ICML 2026 论文列表

Recent work describes what transformers can and cannot compute through connections to boolean circuits, but existing results lack exact characterizations and are sensitive to modeling choices. Padded transformers---to whose input filler symbols such as ``...'' are appended---emerge as a useful gadget for establishing equivalences to circuit classes by providing polynomial space for adaptive parallel computation. However, only a limited set of padded transformer idealizations has been studied, leaving open how robustly these equivalences hold under changes to attention type, model width, and uniformity. We find that, under practical assumptions, padded transformers are surprisingly robust to all of these, and identify numeric precision and model depth as the main factors affecting expressivity. Concretely, we prove that polynomially padded $\text{L-uniform}$ constant-precision transformers are equivalent to $\text{L-uniform AC}^0$, while growing-precision ones achieve $\text{L-uniform TC}^0$ regardless of width. Furthermore, looping enables sequential processing analogous to circuits: $\log^d N$-looped constant-precision transformers reach $\text{FO-uniform AC}^d$, and growing-precision ones reach $\text{FO-uniform TC}^d$. Interestingly, growing width or precision beyond logarithmic does not increase expressivity, and all our results hold for both softmax and average hard attention transformers.

Partial Differential Equations (PDEs) play a fundamental role in scientific computing, and recent efforts have sought to extend the success of foundation models to PDE solving. However, multi-physics PDE pre-training faces the unique challenge of disentangling dynamic heterogeneity to learn universal, elementary patterns that generalize to new PDEs. Additionally, cross-physics transfer lacks a theoretical framework for interpretability—specifically, understanding which pre-trained operator knowledge is effectively transferred to target PDEs. To bridge these gaps, we introduce the theory of neural operator splitting, which decomposes PDE evolution into a modulated global spectral operator and sparse local constitutive mechanisms. A key innovation is Origo, which provides a neural operator bank that enables the identification of operator-level generalization patterns. Extensive experiments demonstrate strong zero-shot generalization and mechanism-level interpretability on unseen PDEs.

Representing the past in a compressed, efficient, and informative manner is a central problem for systems trained on sequential data. The \textit{HiPPO} framework, originally proposed by Gu \& Dao et al., provides a principled approach to sequential compression by projecting signals onto orthogonal polynomial (OP) bases via structured linear ordinary differential equations. Subsequent works have embedded these dynamics in state space models (SSMs), where HiPPO structure serves as an initialization. Nonlinear successors of these SSM methods such as Mamba are state-of-the-art for many tasks with long-range dependencies, but the mechanisms by which they represent and prioritize history remain largely implicit. In this work, we revisit the HiPPO framework with the goal of making these mechanisms explicit. We show how polynomial representations of history can be extended to support capabilities of modern SSMs such as adaptive memory allocation and associative memory, while retaining direct interpretability in the OP basis. We introduce a unified framework comprising five such extensions, which we collectively refer to as a ``HiPPO zoo.'' Each extension exposes a specific modeling capability through an explicit, interpretable modification of the HiPPO framework. The resulting models adapt their memory online and train in streaming settings with efficient updates. We illustrate the behaviors and modeling advantages of these extensions through a range of synthetic sequence modeling tasks, demonstrating that capabilities typically associated with modern SSMs can be realized through explicit, interpretable polynomial memory structures.

Finding mathematical relations underlying natural phenomena is a fundamental task in scientific discovery. Recent advances in evolutionary search with Large Language Models (LLMs) show great promise by leveraging their embedded scientific knowledge. However, discovering governing equations remains challenging due to vast combinatorial hypothesis spaces with exponentially many possible relations. Existing LLM-based approaches treat LLMs as static hypothesis generators unaware of the observed scientific system, leading to suboptimal and inefficient exploration that over-relies on internal priors. To address this, we introduce Decompose, Adapt, and Evolve (DecAEvolve), a framework that combines granular feedback from symbolic term decomposition with LLM refinement through reinforcement learning fine-tuning. DecAEvolve unifies symbolic decomposition with test-time RL adaptation, enabling adaptive rather than static hypothesis generation. Our experiments across diverse scientific benchmarks demonstrate that DecAEvolve significantly improves both the accuracy of discovered equations and the efficiency of the discovery process, reducing error by up to an order of magnitude compared to state-of-the-art baselines.

Reinforcement learning with verifiable rewards (RLVR) is a cornerstone of post-training for large reasoning models, yet widely used algorithms such as Group Relative Policy Optimization (GRPO) often exhibit \textbf{diversity collapse}. We provide a geometric diagnosis by formalizing GRPO as a dynamical flow on the probability simplex. Under a mode-based coarse-graining of rollouts, we show that GRPO induces a \textbf{collision field} over correct modes, monotonically pushing towards simplex vertices and thus yielding a \textbf{winner-take-all} regime. To address this systematically, we introduce \textbf{G$^2$RPO (Geometric GRPO)}, which reshapes RLVR via principled \textbf{vector-field editing}. Concretely, we intervene at the advantage level by adding granularity bonuses inversely proportional to mode probabilities, encouraging underrepresented correct modes. The bonus has a natural geometric interpretation, and its potential performance side effects can be mitigated, thereby avoiding the usual accuracy--diversity trade-off. In experiments with 7B and 14B models trained on a math reasoning task and evaluated on \textbf{AIME 2024/2025}, GRPO loses up to \textbf{57\%} of active correct modes. In contrast, G$^{2}$RPO increases active correct-mode coverage by \textbf{172\%--205\%}, reduces concentration on any single correct mode, prevents the late-stage \emph{entropy crash}, and improves \texttt{pass@1} by \textbf{+1.4} to \textbf{+7.9} points relative to GRPO. Overall, diversity is not merely a regularizer but a \textbf{geometric property} to be controlled to improve the model without trapping it in a single dominant strategy.

Sentence representations are foundational to many Natural Language Processing (NLP) applications. While recent methods leverage Large Language Models (LLMs) to derive sentence representations, most rely on final-layer hidden states, which are optimized for next-token prediction and thus often fail to capture global, sentence-level semantics. This paper introduces a novel perspective, demonstrating that attention value vectors capture sentence semantics more effectively than hidden states. We propose *Value Aggregation (VA)*, a simple method that pools token values across multiple layers and token indices. In a training-free setting, VA outperforms other LLM-based embeddings, even matches or surpasses the ensemble-based MetaEOL. Furthermore, we demonstrate that when paired with suitable prompts, the layer attention outputs can be interpreted as aligned weighted value vectors. Specifically, the attention scores of the last token function as the weights, while the output projection matrix ($W_O$) aligns these weighted value vectors with the common space of the LLM residual stream. This refined method, termed *Aligned Weighted VA (AlignedWVA)*, achieves state-of-the-art performance among training-free LLM-based embeddings, outperforming the high-cost MetaEOL by a substantial margin. Finally, we highlight the potential of obtaining strong LLM embedding models through fine-tuning Value Aggregation.

The deployment of Large Language Models is constrained by the memory and bandwidth demands of static weights and dynamic Key-Value cache. SVD-based compression provides a hardware-friendly solution to reduce these costs. However, existing methods suffer from two key limitations: some are suboptimal in reconstruction error, while others are theoretically optimal but practically inefficient. In this paper, we propose Swift-SVD, an activation-aware, closed-form compression framework that simultaneously guarantees theoretical optimum, practical efficiency and numerical stability. Swift-SVD incrementally aggregates covariance of output activations given a batch of inputs and performs a single eigenvalue decomposition after aggregation, enabling training-free, fast, and optimal layer-wise low-rank approximation. We employ effective rank to analyze local layer-wise compressibility and design a dynamic rank allocation strategy that jointly accounts for local reconstruction loss and end-to-end layer importance. Extensive experiments across six LLMs and eight datasets demonstrate that Swift-SVD outperforms state-of-the-art baselines, achieving optimal compression accuracy while delivering 3–70$\times$ speedups in end-to-end compression time. Our code is available at https://github.com/hiahei/Swift-SVD.

Diffusion-based language models(dLLMs) have emerged as a promising alternative to autoregressive language models, offering the potential for parallel token generation and bidirectional context modeling. However, harnessing this flexibility for fully non-autoregressive decoding remains an open question, particularly for reasoning and planning tasks. In this work, we investigate non-autoregressive decoding in dLLMs by systematically analyzing its inference dynamics along the temporal axis. Specifically, we uncover an inherent failure modes in confidence-based non-autoregressive generation stem from a strong proximity bias—the denoising order tends to concentrate on spatially adjacent tokens. This local dependency leads to spatial error propagation, rendering the entire trajectory critically contingent on the initial unmasking position. Leveraging this insight, we present a minimal-intervention approach that guides early token selection, employing a lightweight planner and end-of-sequence temperature annealing. We thoroughly evaluate our method on various reasoning and planning tasks and observe substantial overall improvement over existing heuristic baselines without significant computational overhead.

Recent studies have attempted to fine-tune time-series foundation models to enhance a target dataset's forecasting performance. However, these approaches proceed without a clear criterion for identifying complex datasets that require fine-tuning due to performance degradation in zero-shot forecasting. To distinguish datasets that are more challenging than standard benchmarks, we introduce data-driven temporal complexity and multichannel complexity. *Temporal complexity* captures the difficulty of identifying distinct patterns by quantifying spectral entropy in the frequency domain, while *multichannel complexity* captures cross-channel information flow that can impact predictive uncertainty. These metrics serve as *effective proxies for performance gains* achievable through fine-tuning. Based on the two metrics, we develop *Time-PEFT*, a parameter-efficient fine-tuning framework that incorporates a frequency adapter for top-$k$ filtering and a channel adapter for multichannel modeling. With the base variant of MOMENT as a backbone, *Time-PEFT* improves performance by up to 38\% over LoRA on complex datasets.

Estimating the quality of register transfer level (RTL) designs is crucial in the electronic design automation (EDA) workflow, as it enables instant feedback on key performance metrics like area and delay without the need for time-consuming logic synthesis. While recent approaches have leveraged large language models (LLMs) to derive embeddings from RTL code and achieved promising results, they overlook the structural semantics essential for accurate quality estimation. In contrast, the control data flow graph (CDFG) view exposes the design's structural characteristics more explicitly, offering richer cues for representation learning. In this work, we introduce StructRTL, a novel structure-aware graph self-supervised learning framework for improved RTL design quality estimation. By learning structure-informed representations from CDFGs, StructRTL significantly outperforms prior art on various quality estimation tasks. To further boost performance, we incorporate a knowledge distillation strategy that transfers low-level insights from post-mapping netlists into the CDFG-based predictor. Experimental results demonstrate that StructRTL establishes new state-of-the-art results, highlighting the effectiveness of combining structural learning with cross-stage supervision.

Rich-format documents are essential for everyday operations yet costly to author, motivating the need for automated generation to enhance productivity. To this end, we present FormAct, an agentic system that generates professional rich-format documents from scratch. FormAct operates on an HTML source representation and performs iterative source refinement with an *editing agent* that invokes a suite of tools, including a syntax-aware source editor and a template retriever, and a *review agent* that critiques rendered pages to guide refinement. Additionally, we incorporate edit-triggered context compression to maintain a bounded working context and keep multi-round editing efficient. To support development and evaluation, we introduce RichDocBench for end-to-end generation, and RichDocFuzz to evaluate formatting-error recognition for reviewer agents. Through extensive automated evaluation and blind human-preference studies, we show that FormAct consistently outperforms strong baselines, including Codex-CLI, with particularly strong improvements in generating error-free, professional rich-format documents.

Transcendental functions, such as the exponential, are central to scientific computing, yet they cannot be natively calculated by digital hardware. Instead, computers must approximate these functions by combining basic operations, such as $\{+, -, \times, \div\}$, using methods like Taylor series. These methods were developed over centuries by mathematicians, who focused on approaches that could attain arbitrary accuracy. However, computers can handle most applications by using only finite-precision types, like *float32*, where any accuracy beyond the type's precision is effectively discarded. We explore, therefore, whether forgoing arbitrary accuracy can lead to the discovery of more efficient approximations. The evolutionary method of symbolic regression is particularly suitable, as it can search for arbitrary operation combinations and can optimize non-differentiable objectives, such as the number of operations used. Our results show that evolution can discover computer programs that outperform established methods in this setting, despite having no prior mathematical knowledge beyond the calculation of the basic operations. Starting from empty code, symbolic regression constructs programs representing novel mathematical expressions. In particular, we discovered a 10-operation program that approximates the exponential function to 14 significant figures, exceeding the accuracy of previously known approximations of this size by more than 6 orders of magnitude.

Large Language Models (LLMs) are increasingly deployed as autonomous agents that execute tool-augmented, multi-step tasks, where latency is a critical factor for real-world applications. Yet an overlooked threat is Reasoning-Level Denial-of-Service (R-DoS), in which an attacker preserves task correctness but degrades availability by inflating an agent’s reasoning depth or tool-use budget. We introduce OTora, the first unified, two-stage red-teaming framework for instantiating R-DoS attacks. Stage I optimizes an adversarial trigger that induces targeted tool invocations using insertion-aware scoring and dynamic target co-evolution, supporting both black-box and white-box settings. Stage II generates agent-aware reasoning payloads via an ICL-guided genetic search that amplifies overthinking while maintaining correct task outcomes. Across WebShop, Email, and OS agents built on multiple backbone models such as LLaMA-70B and GPT-OSS-120B, OTora achieves up to 10× increases in reasoning tokens and order-of-magnitude latency slowdowns, all while preserving near-baseline task accuracy. Finally, we discuss mitigation strategies for detecting and constraining abnormal reasoning and latency spikes.

Research has repeatedly demonstrated that intermediate hidden states extracted from large language models and speech audio models predict measured brain response to natural language stimuli. Yet, very little is known about the representation properties that enable this high prediction performance. Why is it the intermediate layers, and not the output layers, that are most effective for this unique and highly general transfer task? We give evidence that the correspondence between speech and language models and the brain derives from shared meaning abstraction and not their next-word prediction properties. In particular, models construct higher-order linguistic features in their middle layers, cued by a peak in the layerwise *intrinsic dimension*, a measure of feature complexity. We show that a layer's intrinsic dimension strongly predicts how well it explains fMRI and ECoG signals; that the relation between intrinsic dimension and brain predictivity arises over model pre-training; and finetuning models to better predict the brain causally increases both representations' intrinsic dimension and their semantic content. Results suggest that semantic richness, high intrinsic dimension, and brain predictivity mirror each other, and that the key driver of model-brain similarity is *rich meaning abstraction* of the inputs, where language modeling is a task sufficiently complex (but perhaps not the only) to require it.

We study the revenue of approximate correlated equilibrium in discrete first price auctions - the set of allowable bids is $\mathcal{B} = \{0, 1/k, \dots, 1 - 1/k, 1\}$ for some $k \in \mathbb{N}$. We show that the revenue of any $\epsilon$-\textit{approximate} correlated equilibrium is at least $v_2 - \Theta(1/k)- \Theta(\epsilon k^2)$, where $v_2 \geq 0$ is the second-highest valuation. Our results establish the first polynomial convergence rates on the revenue generated by no-swap regret bidders in first-price auctions. For instance, if bidders admit the optimal swap regret of $\mathcal{O}(\sqrt{k T})$, then the time-averaged revenue is at least $v_2 - \Theta(1/k) - \Theta(\epsilon)$ after $\mathcal{O}(k^5/\epsilon^2)$ rounds.

Tokenization is the first point of contact between large language models (LLMs) and text data, yet it has not been viewed by many as a component of LLMs worth accelerating. During inference, tokenizers typically rely on simple dictionary lookups and are executed on CPUs as standard processes. This approach, however, introduces significant overhead from scheduling delays, core selection, data copying, and other system-level costs. These inefficiencies become problematic in latency-sensitive applications such as embedding, small language models, and agentic AI. In this paper, we present the Pinned Tokenizer (PinTok), a novel tokenizer architecture that reduces redundant hardware, operating system, and networking overhead through three key innovations: core and memory pinning, scheduling and context switch avoidance, and duplicate network packet copy and processing avoidance. Our implementation of PinTok can serve as a drop-in replacement for existing tokenizer deployments, delivering latency reductions of up to 95% (average), 97% (P50), 94% (P90), and 87% (P99) along with throughput improvements of up to 2,084%.

Offline meta-reinforcement learning leverages static datasets to enable agents to generalize to unseen environments by combining offline efficiency with meta-learning adaptability, yet it faces key challenges from context and policy distribution shifts. These issues hinder agents from adapting to online environments, and are further exacerbated under sparse-reward settings. As a result, agents often become trapped in an inherent pattern dilemma, failing to achieve robust generalization. In this work, we propose a novel framework that integrates information-theoretic task representation learning with a Transformer-based stochastic world model. Our approach extracts task-defining latent variables that are invariant to behavior policy, thereby effectively mitigating the context distribution shift. To further handle policy shift and model exploitation, we apply a conservative value penalty to imagination-based rollouts, preventing the policy from exploiting model inaccuracies while maintaining robust adaptation. Extensive evaluations demonstrate that our method outperforms state-of-the-art approaches, with superior stability and generalization under out-of-distribution and sparse-reward settings.

Sign-based and LMO-inspired optimizers have recently attracted substantial attention in deep learning due to their strong performance and low memory footprint. However, their fixed-magnitude updates can hurt terminal convergence: they decouple update mechanisms from gradient magnitudes and fail to account for parameter heterogeneity, often leading to oscillation rather than convergence. We propose SoftSignum, a smooth relaxation of sign-based optimization that replaces the hard sign map with a temperature-controlled soft-sign transformation, enabling a parameter-wise transition from sign-like updates to magnitude-sensitive SGD-like steps. We complement it with an adaptive quantile-based temperature schedule and extend the same principle to matrix-valued optimizers, obtaining SoftMuon. We also develop a generalized geometry-relaxation framework based on strongly convex regularizers and Fenchel conjugates, proving convergence in stochastic non-convex setting. Experiments on diverse deep learning tasks, including LLM pretraining, show that SoftSignum and SoftMuon consistently improve over their hard sign-based counterparts and standard AdamW.

Brain network classification is pivotal for diagnosing neurological disorders, yet identifying interpretable functional biomarkers fundamentally relies on precise parcellation. Unfortunately, conventional deep graph encoders applied to brain networks suffer from a critical theoretical limitation termed Spectral Unreachability. Through graph spectral analysis, this paper demonstrates that standard coupled encoder-pooling architectures inevitably oversmooth node representations, corrupting the high-frequency topological signals strictly required to delineate sharp module boundaries. To provide a structural remedy, the Hierarchical Spectral Parcellation Network (HiSP-Net) is proposed, which structurally decouples partition learning from feature smoothing via a project-then-align paradigm. Specifically, HiSP-Net maps representations directly into a partition space using a topology-agnostic projection block to preserve all-frequency details, while a Topology-Aware Alignment regularizer subsequently enforces spatial coherence. Extensive evaluations demonstrate that HiSP-Net consistently outperforms state-of-the-art baselines in classification, while successfully extracting stable and structurally coherent functional biomarkers. Source code is available at https://github.com/Kevin-916/HiSP-Net-demo/.