ICLR 2026 Papers

Behavior generation in autonomous driving aims to simulate dynamic driving scenarios from recorded driving logs. A popular approach is to apply next-token-prediction with discrete trajectory tokenization. In this work, we explore what makes a good trajectory tokenizer from the perspective of logged data usage. We first analyze the four properties (coverage, utilization, symmetry and robustness) of vocabularies of data-driven and rule-based trajectory tokenizers and their impact on performance and generalization. Data-driven tokenizers often build vocabularies with better utilization but suffer from insufficient coverage and sensitivity to noise, while rule-based methods have better coverage but contain too many useless tokens. With these insights, we propose TrajTok, a trajectory tokenizer that combines the two methods with rule-based vocabulary candidate setup and data-driven filtering and selection processes. The tokenizer has balanced coverage and utilization as well as good symmetry and robustness. Furthermore, we propose a spatial-aware label smoothing method for the cross-entropy loss to better model the similarities between the trajectory tokens. Our method wins first place in the 2025 Waymo Open Sim Agents Challenge.

The utilization of large language models (LLMs) has become increasingly widespread, and has attracted considerable attention. Although the emergence of discrete diffusion large language models (dLLMs) mitigates the inference latency inherent in autoregressive LLM decoding, its computational overhead remains substantial. To address this challenge, we propose Hierarchy-dLLM, a hierarchical decoding framework inspired by the divide-and-conquer principle. Our method recursively partitions masked spans into smaller sub-decoding areas and decodes tokens according to their confidence, which substantially increases the number of tokens generated per forward pass and improves information utilization. Extensive experiments conducted on multiple benchmarks demonstrate that Hierarchy-dLLM achieves accuracy comparable to or even surpassing existing baselines. Meanwhile, it is up to 17× faster than vanilla decoding and about 1.5× faster than the Fast-dLLM. These results establish hierarchical decoding as a practical solution for efficient dLLMs inference.

Drag-based image editing has long suffered from distortions in the target region, largely because the priors of earlier base models, Stable Diffusion, are insufficient to project optimized latents back onto the natural image manifold. With the shift from UNet-based DDPMs to more scalable DiT with flow matching (e.g., SD3.5, FLUX), generative priors have become significantly stronger, enabling advances across diverse editing tasks. However, drag-based editing has yet to benefit from these stronger priors. This work introduces DragFlow, the first framework to effectively harness FLUX’s rich prior via region-based supervision, enabling full use of its finer-grained, spatially precise features for drag-based editing and achieving substantial improvements over existing baselines. We first show that directly applying point-based drag editing to DiTs performs poorly: unlike the highly compressed features of UNets, DiT features are insufficiently structured to provide reliable guidance for point-wise motion supervision. To overcome this limitation, DragFlow introduces a region-based editing paradigm, where affine transformations enable richer and more consistent feature supervision. Additionally, we integrate pretrained open-domain personalization adapters (e.g., IP-Adapter) to enhance subject consistency, while preserving background fidelity through gradient mask-based hard constraints. Multimodal large language models (MLLMs) are further employed to resolve task ambiguities. For evaluation, we curate a novel Region-based Dragging benchmark (ReD Bench) featuring region-level dragging instructions. Extensive experiments on DragBench-DR and ReD Bench show that DragFlow surpasses both point-based and region-based baselines, setting a new state-of-the-art in drag-based image editing. Code and datasets will be publicly available upon publication.

This paper studies the computational challenges of large-scale attention-based models in artificial intelligence by introducing innovative sampling methods in the streaming setting. Inspired by the classical definition of the $\ell_2$ sampler and the recent progress of the attention scheme in Large Language Models (LLMs), we propose the definition of the attention sampler. Our approach significantly reduces the computational burden of traditional attention mechanisms. We demonstrate the effectiveness of the attention sampler from a theoretical perspective, including space and update time. Additionally, our framework exhibits scalability and broad applicability across various model architectures and domains.

Multi-Head Latent Attention (MLA) is a recent attention mechanism adopted in state-of-the-art LLMs such as DeepSeek-v3 and Kimi K2. Thanks to its novel formulation, MLA allows two functionally equivalent but computationally distinct kernel implementations: naive and absorb. While the naive kernels (e.g., FlashAttention) are typically preferred in training and prefill for their computational efficiency, existing decoding kernels (e.g., FlashMLA) rely on the absorb method to minimize HBM bandwidth usage. However, the compute-bound nature of the absorb implementations prohibits performance benefits from data reuse opportunities in attention calculations, such as shared prefixes. In this work, we introduce TyphoonMLA, a hybrid approach that combines naive and absorb formulations to harness the strengths of both. TyphoonMLA effectively leverages the shared prefix by applying the naive formulation to the compute-bound parts of attention calculations, while reducing the bandwidth requirements for non-shared parts by using the absorb formulation. As a result, TyphoonMLA improves the throughput of attention calculations in MLA architectures by up to 3× and 3.24× on NPU and GPUs, and boosts end-to-end throughput by up to 1.48× in tokens per second, with only a 3\% overhead in HBM size.

Algorithms with predictions leverage predictions for unknown inputs in online decision-making. These algorithms are analyzed by consistency, i.e., competitive ratio under perfect predictions, and robustness, i.e., competitive ratio under worst-case predictions. Smooth degrading performance with an increased prediction error is also desirable. This paper refines the notion of smoothness, a function of prediction error, defined as the competitive ratio over the problem instances where predictions are guaranteed to provide additional information. With our refined smoothness metric, we establish smoothness bounds for a few scheduling problems, including online total completion time minimization and makespan minimization. For a single machine to minimize the total completion time, we show a lower bound of $\eta$ and a $\eta^2$-smooth algorithm, where $\eta$ is the prediction error ($\eta \geq 1$); the bound holds for small errors. For parallel identical machines to minimize the makespan, we show a lower bound of $2 - O(\eta^{-2})$ and present an $O(\eta^2)$-smooth algorithm for small errors. Both bounds are tighter than the existing ones. For uniformly-related machines to minimize the makespan, we show a tight lower bound of $\lceil \log \eta \rceil$, matched by an $O(\log \eta)$-smooth algorithm.

The Maximum Common Edge Subgraph (MCES) problem is a crucial challenge with significant implications in domains such as biology and chemistry. Traditional approaches, which include transformations into max-clique and search-based algorithms, suffer from scalability issues when dealing with larger instances. This paper introduces ``Neural Graduated Assignment'' (NGA), a simple, scalable, unsupervised-training-based method that addresses these limitations. Central to NGA is stacking of differentiable assignment optimization with neural components, enabling high-dimensional parameterization of the matching process through a learnable temperature mechanism. We further theoretically analyze the learning dynamics of NGA, showing its design leads to fast convergence, better exploration-exploitation tradeoff, and ability to escape local optima. Extensive experiments across MCES computation, graph similarity estimation, and graph retrieval tasks reveal that NGA not only significantly improves computation time and scalability on large instances but also enhances performance compared to existing methodologies. The introduction of NGA marks a significant advancement in the computation of MCES and offers insights into other assignment problems. Code is open-sourced at https://anonymous.4open.science/r/NGA-10E3.

Large language models deployed in sensitive applications increasingly require the ability to unlearn specific knowledge, such as user requests, copyrighted materi- als, or outdated information, without retraining from scratch to ensure regulatory compliance, user privacy, and safety. This task, known as machine unlearning, aims to remove the influence of targeted data (forgetting) while maintaining per- formance on the remaining data (retention). A common approach is to formu- late this as a multi-objective problem and reduce it to a single-objective prob- lem via scalarization, where forgetting and retention losses are combined using a weighted sum. However, this often results in unstable training dynamics and degraded model utility due to conflicting gradient directions. To address these challenges, we propose OFMU, a penalty-based bi-level optimization framework that explicitly prioritizes forgetting while preserving retention through a hierar- chical structure. Our method enforces forgetting via an inner maximization step that incorporates a similarity-aware penalty to decorrelate the gradients of the for- get and retention objectives, and restores utility through an outer minimization step. To ensure scalability, we develop a two-loop algorithm with provable conver- gence guarantees under both convex and non-convex regimes. We further provide a rigorous theoretical analysis of convergence rates and show that our approach achieves better trade-offs between forgetting efficacy and model utility compared to prior methods. Extensive experiments across vision and language benchmarks demonstrate that OFMU consistently outperforms existing unlearning methods in both forgetting efficacy and retained utility.

Accurately estimating the Q-function is a central challenge in offline reinforcement learning. However, existing approaches often rely on a single global Q-function, which struggles to capture the compositional nature of tasks involving diverse subtasks. We propose In-context Compositional Q-Learning ($\texttt{ICQL}$), the first offline RL framework that formulates Q-learning as a contextual inference problem, using linear Transformers to adaptively infer local Q-functions from retrieved transitions without explicit subtask labels. Theoretically, we show that under two assumptions—linear approximability of the local Q-function and accurate weight inference from retrieved context—$\texttt{ICQL}$ achieves bounded Q-function approximation error, and supports near-optimal policy extraction. Empirically, $\texttt{ICQL}$ substantially improves performance in offline settings: improving performance in Kitchen tasks by up to 29. 46\%, and in Gym and Adroit tasks by up to 6\%. These results highlight the underexplored potential of in-context learning for robust and compositional value estimation, positioning $\texttt{ICQL}$ as a principled and effective framework for offline RL.

A recent breakthrough in nonconvex optimization is the online-to-nonconvex conversion framework of Cutkosky et al. (2023), which reformulates the task of finding an $\varepsilon$-first-order stationary point as an online learning problem. When both the gradient and the Hessian are Lipschitz continuous, instantiating this framework with two different online learners achieves a complexity of $ \mathcal{O}(\varepsilon^{-1.75}\log(1/\varepsilon)) $ in the deterministic case and a complexity of $ \mathcal{O}(\varepsilon^{-3.5}) $ in the stochastic case. However, this approach suffers from several limitations: (i) the deterministic method relies on a complex double-loop scheme that solves a fixed-point equation to construct hint vectors for an optimistic online learner, introducing an extra logarithmic factor; (ii) the stochastic method assumes a bounded second-order moment of the stochastic gradient, which is stronger than standard variance bounds; and (iii) different online learning algorithms are used in the two settings. In this paper, we address these issues by introducing an online optimistic gradient method based on a novel **doubly optimistic hint function**. Specifically, we use the gradient at an extrapolated point as the hint, motivated by two optimistic assumptions: that the difference between the hint and the target gradient remains near constant, and that consecutive update directions change slowly due to smoothness. Our method eliminates the need for a double loop and removes the logarithmic factor. Furthermore, by simply replacing full gradients with stochastic gradients and under the standard assumption that their variance is bounded by $\sigma^2$, we obtain a unified algorithm with complexity $\mathcal{O}(\varepsilon^{-1.75} + \sigma^2 \varepsilon^{-3.5})$, smoothly interpolating between the best-known deterministic rate and the optimal stochastic rate.

We present a novel framework for solving optimal transport (OT) problems based on the Hamilton--Jacobi (HJ) equation, whose viscosity solution uniquely characterizes the OT map. By leveraging the method of characteristics, we derive closed-form, bidirectional transport maps, thereby eliminating the need for numerical integration. The proposed method adopts a pure minimization framework: a single neural network is trained with a loss function derived from the method of characteristics of the HJ equation. This design guarantees convergence to the optimal map while eliminating adversarial training stages, thereby substantially reducing computational complexity. Furthermore, the framework naturally extends to a wide class of cost functions and supports class-conditional transport. Extensive experiments on diverse datasets demonstrate the accuracy, scalability, and efficiency of the proposed method, establishing it as a principled and versatile tool for OT applications with provable optimality.

In this paper, we address post-training quantization (PTQ) for large language models (LLMs) from an overlooked perspective: given a pre-trained high-precision LLM, the predominant sequential quantization framework treats different layers equally, but this may be not optimal in challenging bit-width settings. We empirically study the quantization impact of different layers on model accuracy, and observe that: (1) shallow/deep layers are usually more sensitive to quantization than intermediate layers; (2) among shallow/deep layers, the most sensitive one is the first/last layer, which exhibits significantly larger quantization error than others. These empirical observations imply that the quantization design for different layers of LLMs is required on multiple levels instead of a single level shared to all layers. Motivated by this, we propose a new PTQ framework termed **Sliding**-lay**er** **Quant**ization (SliderQuant) that relies on a simple adaptive sliding quantization concept facilitated by few learnable parameters. The base component of SliderQuant is called inter-layer sliding quantization, which incorporates three types of novel sliding window designs tailored for addressing the varying quantization sensitivity of shallow, intermediate and deep layers. The other component is called intra-layer sliding quantization that leverages an incremental strategy to quantize each window. As a result, SliderQuant has a strong ability to reduce quantization errors across layers. Extensive experiments on basic language generation, zero-shot commonsense reasoning and challenging math and code tasks with various LLMs, including Llama/Llama2/Llama3/Qwen2.5 model families, DeepSeek-R1 distilled models and large MoE models, show that our method outperforms existing PTQ methods (including the latest PTQ methods using rotation transformations) for both weight-only quantization and weight-activation quantization under diverse bit width settings. Code will be made publicly available.

Large Language Models (LLMs) exhibit impressive capabilities but often hallucinate, confidently providing incorrect answers instead of admitting ignorance. Prior work has shown that models encode linear representations of their own knowledge and that activation steering can reduce hallucinations. These approaches, however, require real-time monitoring and intervention during inference. We introduce Contrastive Activation Steering for Amortized Learning (CASAL), an efficient algorithm that connects interpretability with amortized optimization. CASAL directly bakes the benefits of activation steering into model's weights. Once trained, LLMs answer questions they know while abstaining from answering those they do not. CASAL's light-weight design requires training only a submodule of a single transformer layer and yet reduces hallucination by $\sim30\%$-$40 \%$ across multiple short-form QA benchmarks. CASAL is $\sim$30x more compute-efficient and $\sim$20x more data-efficient than strong LoRA-based baselines such as SFT and DPO, boosting its practical applicability in data scarce domains. Importantly, CASAL also generalizes effectively to out-of-distribution (OOD) domains. We showcase CASAL's flexibility in mitigating hallucinations in both text-only and vision-language models. To our knowledge, CASAL is the first steering-based training method that has been shown to be effective for both dense and Mixture-of-Experts (MoE) models. CASAL represents a promising step forward for applying interpretability-inspired method for practical deployment in production systems.

Diffusion models have recently emerged as powerful priors for solving inverse problems. While computed tomography (CT) is theoretically a linear inverse problem, it poses many practical challenges. These include correlated noise, artifact structures, reliance on system geometry, and misaligned value ranges, which make the direct application of diffusion models more difficult than in domains like natural image generation. To systematically evaluate how diffusion models perform in this context and compare them with established reconstruction methods, we introduce DM4CT, a comprehensive benchmark for CT reconstruction. DM4CT includes datasets from both medical and industrial domains with sparse-view and noisy configurations. To explore the challenges of deploying diffusion models in practice, we additionally acquire a high-resolution CT dataset at a high-energy synchrotron facility and evaluate all methods under real experimental conditions. We benchmark nine recent diffusion-based methods alongside seven strong baselines, including model-based, unsupervised, and supervised approaches. Our analysis provides detailed insights into the behavior, strengths, and limitations of diffusion models for CT reconstruction. The real-world dataset is publicly available at zenodo.org/records/15420527, and the codebase is open-sourced at github.com/DM4CT/DM4CT.

Despite recent advances in state space models (SSMs) such as Mamba across various sequence domains, research on their standalone capacity for time series classification (TSC) has remained limited. We propose MambaSL, a framework that minimally redesigns the selective SSM and projection layers of a single-layer Mamba, guided by four TSC-specific hypotheses. To address benchmarking limitations—restricted configurations, partial University of East Anglia (UEA) dataset coverage, and insufficiently reproducible setups—we re-evaluate 20 strong baselines across all 30 UEA datasets under a unified protocol. Our results show that MambaSL achieves state-of-the-art performance on the UEA benchmark among 21 models, with statistically significant average improvements over baselines while ensuring reproducibility via public checkpoints.

This paper investigates how multilayer perceptrons (MLPs) can be represented in non-Euclidean spaces, with emphasis on the Poincaré model of hyperbolic geometry. We aim to capture the geometric evolution of their weighted topology and self-organization over time. Instead of restricting analysis to single checkpoints, we construct temporal parameter-graphs across $T$ snapshots of the optimization process. This reflects the view that neural networks encode information not only in their weights but also in the trajectory traced during training. Drawing on the idea that many complex networks admit embeddings in hidden metric spaces where distances correspond to connection likelihood, we present a geometric and temporal graph-based meta learning framework for obtaining dynamic hyperbolic representations of the underlying neural parameter graphs. Our model embeds temporal parameter-graphs in the Poincaré ball and learns from them while maintaining equivariance to within-snapshot neuron permutations and invariance to permutations of past snapshots. In doing so, it preserves functional equivalence across time and recovers the network’s latent geometry. Experiments on regression and classification tasks with trained MLPs show that hyperbolic temporal representations expose how structure emerges during training, offering intrinsic explanations of self-organisation in a given model training environment.

Large language models (LLMs) sometimes memorize undesirable knowledge, which must be removed after deployment. Prior work on machine unlearning has focused largely on optimization methods that adjust parameters to enforce forgetting while preserving retention. However, these approaches assume that the forget and retain sets are readily available, which rarely holds in practice. Unlearning is typically triggered by an undesired generation at inference time, making the retrieval of relevant data the central challenge. We introduce the notion of \emph{data Pareto improvement} for LLM unlearning, which formalizes how retrieval can expand the achievable trade-off frontier between forgetting and retention. To realize this principle, we propose Randomized Antipodal Search on Linearized Influence Kernel (RASLIK), a retrieval algorithm that combines permutation–projection hashing with randomized antipodal search. RASLIK reduces selection variance, achieves sublinear complexity, and yields a double gain in both quality and efficiency. Across multiple models, datasets, and unlearning algorithms, RASLIK consistently outperforms deterministic baselines and even oracle sampling, establishing randomized search as a principled and scalable solution for data-centric unlearning.

Deep learning-based approaches, particularly graph neural networks (GNNs), have gained prominence in simulating flexible deformations and contacts of solids, due to their ability to handle unstructured physical fields and nonlinear regression on graph structures. However, existing GNNs commonly represent meshes with graphs built solely from vertices and edges. These approaches tend to overlook higher-dimensional spatial features, e.g. 2D facets and 3D cells, from the original geometry. As a result, it is challenging to accurately capture boundary representations and volumetric characteristics, though this information is critically important for modeling contact interactions and internal physical quantity propagation, particularly under sparse mesh discretization. In this paper, we introduce MAVEN, a mesh-aware volumetric encoding network for simulating 3D flexible deformation, which explicitly models geometric mesh elements of higher dimension to achieve a more accurate and natural physical simulation. MAVEN establishes learnable mappings among 3D cells, 2D facets, and vertices, enabling flexible mutual transformations. Explicit geometric features are incorporated into the model to alleviate the burden of implicitly learning geometric patterns. Experimental results show that MAVEN consistently achieves state-of-the-art performance across established datasets and a novel metal stretch-bending task featuring large deformations and prolonged contacts.

Error correcting codes (ECCs) are a fundamental technique for ensuring reliable communication over noisy channels. Recent advances in deep learning have enabled transformer-based decoders to achieve state-of-the-art performance on short codes; however, their computational complexity remains significantly higher than that of classical decoders due to the attention mechanism. To address this challenge, we propose EfficientMPT, an efficient message-passing transformer that significantly reduces computational complexity while preserving decoding performance. A key feature of EfficientMPT is the Efficient Error Correcting (EEC) attention mechanism, which replaces expensive matrix multiplications with lightweight vector-based element-wise operations. Unlike standard attention, EEC attention relies only on query-key interaction using global query vector, efficiently encode global contextual information for ECC decoding. Furthermore, EfficientMPT can serve as a foundation model, capable of decoding various code classes and long codes by fine-tuning. In particular, EfficientMPT achieves 85% and 91% of significant memory reduction and 47% and 57% of FLOPs reduction compared to ECCT for $(648,540)$ and $(1056,880)$ standard LDPC code, respectively.