ICML 2026 Papers

We propose *refined GRFs* (GRFs++), a new class of *Graph Random Features* (GRFs) for efficient and accurate computations involving kernels defined on the nodes of a graph. GRFs++ resolve some of the long-standing limitations of regular GRFs, including difficulty modeling relationships between more distant nodes. They reduce dependence on sampling long graph random walks via a novel *walk-stitching* technique, concatenating several shorter walks without breaking unbiasedness. By applying these techniques, GRFs++ inherit the approximation quality provided by longer walks but with greater efficiency, trading sequential inefficient sampling of a long walk for parallel computation of short walks and matrix-matrix multiplication. Furthermore, GRFs++ extend the simplistic GRFs walk termination mechanism (Bernoulli schemes with fixed halting probabilities) to a broader class of strategies, applying general distributions on the walks' lengths. This improves approximation accuracy of graph kernels, without incurring extra computational cost. We provide empirical evaluations to showcase our claims and complement our results with theoretical analysis.

Evaluating generative models is challenging because standard metrics often fail to reflect human preferences. Human evaluations are more reliable but costly and noisy, as participants vary in expertise, attention, and diligence. Pairwise comparisons improve consistency, yet aggregating them into overall quality scores requires careful modeling. Bradley-Terry-based methods update item scores from comparisons, but existing approaches either ignore rater variability or lack convergence guarantees, limiting robustness and interpretability. We introduce BBQ, a Bayesian Bradley-Terry variant that explicitly models rater quality, downweighting or removing unreliable participants, and provides guaranteed monotonic likelihood convergence through an Expectation-Maximization algorithm. Empirical results show that BBQ provides efficient inference, well-calibrated uncertainty estimates, and more robust, interpretable rankings compared to baseline Bradley-Terry models, even with noisy or crowdsourced raters. This framework enables more reliable and cost-effective human evaluation of generative models.

State-space language models such as Mamba match Transformer quality while permitting linear complexity inference, yet still comprise billions of parameters that hinder deployment. While existing one-shot pruning methods are effective for generic linear and attention blocks, they are not designed with the overall Mamba architecture in mind and fail to account for the time-shared and discretized state-transition matrix at the heart of the selective state-space module (SSM). In this paper, we introduce SparseSSM, the first training-free pruning framework that extends the classic optimal brain surgeon (OBS) framework to state space architectures. Our layer-wise algorithm (i) derives an approximate second-order saliency score that aggregates Hessian-trace information across time steps, (ii) incorporates a component sensitivity analysis to guide feed-forward network (FFN) pruning, which also sheds light on where redundancy resides in mamba architecture, (iii) can be easily extended to semi-structured and structured sparsity, and generalized to other SSM-based architectures. Empirically, we prune 50% of SSM weights without fine-tuning and observe only limited performance degradation, achieving the current state-of-the-art one-shot pruning algorithm for Mamba-based LLMs.

Large language models have transformed many applications but remain expensive to train. Sparse Mixture of Experts (MoE) addresses this through conditional computation, with Expert Parallel (EP) as the standard distributed training method. However, EP has three limitations: communication cost grows linearly with the number of activated experts $k$, load imbalance affects latency and memory usage, and data-dependent communication requires metadata exchange. We propose Multi-Head LatentMoE and Head Parallel (HP), a new architecture and parallelism that achieve $O(1)$ communication cost regardless of $k$, completely balanced traffic, and deterministic communication, all while remaining compatible with EP. To accelerate Multi-Head LatentMoE, we propose IO-aware routing and expert computation. Compared to MoE with EP, Multi-Head LatentMoE with HP trains up to $1.82\times$ faster while having better performance. With double the granularity, the performance is even better while being $1.08\times$ faster. Our method makes multi-billion-parameter foundation model research more accessible.

Upcycling, a strategy that initializes Mixture-of-Experts (MoE) by replicating pre-trained feed-forward or MoE networks to expand model capacity, has become a popular method in continual learning due to its effectiveness in mitigating catastrophic forgetting. However, existing paradigms indiscriminately expand capacity to prioritize performance at the cost of severe inefficiency, introducing severe parameter redundancy and failing to exploit structural heterogeneity. To address this, we investigate the determinants of forgetting in training dynamics using Centered Kernel Alignment (CKA) and loss landscape flatness to analyze the behavior of pre- and post-expansion MoE layers, uncovering instability in deep-layer representations and heterogeneous expert sensitivity to new tasks, thereby demonstrating the potential of selective upcycling to eliminate redundancy. Consequently, we propose a dynamic bilevel optimization framework to guide adaptive upcycling, featuring an outer loop employing a Gumbel-Softmax differentiable mask to perform Neural Architecture Search (NAS) for adaptive growth, while an inner loop optimizes weight updates via task objectives and CKA-regularized replay. Experiments on the TRACE benchmark demonstrate that our proposed method achieves better average accuracy with 80\% forgetting reduction, while effectively eliminating 60\% of redundant parameter expansion that standard upcycling would introduce.

Language models famously improve under a smooth scaling law, but some specific capabilities exhibit sudden breakthroughs in performance. Advocates of "emergence" view these capabilities as unlocked at a specific scale, but others attribute breakthroughs to superficial metric thresholding effects. We propose that breakthroughs are instead driven by continuous changes in the probability distribution of training outcomes when performance is bimodally distributed across random seeds. We show that different random seeds can produce either smooth or emergent scaling trends in synthetic length generalization tasks, multiple choice question answering, and grammatical generalization. We reveal that sharp breakthroughs in metrics are produced by underlying continuous changes in their distribution across seeds. These distributions may become abruptly bimodal at a capacity threshold—but this threshold appears at scales well before most seeds achieve breakthrough. Our observations hold true even under continuous loss metrics, confirming that random variation must be considered when predicting a model's performance from its scale.

Evaluating a new model on an existing benchmark is often necessary to understand its behavior before deployment. For modern evaluation frameworks, generating and evaluating a response for all queries can be prohibitively expensive. In practice, responses from previously-evaluated models are often cached -- creating a potential opportunity to use this additional information to decrease the number of queries required to accurately evaluate a new model. In this paper, we introduce an approach for predicting benchmark performance that leverages cached model responses based on the Data Kernel Perspective Space (DKPS), a method for quantifying the relationship between models in the black-box setting. Theoretically, we show that DKPS-based methods are query-efficient under certain conditions. Empirically, we demonstrate that DKPS-based methods achieve the same mean absolute error as baselines with a substantially decreased query budget. We conclude by proposing an offline method for selecting a set of queries that maximizes the goodness-of-fit on reference models, improving prediction accuracy over random query selection.

Deep learning approaches typically require prohibitive amounts of data to approximate known-constraint Exact Cover tasks, while existing neuro-symbolic methods often face training infeasibility and scalability bottlenecks. To bridge this divide, we propose the Hypergraph Optimization Network (HONet), an end-to-end framework integrating a structure-preserving Deep Residual Hypergraph Encoder with a differentiable fixed-constraint Quadratic Programming layer. By adopting a ``Fixed Polytope'' paradigm guided by the Geometric Consistency Loss, HONet explicitly shapes the objective landscape, encouraging the valid discrete solution to align with the global energy minimum. Empirical results show that HONet rapidly achieves 100\% accuracy on $9 \times 9$ Sudoku using limited samples, exhibiting superior data efficiency over baselines while maintaining exceptional robustness in highly sparse regimes and additional tasks.

Test-time scaling can improve model performance by aggregating stochastic reasoning trajectories. However, achieving sample-efficient test-time self-consistency under a limited budget remains an open challenge. We introduce PETS (**P**rincipled and **E**fficient **T**est-Time **S**elf-Consistency), which initiates a principled study of trajectory allocation through an optimization framework. Central to our approach is the *self-consistency rate*}, a new measure defined as agreement with the infinite-budget majority vote. This formulation makes sample-efficient test-time allocation theoretically grounded and amenable to rigorous analysis. We study both offline and online settings. In the offline regime, where all questions are known in advance, we connect trajectory allocation to crowdsourcing, a classic and well-developed area, by modeling reasoning traces as workers. This perspective allows us to leverage rich existing theory, yielding theoretical guarantees and an efficient majority-voting-based allocation algorithm. In the online streaming regime, where questions arrive sequentially and allocations must be made on the fly, we propose a novel method inspired by the offline framework. Our approach adapts budgets to question difficulty while preserving strong theoretical guarantees and computational efficiency. Experiments show that PETS consistently outperforms uniform allocation. On GPQA, PETS achieves perfect self-consistency in both settings while reducing the sampling budget by up to $75\\%$ (offline) and $55\\%$ (online) relative to uniform allocation.

Fine-tuning large language models (LLMs) with backpropagation achieves high performance but incurs substantial memory overhead, limiting scalability on resource-constrained hardware. Zeroth-order (ZO) optimization provides a memory-efficient alternative by relying solely on forward passes, yet it typically suffers from slow or unstable convergence due to high-variance gradient estimates. Sparse ZO updates partially address this issue by perturbing only a subset of parameters, but their effectiveness hinges on selecting informative parameters, which is challenging in ZO optimization because each query yields only scalar feedback. We propose Adaptive Curvature-Guided Sparse Zeroth-Order Optimization (CurvZO), which tracks curvature signals online from scalar ZO feedback and leverages these signals to construct a parameter-wise sampling distribution for selecting coordinates at each update, reducing the variance of the sparse ZO gradient estimator. Moreover, CurvZO dynamically adapts the perturbation budget to the evolving curvature signal distribution, yielding sparse ZO updates that remain both focused and sufficiently exploratory. Extensive experiments on OPT and Llama across diverse NLP tasks show that CurvZO consistently improves fine-tuning performance and reduces training time over ZO baselines. It improves accuracy by up to 4.4 points and achieves up to a $2\times$ speedup, while preserving memory efficiency.

Conformal prediction provides rigorous, distribution-free uncertainty guarantees, but often yields prohibitively large prediction sets in structured domains such as routing, planning, or sequential recommendation. We introduce graph-based conformal compression, a framework for constructing compact subgraphs that preserve statistical validity while reducing structural complexity. We formulate compression as selecting a smallest subgraph capturing a prescribed fraction of the probability mass, and reduce to a weighted version of densest-k-subgraphs in hypergraphs, in the regime where the subgraph has a large fraction of edges. We design efficient approximation algorithms that achieve constant factor coverage and size trade-offs. Crucially, we prove that our relaxation satisfies a monotonicity property, derived from a connection to parametric minimum cuts, which guarantees the nestedness required for valid conformal calibration. Our results therefore not only highlight an algorithmic regime, distinct from classical densest-k-subgraph hardness settings, where the problem can be approximated efficiently, but also bridge conformal prediction with combinatorial graph compression via monotonicity. We finally validate our algorithmic approach via simulations for trip planning and navigation, and compare to natural baselines.

Time series forecasting plays a crucial role in a wide range of real-world applications and has become increasingly complex with the growth of multivariate dimensions and extended historical observations, leading to the prosperity of deep forecasting models. Previous models are hindered by three major challenges: high computational complexity, inefficient token utilization caused by redundancy and scarcity, and temporal distribution shifts resulting from non-stationary dynamics. Inspired by Koopman theory and the success of multilevel encoder–decoder architectures with skip connections, we design an input-dependent Koopman module to decompose time series into Koopman dynamics and residual dynamics. Building upon this formulation, we propose a U-shaped Multilevel Attention module (UMA) that integrates element-wise attention filtering and linear attention, giving rise to KUMA. The input-dependent Koopman operator mitigates the issue of operator mixture and alleviates temporal distribution shifts, while UMA achieves a favorable balance between token redundancy and token scarcity with acceptable computational efficiency. Comprehensive evaluations across 12 benchmark datasets demonstrate that KUMA achieves superior performance compared to existing excellent approaches.

Partial participation is essential for communication-efficient federated learning at scale, yet existing Byzantine-robust methods typically assume full client participation. In the partial participation setting, a majority of the sampled clients may be Byzantine, once Byzantine clients dominate, existing methods break down immediately. We introduce delayed momentum aggregation, a principle where the central server aggregates cached momentum from non-sampled clients along with fresh momentum from sampled clients. This principle ensures Byzantine clients remain a minority from the server's perspective even when they dominate the sampled set. We instantiate this principle in our optimizer DeMoA. We analyze the convergence rate of DeMoA, showing that DeMoA is Byzantine-robust under partial participation. Experiments show that, with 20% Byzantine ratio and only 10% partial participation rate, DeMoA achieves the best accuracy even when existing methods fail empirically.

A simple strategy for improving LLM accuracy, especially in math and reasoning problems, is to sample multiple responses and submit the answer most consistently reached. In this paper we leverage Bayesian prior information to save on sampling costs, stopping once sufficient consistency is reached. Although the exact posterior is computationally intractable, we further introduce an efficient ``$L$-aggregated'' stopping policy that tracks only the $L-1$ most frequent answer counts. Theoretically, we prove that $L=3$ is all you need: this coarse approximation is sufficient to achieve asymptotic optimality, and strictly dominates prior-free baselines, while having a fast posterior computation. Empirically, this identifies the most consistent (i.e., mode) LLM answer and achieves similar answer accuracy using fewer samples.

Foundation models for genomics have the potential to revolutionize therapeutic design, yet the optimal architectural choices for modeling the vast and diverse distribution of metagenomic data remain under-explored. In this work, we present the machine learning methodology behind EDEN, a family of metagenomic foundation models scaled up to 28 billion parameters and trained on 9.7 trillion nucleotide tokens. We provide a systematic empirical study of architectural trade-offs between autoregressive Transformers (Llama-style), State-Space Models (Mamba), and Long-convolutional architectures (Hyena) for nucleotide-level modeling. Contrary to recent trends favoring linear-time sequence models for long-range biological data, we demonstrate that the Llama architecture exhibits superior scaling efficiency and semantic retrieval capabilities as the model capacity grows. We derive a set of quality-aware scaling laws for metagenomics, showing how model performance follows predictable power-law behavior across three orders of magnitude in parameters and data. Through extensive benchmarking, spanning unsupervised zeroshot fitness prediction, semantic completion, and gene recovery, we establish a blueprint for scaling biological foundation models and provide empirical evidence demonstrating why Transformer-based architectures define the current frontier.

While widespread, Transformers lack inductive biases for geometric symmetries common in science and computer vision. Existing equivariant methods often sacrifice the efficiency and flexibility that make Transformers so effective through complex, computationally intensive designs. We introduce the Platonic Transformer to resolve this trade-off. By defining attention relative to reference frames from the Platonic solid symmetry groups, our method induces a principled weight-sharing scheme. This enables combined equivariance to continuous translations and Platonic symmetries, while preserving the exact architecture and computational cost of a standard Transformer. Furthermore, we show that this attention is formally equivalent to a dynamic group convolution, which reveals that the model learns adaptive geometric filters and enables a *highly scalable, linear-time convolutional variant*. Across diverse benchmarks in computer vision (CIFAR-10), 3D point clouds (ScanObjectNN), and molecular dynamics, property prediction and generation (OMol25, ProteinMD, QM9), the Platonic Transformer achieves competitive performance by leveraging these geometric constraints at no additional cost.

Transformers serve as the foundation of most modern large language models. To mitigate the quadratic complexity of standard full attention, various efficient attention mechanisms, such as linear and hybrid attention, have been developed. A fundamental gap remains: their expressive power relative to full attention lacks a rigorous theoretical characterization. In this work, we theoretically characterize the performance differences among these attention mechanisms. Our theory applies to all linear attention variants that can be formulated as a recurrence, including Mamba, DeltaNet, etc. Specifically, we establish an expressiveness hierarchy: for the sequential function composition-a multi-step reasoning task that must occur within a model’s forward pass, an $(L+1)$-layer full attention network is sufficient, whereas any hybrid network interleaving $L-1$ layers of full attention with a substantially larger number ($2^{3L^2}$) of linear attention layers cannot solve it. This result demonstrates a clear separation in expressive power between the two types of attention. Our work provides the first provable separation between hybrid attention and standard full attention, offering a theoretical perspective for understanding the fundamental capabilities and limitations of different attention mechanisms.

Existing time series tokenization methods predominantly encode a constant number of samples into individual tokens. This inflexible approach can generate excessive tokens for even simple patterns like extended constant values, resulting in substantial computational overhead. Inspired by the success of byte pair encoding, we propose the first pattern-centric tokenization scheme for time series analysis. Based on a discrete vocabulary of frequent motifs, our method merges samples with underlying patterns into tokens, compressing time series adaptively. Exploiting our finite set of motifs and the continuous properties of time series, we further introduce conditional decoding as a lightweight yet powerful post-hoc optimization method, which requires no gradient computation and adds no computational overhead. On recent time series foundation models, our motif-based tokenization improves forecasting performance by 40% and boosts efficiency by 2314% on average. Conditional decoding further reduces MSE by up to 48%. In an extensive analysis, we demonstrate the adaptiveness of our tokenization to diverse temporal patterns, its generalization to unseen data, and its meaningful token representations capturing distinct time series properties, including statistical moments and trends.

Dataset pruning (DP) and dataset distillation (DD) fundamentally differ in their outputs: DP selects original image subsets, while DD generates synthetic images. Recently, DD's increasing reliance on original images suggests a convergence of the two directions. To investigate this convergence trend, we propose a unified dataset compression (DC) benchmark. This benchmark reveals an interesting trade-off for soft-label-DD: while soft labels provide valuable information, they can make the distillation process less essential, as distilled images may not always outperform random subsets. In addition, the benchmark reveals that in current stages, dataset pruning outperforms dataset distillation at small dataset sizes. Given these observations, we explore hard-label-DC as a complementary approach that emphasizes image quality while offering substantial storage efficiency. Our PCA (Prune, Combine, and Augment) is the first framework that does not rely on soft labels but instead focuses on image quality. (1) "P'' means selecting easy samples based on dataset pruning metrics, (2) "C'' indicates combining these samples effectively, and (3) "A'' is to apply constrained image augmentation during training.