ICML 2026 Papers

Diffusion Language Models offer a compelling alternative to autoregressive (AR) models by enabling parallel and controllable generation. Within this family, Masked Diffusion Models (MDMs) currently perform best but still underperform AR models in perplexity and lack key inference-time efficiency features, most notably KV caching. We introduce Esoteric Language Models (Eso-LMs), a new family of models that fuses AR and MDM paradigms, smoothly interpolating between their perplexities while overcoming their respective limitations. Unlike prior work, which uses transformers with bidirectional attention as MDM denoisers, we exploit the connection between MDMs and Any-Order autoregressive models and adopt causal attention. This design lets us (1) **compute the exact likelihood of MDMs for the first time** and, crucially, (2) **allows exact KV caching for MDMs** while preserving parallel generation over the full sequence length for the first time, significantly improving inference efficiency. Combined with an optimized sampling schedule, Eso-LMs establish a new state of the art on the speed-quality Pareto frontier for unconditional generation.

Retrieval-augmented generation (RAG) systems suffer from a fundamental functional misalignment where retrievers optimize for semantic relevance, often recalling documents with high background utility but factually erroneous answer spans that generators blindly adopt as cognitive shortcuts. To resolve this, we propose the collaborative Critic-Reasoner framework that shifts robustness control from coarse-grained filtering to fine-grained cognitive decoupling. We disentangle the generation process into two serialized roles by deploying a Critic to perform surgical evidence purification through identifying and masking misleading entities while preserving supportive background context, followed by a Reasoner that switches from rote extraction to deductive reasoning based on the residual evidence. We operationalize this framework via a two-stage alignment strategy combining supervised fine-tuning (SFT) with path-aware direct preference optimization (DPO) to enforce strict behavioral synergy. Experimental results on adversarial benchmarks such as ConFiQA demonstrate that our method significantly outperforms baselines, achieving a 25.99\% accuracy gain in conflicting scenarios and effectively resolving the trust bias dilemma in real-world RAG.

Post-training adaptation has become the central paradigm for leveraging large language models (LLMs) in recommendation. While recent preference optimization methods, such as Direct Preference Optimization (DPO), enhance pairwise preference discrimination, they remain vulnerable to \emph{reward hacking}: models exploit imperfections in reward signals, leading to inflated training metrics without genuine recommendation gains. We analyze this issue from a gradient perspective and formalize the concept of the \emph{$\varepsilon$-insensitive region}, where pairwise updates exert little influence on the ordering between positives and unsampled negatives. Under the Bradley–Terry model, we further show that these regions can occupy a substantial fraction of the preference space, inevitably leading to misaligned rankings. To address this issue, we propose Simulated Preference Optimization for Reward-hacking mitigation using Pseudo-negatives (SIRIUS). Our framework introduces pseudo-negative samples to enrich contrastive signals and reduce the prevalence of $\varepsilon$-insensitive regions. Extensive experiments on three public benchmarks show that \our{} consistently improves ranking quality and effectively mitigates reward hacking, providing both theoretical and practical insights for advancing LLM-based recommendation. Our code is available at \url{https://anonymous.4open.science/r/C557-id}

State Space Models (SSMs) have emerged as a powerful and efficient alternative to Transformers, demonstrating linear-time complexity and exceptional sequence modeling capabilities. However, their application to vision tasks remains challenging. First, existing vision SSMs largely depend on manually designed fixed scanning methods to flatten image patches into sequences, which imposes predefined geometric structures and increases the complexity. Second, the broader adoption of vision SSMs is hindered in domains that require query-based interactions between distinct information streams. This is a result of the inherently causal and self-referential nature of SSMs designed for 1D sequence modeling tasks. This fusion mechanism is indispensable for critical perception tasks such as multi-view 3D fusion. To address these limitations, we propose Deformba, a context adaptive method that dynamically augments the spatial structural information while maintaining the linear complexity of SSMs. Deformba also allows multi-modal fusion, analogous to standard cross attention. To demonstrate the effectiveness and general applicability of Deformba, we test its performance on general 2D vision tasks such as image classification, object detection, instance segmentation, and semantic segmentation, as well as 3D vision tasks like BEV perception. Extensive experiments show that Deformba achieves strong performance across various visual perception benchmarks.

Neighborhood search operators are critical to the performance of Multi-Objective Evolutionary Algorithms (MOEAs) and rely heavily on expert design. Although recent LLM-based Automated Heuristic Design (AHD) methods have made notable progress, they primarily optimize individual heuristics or components independently, lacking explicit exploration and exploitation of dynamic coupling relationships between operators. In this paper, multi-operator optimization in MOEAs is formulated as a Markov decision process, enabling the improvement of interdependent operators through sequential decision-making. To address this, we propose the Evolution of Operator Combination (E2OC) framework for MOEAs, which achieves the co-evolution of design strategies and executable codes. E2OC employs Monte Carlo Tree Search to progressively search combinations of operator design strategies and adopts an operator rotation mechanism to identify effective operator configurations while supporting the integration of mainstream AHD methods as the underlying designer. Experimental results across AHD tasks with varying objectives and problem scales show that E2OC consistently outperforms state-of-the-art AHD and other multi-heuristic co-design frameworks, demonstrating strong generalization and sustained optimization capability.

Experience learning has achieved promising results in enhancing LLM agent planning and reasoning by integrating past interactions as reusable knowledge. However, existing methods remain confined to explicit text space—retrieving experiences via semantic similarity and concatenating them into the context window, leading to substantial token overhead and a decoupled architecture that separates retrieval from generation. To address these limitations, we propose ExpWeaver, a framework that enables LLM agents to learn from experience via latent retrieval-augmented generation, without requiring a separate RAG module. ExpWeaver encodes experiences using the LLM’s own hidden states, retrieves relevant experiences directly in latent space at each decoding step, and integrates them through cross-attention aggregation and gated residual mechanisms. The entire pipeline is optimized end-to-end with reinforcement learning, supporting both generative and ranking tasks. We evaluate ExpWeaver on 13 diverse tasks spanning question answering, reasoning, coding, scientific prediction, and recommendation. Results demonstrate that: (1) ExpWeaver achieves state-of-the-art on 12 out of 13 tasks, outperforming the strongest baseline by over 6.8%; (2) ExpWeaver maintains token efficiency comparable to non-retrieval baselines while text-based retrieval methods require 1.5–2× more tokens; and (3) ExpWeaver exhibits superior cross-domain generalization, outperforming the strongest baseline by 16.32% under zero-shot transfer and 15.21% under few-shot transfer. Our code for ExpWeaver is released at https://github.com/ulab-uiuc/ExpWeaver.

Transferable jailbreaking attacks enable red-teaming of black-box large language models by optimizing adversarial prompts on open-source surrogates. A natural approach to improve transferability is multi-model training---optimizing against multiple source models simultaneously. Yet this approach has been largely abandoned, as it yields only marginal gains with standard optimizers. We argue the root cause is poor generalization: standard gradient descent lacks stability when aggregating gradients from diverse models. Since GCG and its variants implicitly perform SGD in discrete token space, they inherit this instability in multi-model settings. We address this with Lookahead-GCG, which combines: (1) Stochastic Nesterov Accelerated Gradient (SNAG), whose lookahead mechanism reduces sensitivity to individual gradient updates, providing stability for multi-model optimization; (2) embedding-space momentum accumulation, which enables SNAG in discrete token optimization; and (3) maximally distant initialization, which exploits SNAG's improved generalization by starting from a universally beneficial point. Experiments show our method achieves 50.37% ASR on open-source and 34.03\% on closed-source LLMs, outperforming GCG and TransferAttack with +11.78% gains from multi-model optimization.

Diffusion Transformers (DiT) have demonstrated remarkable generative capabilities but remain highly computationally expensive. Previous acceleration methods, such as pruning and distillation, typically rely on a fixed computational capacity, leading to insufficient acceleration and degraded generation quality. To address this limitation, we propose \textbf{Elastic Diffusion Transformer (E-DiT)}, an adaptive acceleration framework for DiT that effectively improves efficiency while maintaining generation quality. Specifically, we observe that the generative process of DiT exhibits substantial sparsity (i.e., some computations can be skipped with minimal impact on quality), and this sparsity varies significantly across samples. Motivated by this observation, E-DiT equips each DiT block with a lightweight router that dynamically identifies sample-dependent sparsity from the input latent. Each router adaptively determines whether the corresponding block can be skipped. If the block is not skipped, the router then predicts the optimal MLP width reduction ratio within the block. During inference, we further introduce a block-level feature caching mechanism that leverages router predictions to eliminate redundant computations in a training-free manner. Extensive experiments across 2D image (Qwen-Image and FLUX) and 3D asset (Hunyuan3D-3.0) demonstrate the effectiveness of E-DiT, achieving up to $\sim$2$\times$ speedup with negligible loss in generation quality.

We propose to model text generation as a continuous-time latent dynamical process, where token generation is formulated as a Markov decision process whose internal state evolves via a neural ODE. This formulation bridges discrete token sequences and continuous semantic evolution, providing a theoretically grounded framework for text generation with continuous-time latent states. The framework is optimized via reinforcement learning, maximizing a composite objective that integrates cumulative rewards with a Kullback–Leibler divergence regularization term from a pre-trained language model. Both theoretical and empirical results demonstrate that our Continuous-Time Latent Language Model (CT-LLM) achieves superior effectiveness and efficiency in text generation, establishing a new paradigm for continuous-time language modeling.

Shapley additive explanations (SHAP) are widely recognised as computationally intractable for neural networks, since they induce an exponential search space over the input features. In this work, we take a first step towards scaling exact SHAP computation to larger search spaces by introducing an algorithm that leverages recent advances in neural network verification to compute arbitrarily tight exact lower and upper bounds on SHAP values for neural networks, ultimately recovering the exact SHAP values. We demonstrate that our approach scales to orders of magnitude larger search spaces than state-of-the-art exact methods. This provides an important first step towards exact SHAP computation and establishes a principled cornerstone for evaluating statistical approximation methods on larger search spaces.

Continual Pre-Training (CPT) is essential for enabling Language Models (LMs) to integrate new factual knowledge without erasing old. While classical CPT techniques like data replay have become the standard paradigm, the mechanisms underlying how LMs acquire and retain facts over time, termed as continual Factual Knowledge Acquisition (cFKA), remain unclear. In this work, we present a theoretical framework that characterizes the training dynamics of cFKA using a single-layer Transformer with linear attention, offering a unified explanation for the behavior of popular CPT methods. Our analysis reveals that regularization-based methods merely adjust the convergence rate of parameters without altering the inherent forgetting tendency, whereas data replay methods shift convergence dynamics and stabilize pretrained knowledge. Building on these insights, we propose a novel generative data replay approach, called Selecting Tokens via attentiOn Contribution (STOC), which identifies influential factual snippets to guide replay generation. Extensive experiments on both synthetic and real-world datasets validate our theoretical findings and demonstrate that STOC effectively enhances cFKA by mitigating catastrophic forgetting.

Label noise is pervasive in real-world datasets and significantly compromises model generalization, fueling extensive research into Learning with Noisy Labels (LNL). Most LNL methods focus on robust discriminative learning, while recent generative classifiers such as label diffusion models (LDMs) show superior robustness by modeling class posteriors. However, current LDMs predominantly rely on standard $\epsilon$-prediction, where Gaussian pnoise lacks explicit class semantics, limiting both optimization and inference under label noise environments. To address this issue, we propose just y-prediction (JYP), a novel training paradigm that enables LDMs to directly characterize the label manifold and leverage explicit class-semantic guidance. Theoretically, we prove that JYP converges to an optimal solution equivalent to that of $\epsilon$-prediction within the label diffusion framework, while facilitating accelerated convergence and enabling one-step inference. Leveraging JYP as a foundation, we further incorporate historical cumulative inconsistency to adaptively tailor optimization strategies for clean, noisy, and hard samples. Extensive experiments demonstrate that our method consistently outperforms competitors across diverse synthetic noisy datasets and achieves state-of-the-art performance on multiple real-world benchmarks.

Bayesian experimental design (BED) is a principled framework for data-efficient design of sequential experiments. However, existing BED methods are unable to adapt to dynamic constraints inherent in real-world tasks due to budget limitations, varying costs, or physical constraints that restrict how designs evolve over time. In this paper, we introduce a novel approach to BED that enables constrained optimization of experimental designs by combining offline pre-training of an amortized policy and a posterior network with online multi-step lookahead planning using scenario trees. We empirically demonstrate that our method yields substantially more informative design sequences than existing methods across a range of constrained BED tasks, while incurring only a modest additional computational overhead.

As language models are increasingly deployed for complex autonomous tasks, their ability to reason accurately over longer horizons becomes critical. An essential component of this ability is planning and managing a long, complex chain-of-thought (CoT). We introduce LongCoT, a scalable benchmark of 2,500 expert-designed problems spanning chemistry, mathematics, computer science, chess, and logic to isolate and directly measure the long-horizon CoT reasoning capabilities of frontier models. Problems consist of a short input with a verifiable answer; solving them requires navigating a graph of interdependent steps that span tens to hundreds of thousands of reasoning tokens. Each local step is individually tractable for frontier models, so failures reflect long-horizon reasoning limitations. At release, the best models achieve <10% accuracy (GPT 5.2: 9.8%; Gemini 3 Pro: 6.1%) on LongCoT, revealing a substantial gap in current capabilities. Overall, LongCoT provides a rigorous measure of long-horizon reasoning, tracking the ability of frontier models to reason reliably over extended periods.

Diffusion Transformers (DiTs) can synthesize high-fidelity images, but training at ultra-high resolutions is expensive, making inference-time extrapolation essential. Existing methods are typically \textit{scale-agnostic}, applying the same positional-modulation schedule regardless of target resolution. We show that this misses a scale-sensitive property of denoising: resizing shifts a fixed semantic pattern toward lower \emph{normalized} spatial frequencies while enlarging the spatial support over which global structure must be coordinated. The latter delays structural lock-in at high resolutions, so schedules may either relax guidance before global layout has stabilized, causing \textit{structural collapse}, or retain excessive intervention into late denoising, causing \textit{textural degradation}. We introduce \textbf{SigMa ($\sigma$)}, a training-free framework that uses Sigmoid Modulation for \textit{scale-adaptive} extrapolation through two scaling laws: \textit{Decoupled Geometric Center Alignment} and \textit{Iso-Variance Rate Adaptation}. Experiments show that SigMa reduces this mismatch, enabling training-free extrapolation up to 16 megapixels and achieving the best or competitive performance among the tested training-free extrapolation baselines. Code is available at \href{https://github.com/bxuanz/SigMa.git}{github.com/bxuanz/SigMa}.

Conditional generative models have recently achieved remarkable success in various applications. However, a suitable metric for evaluating the reliability of these models, which takes into account their inherent uncertainty, is still lacking. Existing metrics, which typically assess a single output, may fail to capture the variability or potential risks in generation. In this paper, we propose a novel evaluation metric called reliability score based on conformal prediction, which measures the worst-case performance within the prediction set at a pre-specified confidence level. However, computing this score is challenging due to the high-dimensional nature of the output space and the nonconvexity of both the metric function and the prediction set. To efficiently compute this score, we introduce Conformal ReLiability (CReL), a framework that can (i) construct the prediction set with desired coverage; and (ii) accurately optimize the reliability score within the constructed prediction set. We provide theoretical results on coverage and demonstrate empirically that our method produces more informative prediction sets than existing approaches. Experiments on synthetic data and the image-to-text and text-to-image tasks further demonstrate the interpretability of our new metric, and the validity and effectiveness of our computational framework. Source code can be found at https://ggc29.github.io/CReL/.

Advances in data collection are producing growing volumes of temporal count observations, making adapted modeling increasingly necessary. In this work, we introduce a generative framework for independent component analysis of temporal count data, combining regime-adaptive dynamics with Poisson log-normal emissions. The model identifies disentangled components with regime-dependent contributions, enabling representation learning and perturbations analysis. Notably, we establish the identifiability of the model, supporting principled interpretation. To learn the parameters, we propose an efficient amortized variational inference procedure. Experiments on simulated data evaluate recovery of the mixing function and latent sources across diverse settings, while real-world applications to gut microbiome and climate datasets reveal co-variation patterns and regime shifts consistent with domain-specific knowledge.

Decentralized Unlearning (DU) aims to remove the influence of specific clients from a collaboratively trained global model. However, existing methods suffer from strong reliance on static, problem-specific hyperparameters or restrictive convexity assumptions, limiting their general applicability. To overcome these limitations, we propose **TRA**jectory-aware **CE**rtified **D**ecentralized **U**nlearning (**TRACE-DU**), a generic unlearning framework for decentralized training. **TRACE-DU** introduces a fine-grained sensitivity analysis that leverages local SGD updates and decentralized training dynamics, thereby eliminating the need for convexity assumptions and reducing dependence on manually tuned parameters. By integrating strategic checkpoint selection with calibrated noise perturbation, the proposed framework enables efficient certified unlearning. Moreover, we exploit historical model trajectories to extend this framework, enabling it to naturally support sequential unlearning requests from an arbitrary number of clients. We provide theoretical guarantees for certified unlearning and derive sensitivity bounds under both convex and non-convex loss functions. Experimental results demonstrate that our framework outperforms state-of-the-art baselines across diverse metrics.

Deep learning surrogates for 3D Partial Differential Equations (PDEs) often fail to generalize across geometric transformations because they depend heavily on specific coordinate systems. While equivariant networks offer a solution, they typically rely on local operations in the spatial domain, making the global receptive field—essential for PDE dynamics—computationally expensive. Conversely, Fourier Neural Operators (FNOs) efficiently capture global interactions, yet establishing 3D equivariance within them remains impractical due to the prohibitive cost of spectral group convolutions. To bridge this gap, we introduce EqGINO, a geometrically robust framework that enforces isotropy in the spectral domain. By design, EqGINO guarantees exact equivariance to the discrete symmetries inherent to the discretized computational domain. Beyond this discrete guarantee, our structural prior enables effective generalization to arbitrary continuous orientations even with a limited number of SE(3)-transformed training samples. Consequently, our method robustly models coordinate-invariant physical laws on complex irregular 3D geometries. Our code is available at https://github.com/sung-won-kim/EqGINO