ICML 2026 论文列表

While state-of-the-art language models (LMs) surpass the vast majority of humans in certain domains, their reasoning remains largely opaque, reducing trust and risking deception and hallucination. In this work, we introduce the Prototype Transformer (ProtoT)—an autoregressive LM architecture that replaces the quadratic-cost self-attention in the transformer with a linear-cost module based on prototypes (parameter vectors). In ProtoT, the prototypes create communication channels aggregating contextual information at different time scales. We show that this leads to the prototypes automatically capturing nameable concepts (e.g. “woman”) during training, and it provides the potential to interpret the model’s reasoning and do targeted edits of its behavior. Compared to baselines, ProtoT scales well with model and data size, shows robustness to input perturbations, and performs well on text generation and downstream tasks (GLUE). These results suggest that ProtoT is a promising step toward autoregressive language models that are more interpretable by design.

Handling missing node features is a key challenge for deploying Graph Neural Networks (GNNs) in real-world domains such as healthcare and sensor networks. Existing studies mostly address relatively benign scenarios, namely benchmark datasets with (a) high-dimensional but sparse node features and (b) incomplete data generated under Missing Completely At Random (MCAR) mechanisms. For (a), we theoretically prove that high sparsity substantially limits the information loss caused by missingness, making all models appear robust and preventing a meaningful comparison of their performance. To overcome this limitation, we introduce one synthetic and three real-world datasets with dense, semantically meaningful features. For (b), we move beyond MCAR and design evaluation protocols with more realistic missingness mechanisms. Moreover, we provide a theoretical background to state explicit assumptions on the missingness process and analyze their implications for different methods. Building on this analysis, we show that a simple baseline adapted to the graph domain is competitive with respect to specialized architectures across diverse datasets and missingness regimes.

We characterize the pre-softmax attention matrix $\mathbf{QK^\top}$ in transformers as an associative memory matrix encoding pairwise associations between input features. By decomposing this matrix into its symmetric and skew-symmetric parts, we interpret the symmetric component as governing the structure of the _energy landscape_, and the skew-symmetric component as driving _circulation_ on that landscape. Leveraging the energy formulation induced by the symmetric component, we derive Hopfield-style stability measures that quantify the stability of retrieved features. We observe meaningful correlations between Hopfield-style stability measures and the fidelity--diversity trade-offs in generation. Finally, we propose a controllable knob to modulate this trade-off by modifying the circulation of the underlying dynamics. Code is available at our Project Page (https://hyeon-cho.github.io/Balancing/).

Inverse design aims to find design parameters $x$ achieving target performance $y^*$. Generative approaches learn bidirectional mappings between designs and labels, enabling diverse solution sampling. However, standard conditional flow matching (CFM), when adapted to inverse problems by pairing labels with design parameters, exhibits strong sensitivity to their arbitrary ordering and scaling, leading to unstable training. We introduce Diagonal Flow Matching (Diag--CFM), which resolves this through a zero-anchoring strategy that pairs design coordinates with noise and labels with zero, making the learning problem provably invariant to coordinate permutations. This yields substantially lower round-trip error than CFM and invertible neural network baselines across design dimensions up to $P{=}784$, including order-of-magnitude gains on several benchmarks. We develop two architecture-intrinsic uncertainty metrics, Zero-Deviation and Self-Consistency, that enable three practical capabilities: selecting the best candidate among multiple generations, abstaining from unreliable predictions, and detecting out-of-distribution targets; consistently outperforming ensemble and general-purpose alternatives across all tasks. We validate on airfoil, gas turbine combustor, scalable analytical benchmarks, a photonics inverse-design task, and an image-statistics benchmark.

Structured State-Space Duality (SSD) [Dao \& Gu, ICML 2024] is an equivalence between a simple Structured State-Space Model (SSM) and a masked attention mechanism. In particular, a state-space model with a scalar-times-identity state matrix is equivalent to a masked self-attention with a $1$-semiseparable causal mask. Consequently, the same sequence transformation (model) has two algorithmic realizations: as a linear-time $O(T)$ recurrence or as a quadratic-time $O(T^2)$ attention. In this note, we formalize and generalize this duality: (i) we extend SSD from the scalar‑identity case to general diagonal SSMs (diagonal state matrices); (ii) we show that these diagonal SSMs match the scalar case's training complexity lower bounds while supporting richer dynamics; (iii) we establish a necessary and sufficient condition under which an SSM is equivalent to $1$-semiseparable masked attention; and (iv) we show that such duality fails to extend to standard softmax attention due to rank explosion. Together, these results tighten the bridge between recurrent SSMs and Transformers, and widen the design space for expressive yet efficient sequence models. Code is available at https://github.com/MAGICS-LAB/state_space_daulity.

The rise of auto-bidding in online advertising has created new challenges for ensuring advertiser incentive compatibility, particularly when advertisers delegate bidding to agents with high-level constraints. One challenge is the multiplicity of equilibria with reported constraints. Alimohammadi et al. (2023) proposed a notion of Auto-bidding Incentive Compatibility (AIC) which serves to highlight that standard auctions may not incentivize truthful reporting of these constraints. However, their definition of AIC is very stringent as it requires that the worst-case outcome of an advertiser's truthful report is at least as good as the best-case outcome of any of the advertiser's possible deviations. In this paper, we introduce two refined and relaxed concepts: Risk-Averse Auto-bidding Incentive Compatibility (RAIC) and Optimistic Auto-bidding Incentive Compatibility (OAIC). RAIC (OAIC) stipulates that truthful reporting is preferred if its least (most) favorable equilibrium outcome is no worse than the least (most) favorable equilibrium outcome from any misreport. We demonstrate that SPA satisfies both RAIC and OAIC. These findings clarify SPA's incentive properties under auto-bidding, specifically regarding advertiser perspectives on equilibrium selection.

While edge GPUs are increasingly used for latency-critical DNN tasks, limited resources often fail to meet strict real-time (RT) requirements under concurrent workloads. Existing preemption and early-exit mechanisms often underutilize GPU resources through single-task queuing and sacrifice excessive accuracy during task bursts. To address this, we propose RTInfer, a novel system that enables concurrent RT task execution while balancing throughput and accuracy. RTInfer integrates an accuracy-calibrated lightweight variant co-optimization to generate efficient models, a memory-layout-aware scheduler to mitigate fragmentation during preemption, and an on-demand loading strategy to minimize host-to-GPU latency. Extensive evaluations demonstrate that RTInfer outperforms state-of-the-art methods by reducing average deadline miss rate (DMR) from 32.8\% to 0\% and improving accuracy by up to 56.5\%.

The specification of the action space plays a pivotal role in imitation-based robotic manipulation policy learning, fundamentally shaping the optimization landscape of policy learning. While recent advances have focused heavily on scaling training data and model capacity, the choice of action space remains guided by ad-hoc heuristics or legacy designs, leading to an ambiguous understanding of robotic policy design philosophies. To address this ambiguity, we conducted a large-scale and systematic empirical study, confirming that the action space does have significant and complex impacts on robotic policy learning. We dissect the action design space along temporal and spatial axes, facilitating a structured analysis of how these choices govern both policy learnability and control stability. Based on 13,000+ real-world rollouts on a bimanual robot and evaluation on 500+ trained models over four scenarios, we examine the trade-offs between absolute vs. delta representations, and joint-space vs. task-space parameterizations. Our large-scale results suggest that properly designing the policy to predict delta actions consistently improves performance, while joint-space and task-space representations offer complementary strengths, favoring control stability and generalization, respectively.

The additive noise mechanism is a foundational tool for differential privacy (DP) of $T$-dimensional real-valued vector queries. The Gaussian mechanism, utilizing Gaussian noise, is the mostly widely used such mechanism, due to its simplicity and strong privacy guarantees. In this work, we provide justification for this choice, showing that as the dimension $T\to\infty$, no additive-noise mechanism can asymptotically improve on the Gaussian mechanism's privacy--utility tradeoff for the strong privacy settings typically used. We also develop a new family of **Spherical Generalized Gamma** DP mechanisms, which contains both the Gaussian mechanism and the recently studied $\ell_2$ mechanism (Joseph \emph{et al.}, ICML 2025). We identify members of this family that outperform both the Gaussian and $\ell_2$ mechanisms in certain low-dimensional settings, and show tight composition of all mechanisms in this family, answering an open question of Joseph \emph{et al.}~regarding the $\ell_2$ mechanism.

Benchmark Design in Black-Box Optimization (BBO) is a fundamental yet open-ended topic. Early BBO benchmarks are predominantly human-crafted, introducing expert bias and constraining diversity. Automating this design process can relieve the human-in-the-loop burden while enhancing diversity and objectivity. We propose Evolution of Benchmark (EoB), an automated BBO benchmark designer empowered by the large language model (LLM) and its program evolution capability. Specifically, we formulate benchmark design as a bi-objective optimization problem towards maximizing (i) landscape similarity to target tasks and (ii) algorithm-differentiation ability across a portfolio of BBO solvers. Under this paradigm, EoB iteratively prompts LLM to evolve a population of benchmark programs and employs a reflection-based scheme to co-evolve the landscape and its corresponding program. Comprehensive experiments validate our EoB is a competitive candidate in multi-dimensional usages: 1) Benchmarking BBO algorithms; 2) Training and testing learning-assisted BBO algorithms; 3) Extending proxy for expensive real-world problems.

We introduce Mahalanobis-Pruned Mixture-of-Experts (MP-MoE), a novel routing framework that approaches expert selection from the perspective of ensemble pruning. Existing Mixture-of-Experts (MoE) routing strategies often suffer from representation collapse due to greedy top-k selection mechanisms or rely on complex auxiliary regularization terms that may compromise model performance. To address these issues, we formulate routing as a diversity-aware subset selection problem and optimize a Mahalanobis-distance-based objective that explicitly enhances expert diversity. Specifically, we demonstrate that the expert co-occurrence matrix effectively captures inter-expert correlations, allowing us to efficiently model the covariance structure required for distance computation without accessing expert parameters. Furthermore, we devise a greedy strategy for the routing mechanism, backed by theoretical approximation guarantees, rendering it a plug-and-play module with negligible overhead. MP-MoE increases wall-clock training time by approximately 3\%, while incurring no additional latency at inference time. Extensive experiments demonstrate that during the pre-training of the large language model, our method consistently outperforms the baseline by 1-3 percentage points across a broad range of benchmarks.

The unordered nature of point cloud data poses significant challenges to conventional architectures primarily designed for structured data. Spiking neural networks (SNN), by virtue of their inherent sparsity and dynamics, are particularly well-suited for processing point clouds to effectively extract meaningful features. We propose SpikeNet, a novel spiking neural network architecture for energy-efficient and robust point cloud analysis. We introduce spiking-driven sparse attention mechanism coined the Spiking Vector Mask Transformer (SVMT). By dynamically aligning the sparsity of point cloud data through binary spiking masks, SVMT eliminates the need for softmax and multiplication operations, significantly improving computational efficiency. We also propose a Dynamic Sparse Spiking Residual (DSSR) structure and integrate it with SVMT to form the Spiking Neural Network (SpikeNet) for point cloud classification and segmentation. SpikeNet overcomes the trade-off between accuracy and efficiency in previous SNN methods, achieving collaborative optimization of performance and energy-efficiency. Experiments on benchmark datasets show that SpikeNet achieves state-of-the-art performance in shape classification and segmentation tasks, comparable to artificial neural network (ANN) based methods. Our source code is in supplementary material and will be made publicly available.

Video editing has evolved toward In-Context Learning (ICL) paradigms, yet the resulting quadratic attention costs create a critical computational bottleneck. In this work, we propose **I**n-context **S**parse **A**ttention (**ISA**), the first experimentally lossless sparse framework tailored for ICL video editing. Our design is grounded in two key insights: __**first**__, context tokens exhibit significantly lower saliency than source tokens; __**second**__, we theoretically prove and empirically validate that Query sharpness correlates with approximation error. Motivated by these findings, ISA implements an efficient pre-selection strategy to prune redundant context, followed by a dynamic query grouping mechanism that routes high-error queries to full attention and low-error ones to a computationally efficient 0-th order Taylor sparse attention. Furthermore, we construct a scalable pipeline to curate a 1M-sample dataset and train __**LIVEditor-14B**__, a novel lightning video editing model via ISA. Extensive experiments demonstrate that LIVEditor-14B achieves a ~60% reduction in latency while surpassing state-of-the-art methods across EditVerseBench, IVE-Bench, and VIE-Bench, delivering experimentally lossless acceleration without compromising visual fidelity.

Floorplanning determines the coordinate and shape of each module in Integrated Circuit. With the advancement of technology node, in floorplanning stage especially 3D scenarios with multiple stacked layers, it has become increasingly challenging to adhere to complex hardware design rules. Current methods are only capable of handling specific and limited design rules, while violations of other rules require manual and meticulous adjustment. This leads to labor-intensive and time-consuming post-processing for expert engineers. In this paper, we propose an all-in-one deep reinforcement learning-based approach to tackle these challenges, and design novel representations for real-world IC design rules that have not been addressed by previous approaches. Specifically, the processing of various hardware design rules is unified into a single framework with three key components: 1) novel matrix representations to model the design rules, 2) constraints on the action space to filter out invalid actions that cause rule violations, and 3) quantitative analysis of constraint satisfaction as reward signals. Experiments on public benchmarks demonstrate the effectiveness and validity of our approach. Furthermore, transferability are well shown on unseen circuits. Our framework is extensible to accommodate new design rules, thus providing flexibility to address emerging challenges in future chip design.

Tertiary structure-based RNA design aims to generate RNA sequences that can fold into desired 3D structures, but remains a challenging problem due to the scarcity of annotated data, structural noise, and the intrinsic complexity of RNA topology. Existing structure-to-sequence frameworks largely rely on static k-nearest neighbor graphs and rigid message passing schemes, which fail to capture the flexible and heterogeneous nature of RNA geometry. To address these issues, we propose a unified framework, ATL-FGD, that integrates Adaptive Topology Learning (ATL) and FiLM-Guided Distillation (FGD) for robust RNA design. ATL introduces a differentiable edge gating mechanism to jointly learn topology and representation, enabling the model to construct data-driven, layer-adaptive graphs that better reflect structural dynamics and biochemical consistency. On top of this, FGD bridges structural and sequence representations via feature-wise linear modulation, softly transferring the semantic knowledge from RNA foundation models without relying on them during inference. Extensive experiments on tertiary structure-based RNA design benchmarks demonstrate that our approach achieves significant improvements in both sequence recovery and structural fidelity.

A wide range of optimization problems can often be written in terms of generalized convex functions (GCFs). When this structure is present, it can convert certain nested bilevel objectives into single-level problems amenable to standard first-order optimization methods. We provide a new differentiable layer with a convex parameter space and show (Theorems 5.1 and 5.2) that it and its gradient are universal approximators for GCFs and their gradients. We demonstrate how this parameterization can be leveraged in practice by (i) learning optimal transport maps with general cost functions and (ii) learning optimal auctions of multiple goods. In both these cases, we show how our layer can be used to convert the existing bilevel or min-max formulations into single-level problems that can be solved efficiently with first-order methods.

Scaling the context length of large language models (LLMs) offers significant benefits but is computationally expensive. This expense stems primarily from the self-attention mechanism, whose $O(N^2)$ complexity with respect to sequence length presents a major bottleneck for both memory and latency. Fortunately, the attention matrix is often sparse, particularly for long sequences, suggesting an opportunity for optimization. Block-sparse attention has emerged as a promising solution that partitions sequences into blocks and skips computation for a subset of these blocks. However, the effectiveness of this method is highly dependent on the underlying attention patterns, which can lead to sub-optimal block-level sparsity. For instance, important key tokens for queries within a single block may be scattered across numerous other blocks, leading to computational redundancy. In this work, we propose Permuted Block-Sparse Attention (**PBS-Attn**), a plug-and-play method that leverages the permutation properties of attention to increase block-level sparsity and enhance the computational efficiency of LLM prefilling. We conduct comprehensive experiments on challenging long-context datasets, demonstrating that PBS-Attn consistently outperforms existing block-sparse attention methods in model accuracy and closely matches the full attention baseline. Powered by our custom permuted-FlashAttention kernels, PBS-Attn achieves an end-to-end speedup of up to $\mathbf{2.75\times}$ in long-context prefilling, confirming its practical viability. Code available at \url{https://github.com/xinghaow99/pbs-attn}.

We introduce ideal attribution mechanisms, a formal abstraction for reasoning about attribution decisions over strings. At the core of this abstraction lies the ledger, an append-only log of the prompt-response interaction history between a model and its user. Each mechanism produces deterministic decisions based on the ledger and an explicit selection criterion, making it well-suited to serve as a ground truth for attribution. We frame the design goal of watermarking schemes as faithful representation of ideal attribution mechanisms. This novel perspective brings conceptual clarity, replacing piecemeal probabilistic statements with a unified language for stating the guarantees of each scheme. It also enables precise reasoning about desiderata for future watermarking schemes, even when no current construction achieves them, as the ideal functionalities are specified first. In this way, the framework provides a roadmap that clarifies which guarantees are attainable in an idealized setting and worth pursuing in practice.

Masked Diffusion Models (MDMs) have emerged as a promising approach for generative modeling in discrete spaces. By generating sequences in any order and allowing for parallel decoding, they enable fast inference and strong performance on non-causal tasks. However, this flexibility comes with a *training complexity* trade-off: MDMs train on an exponentially large set of masking patterns, which is not only computationally expensive, but also creates a train--test mismatch between the random masks used in training and the highly structured masks induced by inference-time unmasking. In this work, we propose Progressive UnMAsking (PUMA), a simple modification of the forward masking process that aligns training-time and inference-time masking patterns, thereby focusing optimization on *inference-aligned masks* and speeding up training. Empirically, PUMA speeds up pretraining at the 125M scale by $\approx 2.3 \times$ and offers complementary advantages on top of common recipes like autoregressive initialization. We open-source our codebase at https://github.com/JaeyeonKim01/PUMA.