability of the likelihood, most methods rely on isotropic Gaussian approximations, which can push estimates off the data manifold and produce inconsistent, poor reconstructions. We propose Equivariance Regularized (EquiReg) diffusion, a general plug-in framework that improves posterior sampling by penalizing trajectories that deviate from the data manifold. EquiReg formalizes manifold-preferential equivariant functions that exhibit low equivariance error for on-manifold samples and high error for off-manifold ones, thereby guiding sampling toward symmetry-preserving regions of the solution space. We highlight that such functions naturally emerge when training non-equivariant models with augmentation or on data with symmetries. EquiReg's largest gains are under reduced sampling and measurement consistency steps, where many methods suffer severe quality degradation. By regularizing trajectories toward the manifold, EquiReg implicitly accelerates convergence and enables high-quality reconstructions. EquiReg consistently improves performance in linear and nonlinear image restoration tasks and solving partial differential equations. Our code is available at https://github.com/Anima-Lab/EquiReg
Bahareh Tolooshams, Aditi Chandrashekar, Rayhan Zirvi et al.· 0 citations
Finetuning open-weight Large Language Models (LLMs) is standard practice for achieving task-specific performance improvements. Until now, finetuning has been regarded as a controlled and secure process in which training on benign datasets leads to predictable behaviors. In this paper, we demonstrate, for the first time, that an adversary can create compromised LLMs that are performant and benign, yet exhibit adversarial behaviors once finetuned by downstream users. To this end, we propose an attack, FAB (Finetuning-activated Adversarial Behaviors), which compromises an LLM via meta-learning techniques that simulate downstream finetuning, explicitly optimizing for the emergence of adversarial behaviors in the finetuned models. At the same time, the compromised LLM is regularized to retain general capabilities and to exhibit no adversarial behaviors prior to finetuning. As a result, when users finetune (e.g., instruction-tuning, distillation, DPO) the seemingly benign model on their own datasets, they unknowingly trigger its dormant adversarial behavior. We experimentally demonstrate the effectiveness of FAB across multiple LLMs and three commonly considered target behaviors: unsolicited advertising, jailbreakability, and over-refusal. We show that FAB-triggers are robust to various finetuning choices made by the user (e.g., dataset, number of steps, scheduler, post-training algorithm). Our findings challenge prevailing assumptions on the security of finetuning, revealing a critical attack vector.
Thibaud Gloaguen, Mark Vero, Robin Staab et al.· 0 citations
