ICLR 2026 Papers

Large Language Models (LLMs) have achieved remarkable success through imitation learning on vast text corpora, but this paradigm creates a training-generation gap and limits robust reasoning. Reinforcement learning (RL) offers a more data-efficient solution capable of bridging this gap, yet its application has been constrained by a critical data bottleneck: existing RL datasets are orders of magnitude smaller and less diverse than web-scale pre-training corpora. To address this, we introduce the \textbf{\texttt{Webscale-RL} pipeline}, a scalable data engine that systematically converts large-scale pre-training documents into millions of diverse, verifiable question-answer pairs for RL. Using this pipeline, we construct the \textbf{\texttt{Webscale-RL} dataset}, containing 1.2 million examples across more than 9 domains. Our experiments show that the model trained on this dataset significantly outperforms continual pretraining and strong data refinement baselines across a suite of benchmarks. Notably, RL training with our dataset proves substantially more efficient, achieving the performance of continual pre-training with up to 100$\times$ fewer tokens. Our work presents a viable path toward scaling RL to pre-training levels, enabling more capable and efficient language models.

Computing next-token likelihood ratios between two language models (LMs) is a standard task in training paradigms such as knowledge distillation. Since this requires both models to share the same probability space, it becomes challenging when the teacher and student LMs use different tokenizers, for instance, when edge-device deployment necessitates a smaller vocabulary size to lower memory overhead. This work addresses this vocabulary misalignment problem by uncovering an implicit recursive structure in the commonly deployed Byte-Pair Encoding (BPE) algorithm and utilizing it to create a probabilistic framework for \textit{cross-tokenizer likelihood scoring}. Our method enables sequence likelihood evaluation for vocabularies different from the teacher model native tokenizer, addressing two specific scenarios: when the student vocabulary is a subset of the teacher vocabulary, and the general case where it is arbitrary. In the subset regime, our framework computes exact likelihoods and provides next-token probabilities for sequential sampling with only $\mathcal{O}(1)$ model evaluations per token. When used for distillation, this yields up to a 12% reduction in memory footprint for the Qwen2.5-1.5B model while also improving baseline performance up to 4\% on the evaluated tasks. For the general case, we introduce a rigorous lossless procedure that leverages BPE recursive structure, complemented by a fast approximation that keeps large-vocabulary settings practical. Applied to GSM8K mathematical reasoning distillation, our method improves accuracy by over 2% the current state of the art. Code: https://github.com/truongbuu/cross-tokenizer-scoring

The possibility of LLM self-awareness and even sentience is gaining increasing public attention and has major safety and policy implications, but the science of measuring them is still in a nascent state. Here we introduce a novel methodology for quantitatively evaluating metacognitive abilities in LLMs. Taking inspiration from research on metacognition in nonhuman animals, our approach eschews model self-reports and instead tests to what degree models can strategically deploy knowledge of internal states. Using two experimental paradigms, we demonstrate that frontier LLMs introduced since early 2024 show increasingly strong evidence of certain metacognitive abilities, specifically the ability to assess and utilize their own confidence in their ability to answer factual and reasoning questions correctly and the ability to anticipate what answers they would give and utilize that information appropriately. We buttress these behavioral findings with an analysis of the token probabilities returned by the models, which suggests the presence of an upstream internal signal that could provide the basis for metacognition. We further find that these abilities 1) are limited in resolution, 2) emerge in context-dependent manners, and 3) seem to be qualitatively different from those of humans. We also report intriguing differences across models of similar capabilities, suggesting that LLM post-training may have a role in developing metacognitive abilities.

Density ratio estimation (DRE) is fundamental in machine learning and statistics, with applications in domain adaptation and two-sample testing. However, DRE methods are highly sensitive to hyperparameter selection, with suboptimal choices often resulting in poor convergence rates and empirical performance. To address this issue, we propose a novel model aggregation algorithm for DRE that trains multiple models with different hyperparameter settings and aggregates them. Our aggregation provably achieves minimax-optimal error convergence without requiring prior knowledge of the smoothness of the unknown density ratio. Our method surpasses cross-validation-based model selection and model averaging baselines for DRE on standard benchmarks for DRE and large-scale domain adaptation tasks, setting a new state of the art on image and text data.

A fine-grained data recipe is crucial for pre-training large language models (LLMs), as it can significantly enhance training efficiency and model performance. One important ingredient in the recipe is to select samples based on scores produced by defined rules, LLM judgment, or statistical information in embeddings, which can be roughly categorized into quality and diversity metrics. Due to the high computational cost when applied to trillion-scale token pre-training datasets such as FineWeb and DCLM, these two or more types of metrics are rarely considered jointly in a single selection process. However, in our empirical study, selecting samples based on quality metrics exhibit severe diminishing returns during long-term pre-training, while selecting on diversity metrics removes too many valuable high-quality samples, both of which limit pre-trained LLMs' capabilities. Therefore, we introduce DATAMASK, a novel and efficient joint learning framework designed for large-scale pre-training data selection that can simultaneously optimize multiple types of metrics in a unified process, with this study focusing specifically on quality and diversity metrics. DATAMASK approaches the selection process as a mask learning problem, involving iterative sampling of data masks, computation of policy gradients based on predefined objectives with sampled masks, and updating of mask sampling logits. Through policy gradient-based optimization and various acceleration enhancements, it significantly reduces selection time by 98.9% compared to greedy algorithm, enabling our study to explore joint learning within trillion-scale tokens. With DATAMASK, we select a subset of about 10% from the 15 trillion-token FineWeb dataset, termed FineWeb-Mask. Evaluated across 12 diverse tasks, this high-quality and diverse subset achieves significant improvements of 3.2% on a 1.5B dense model and 1.9% on a 7B MoE model after pre-training with hundreds of billions of tokens, demonstrating its effectiveness.

Building efficient and effective generative models for neural network weights has been a research focus of significant interest that faces challenges posed by the high-dimensional weight spaces of modern neural networks and their symmetries. Several prior generative models are limited to generating partial neural network weights, particularly for larger models, such as ResNet and ViT. Those that do generate complete weights struggle with generation speed or require finetuning of the generated models. In this work, we present \ours, a Flow Matching model that operates directly in weight space to generate diverse and high-accuracy neural network weights for a variety of architectures, neural network sizes, and data modalities. The neural networks generated by \ours do not require fine-tuning to perform well and can scale to large networks. We apply Git Re-Basin and TransFusion for neural network canonicalization in the context of generative weight models to account for the impact of neural network permutation symmetries and to improve generation efficiency for larger model sizes. The generated networks excel at transfer learning, and ensembles of hundreds of neural networks can be generated in minutes, far exceeding the efficiency of diffusion-based methods. DeepWeightFlow models pave the way for more efficient and scalable generation of diverse sets of neural networks.

Protecting the intellectual property of large language models (LLMs) is crucial, given the substantial resources required for their training. Consequently, there is an urgent need for both model owners and third parties to determine whether a suspect LLM is trained from scratch or derived from an existing base model. However, the intensive post-training processes that models typically undergo—such as supervised fine-tuning, extensive continued pretraining, reinforcement learning, multi-modal extension, pruning, and upcycling—pose significant challenges to reliable identification. In this work, we propose a training-free fingerprinting method based on weight matrices. We leverage the Linear Assignment Problem (LAP) and an unbiased Centered Kernel Alignment (CKA) similarity to neutralize the effects of parameter manipulations, yielding a highly robust and high-fidelity similarity metric. On a comprehensive testbed of 60 positive and 90 negative model pairs, our method demonstrates exceptional robustness against all six aforementioned post-training categories while exhibiting a near-zero risk of false positives. By achieving perfect scores on all classification metrics, our approach establishes a strong basis for reliable model lineage verification. Moreover, the entire computation completes within 30s on an NVIDIA 3090 GPU.

Chain-of-Thought (CoT) reasoning has advanced the capabilities and transparency of language models (LMs); however, reasoning chains can contain inaccurate statements that reduce performance and trustworthiness. To address this, we propose to augment each reasoning step in a CoT with a latent veracity (or correctness) variable. To efficiently explore this expanded space, we introduce Veracity Search (VS), a discrete search algorithm over veracity assignments. It performs otherwise intractable inference in the posterior distribution over latent veracity values by leveraging the LM's joint likelihood over veracity and the final answer as a proxy reward. This efficient inference-time verification method facilitates supervised fine-tuning of an Amortized Veracity Inference (AVI) machine by providing pseudo-labels for veracity. AVI generalizes VS, enabling accurate zero-shot veracity inference in novel contexts. Empirical results demonstrate that VS reliably identifies errors in logical (ProntoQA), mathematical (GSM8K), and commonsense (CommonsenseQA) reasoning benchmarks, with AVI achieving comparable zero-shot accuracy. Finally, we demonstrate the utility of latent veracity inference for providing feedback during self-correction and self-improvement.

Large-scale text-to-image diffusion models, while powerful, suffer from prohibitive computational cost. Existing one-shot network pruning methods can hardly be directly applied to them due to the iterative denoising nature of diffusion models. To bridge the gap, this paper presents \textit{OBS-Diff}, a novel one-shot pruning framework that enables accurate and training-free compression of large-scale text-to-image diffusion models. Specifically, (i) OBS-Diff revitalizes the classic Optimal Brain Surgeon (OBS), adapting it to the complex architectures of modern diffusion models and supporting diverse pruning granularity, including unstructured, N:M semi-structured, and structured (MHA heads and FFN neurons) sparsity; (ii) To align the pruning criteria with the iterative dynamics of the diffusion process, by examining the problem from an error-accumulation perspective, we propose a novel timestep-aware Hessian construction that incorporates a logarithmic-decrease weighting scheme, assigning greater importance to earlier timesteps to mitigate potential error accumulation; (iii) Furthermore, a computationally efficient group-wise sequential pruning strategy is proposed to amortize the expensive calibration process. Extensive experiments show that OBS-Diff achieves state-of-the-art one-shot pruning for diffusion models, delivering inference acceleration with minimal degradation in visual quality.

The use of low-bit quantization has emerged as an indispensable technique for enabling the efficient training of large-scale models. Despite its widespread empirical success, a rigorous theoretical understanding of its impact on learning performance remains notably absent, even in the simplest linear regression setting. We present the first systematic theoretical study of this fundamental question, analyzing finite-step stochastic gradient descent (SGD) for high-dimensional linear regression under a comprehensive range of quantization targets: data, label, parameter, activation, and gradient. Our novel analytical framework establishes precise algorithm-dependent and data-dependent excess risk bounds that characterize how different quantization affects learning: parameter, activation, and gradient quantization amplify noise during training; data quantization distorts the data spectrum and introduces additional approximation error. Crucially, we distinguish the effects of two quantization schemes: we prove that for additive quantization (with constant quantization steps), the noise amplification benefits from a suppression effect scaled by the batch size, while multiplicative quantization (with input-dependent quantization steps) largely preserves the spectral structure, thereby reducing the spectral distortion. Furthermore, under common polynomial-decay data spectra, we quantitatively compare the risks of multiplicative and additive quantization, drawing a parallel to the comparison between FP and integer quantization methods. Our theory provides a powerful lens to characterize how quantization shapes the learning dynamics of optimization algorithms, paving the way to further explore learning theory under practical hardware constraints.

Spectral Graph Neural Networks (GNNs) achieve strong performance in node classification, yet their node distinguishability remains poorly understood. We analyze how graph matrices and node features jointly influence node distinguishability. Further, we derive a theoretical lower bound on the number of distinguishable nodes, which is governed by two key factors: distinct eigenvalues in the graph matrix and nonzero frequency components of node features in the eigenbasis. Based on these insights, we propose AdaSpec, an adaptive graph matrix generation module that enhances node distinguishability of spectral GNNs without increasing the order of computational complexity. We prove that AdaSpec preserves permutation equivariance, ensuring that reordering the graph nodes results in a corresponding reordering of the node embeddings. Experiments across eighteen benchmark datasets validate AdaSpec's effectiveness in improving node distinguishability of spectral GNNs.

Combinatorial generalization (CG)—generalizing to unseen combinations of known semantic factors—remains a grand challenge in machine learning. While symmetry-based methods are promising, they learn from observed data and thus fail at what we term $\textbf{symmetry generalization}$: extending learned symmetries to novel data. We tackle this by proposing a novel framework that endows the latent space with the structure of a $\textbf{symmetric space}$, a class of manifolds whose geometric properties provide a principled way to extend these symmetries. Our method operates in two steps: first, it imposes this structure by learning the underlying algebraic properties via the $\textbf{Cartan decomposition}$ of a learnable Lie algebra. Second, it uses $\textbf{geodesic symmetry}$ as a powerful self-supervisory signal to ensure this learned structure extrapolates from observed samples to unseen ones. A detailed analysis on a synthetic dataset validates our geometric claims, and experiments on standard CG benchmarks show our method significantly outperforms existing approaches.

Recent advances in large-scale 3D scene reconstruction using unmanned aerial vehicles (UAVs) have spurred increasing interest in neural rendering techniques. However, existing approaches with conventional cameras struggle to capture consistent multi-view images of scenes, particularly in extremely blurred and low-light environments, due to the inherent limitations in dynamic range caused by long exposure and motion blur resulting from camera motion. As a promising solution, bio-inspired event cameras exhibit robustness in extreme scenarios, thanks to their high dynamic range and microsecond-level temporal resolution. Nevertheless, dedicated event datasets specifically tailored for large-scale UAV 3D scene reconstruction remain limited. To bridge this gap, we introduce \textbf{SkyEvents}, a pioneering large-scale event-enhanced UAV dataset for 3D scene reconstruction, incorporating RGB, event, and LiDAR data. SkyEvents encompasses 22 sequences, spanning over 8 hours of video, captured across a diverse set of illumination conditions, scenarios, and flight altitudes. To facilitate the event-based 3D scene reconstruction with SkyEvents, we propose the Geometry-constrained Timestamp Alignment (\textbf{GTA}) module to align timestamps between the event and RGB cameras. Furthermore, we introduce Region-wise Event Rendering (\textbf{RER}) loss for supervising the rendering optimization. With SkyEvents, we aim to motivate and equip researchers to advance large-scale 3D scene reconstruction in challenging environments, harnessing the unique strengths of event cameras. Dataset and code will be available at \url{https://}.

While diffusion language models (DLMs) offer a promising alternative to autoregressive models (ARs), existing open-source DLMs suffer from high inference latency. This bottleneck is mainly due to the attention’s quadratic complexity with respect to context length in computing all query–key pairs. Intuitively, to reduce this complexity, restricting computation to sparse attention patterns that retain only the most important ones offers an effective solution. This type of method is widely used in ARs, where the attention mechanism exhibits clear and fixed sparse patterns. In DLMs, our analysis also reveals the presence of sparse patterns and further highlights three unique observations: (1) attention patterns vary across heads, (2) attention patterns in each head remain highly similar across denoising steps, and (3) early denoising steps are critical for generation. These unique findings render well-studied fixed sparse attention methods in ARs largely incompatible with DLMs, as their fixed patterns fail to capture head-specific patterns in DLMs, and sparse attention applied in the early steps can lead to degradation in generation. To address these challenges, we propose **SparseD**, a novel sparse attention method for DLMs. Leveraging the observations in DLMs, SparseD only pre-computes and selects the most important query–key pairs once as head-specific sparse patterns for reusing across denoising steps. This manner can handle head-specific patterns without incurring the high latency associated with recomputing sparse patterns at each denoising step. Meanwhile, SparseD skips sparse attention and uses full attention in the early steps to preserve generation quality. Together, these establish SparseD as a practical and efficient solution for deploying DLMs in long-context applications. Experimental results demonstrate that SparseD achieves lossless acceleration, delivering up to $1.50\times$ speedup over FlashAttention at a 64k context length with 1,024 denoising steps. Anonymous code is available at https://anonymous.4open.science/r/SparseD-8C76/.

Kolmogorov–Arnold Networks (KANs) are a recently introduced neural architecture that replace fixed nonlinearities with trainable activation functions, offering enhanced flexibility and interpretability. While KANs have been applied successfully across scientific and machine learning tasks, their initialization strategies remain largely unexplored. In this work, we study initialization schemes for spline-based KANs, proposing two theory-driven approaches inspired by LeCun and Glorot, as well as an empirical power-law family with tunable exponents. Our evaluation combines large-scale grid searches on function fitting and forward PDE benchmarks, an analysis of training dynamics through the lens of the Neural Tangent Kernel, and evaluations on a subset of the Feynman dataset. Our findings indicate that the Glorot-inspired initialization significantly outperforms the baseline in parameter-rich models, while power-law initialization achieves the strongest performance overall, both across tasks and for architectures of varying size. This work underscores initialization as a key factor in KAN performance and introduces practical strategies to improve it.

Diffusion-based planners have shown great promise for autonomous driving due to their ability to capture multi-modal driving behaviors. However, guiding these models effectively in reactive, closed-loop environments remains a significant challenge. Simple conditioning often fails to provide sufficient guidance in complex and dynamic driving scenarios. Recent work attempts to use typical expert driving behaviors (i.e., anchors) to guide diffusion models but relies on a truncated schedule, which introduces theoretical inconsistencies and can compromise performance. To address this, we introduce BridgeDrive, a novel anchor-guided diffusion bridge policy for closed-loop trajectory planning. Our approach provides a principled diffusion framework that effectively translates anchors into fine-grained trajectory plans, appropriately responding to varying traffic conditions. Our planner is compatible with efficient ODE solvers, a critical factor for real-time autonomous driving deployment. We achieve state-of-the-art performance on the Bench2Drive benchmark, improving the success rate by 7.72\% over prior arts.

Despite the impressive capabilities of large language models across various tasks, their continued scaling is severely hampered not only by data scarcity but also by the performance degradation associated with excessive data repetition during training. To overcome this critical bottleneck, we introduce the Massive Genre-Audience (MGA) reformulation method, a framework designed to augment corpora in a way that supports more effective model performance scaling. Instead of relying on complex, predefined seed systems, MGA systematically reformulates existing corpora into diverse, contextually-rich variations by adaptively generating genre-audience pairs. We present this framework and the resulting 770 billion token MGACorpus, created as a practical instantiation of our methodology. We experimentally validate MGA's core benefits by demonstrating superior scaling properties, in terms of both model size and data budget, against data repetition and upsampling (up to 13B parameters). Furthermore, our comprehensive analysis investigates the role of synthesis principles in generation quality and reveals nuances in evaluating model capabilities using standard loss metrics. Our work shows that a systematic framework like MGA provides a reliable pathway to substantially augment training datasets, effectively alleviating repetition bottlenecks and enabling more efficient scaling of large language models.

With the increasing demand for interpretability in machine learning, functional ANOVA decomposition has gained renewed attention as a principled tool for breaking down high-dimensional function into low-dimensional components that reveal the contributions of different variable groups. Recently, Tensor Product Neural Network (TPNN) has been developed and applied as basis functions in the functional ANOVA model, referred to as ANOVA-TPNN. A disadvantage of ANOVA-TPNN, however, is that the components to be estimated must be specified in advance, which makes it difficult to incorporate higher-order TPNNs into the functional ANOVA model due to computational and memory constraints. In this work, we propose Bayesian-TPNN, a Bayesian inference procedure for the functional ANOVA model with TPNN basis functions, enabling the detection of higher-order components with reduced computational cost compared to ANOVA-TPNN. We develop an efficient MCMC algorithm and demonstrate that Bayesian-TPNN performs well by analyzing multiple benchmark datasets. Theoretically, we prove that the posterior of Bayesian-TPNN is consistent.

In traditional information retrieval, corpus items compete with each other to occupy top ranks in response to a query. In contrast, in many recent retrieval scenarios associated with complex, multi-hop question answering or text-to-SQL, items are not self-complete: they must instead collaborate, i.e., information from multiple items must be combined to respond to the query. In the context of modern dense retrieval, this need translates into finding a small collection of corpus items whose contextual word vectors collectively cover the contextual word vectors of the query. The central challenge is to retrieve a near-optimal collection of covering items in time that is sublinear in corpus size. By establishing coverage as a submodular objective, we enable successive dense index probes to quickly assemble an item collection that achieves near-optimal coverage. Successive query vectors are iteratively `edited', and the dense index is built using random projections of a novel, lifted dense vector space. Beyond rigorous theoretical guarantees, we report on a scalable implementation of this new form of vector database. Extensive experiments establish the empirical success of DISCo, in terms of the best coverage vs. query latency tradeoffs.