ICLR 2026 Papers

While Masked Diffusion Language Models (MDLMs) relying on token masking and unmasking have shown promise in language modeling, their computational efficiency and generation flexibility remain constrained by the masking paradigm. In this paper, we propose Deletion-Insertion Diffusion language models (DID) that rigorously formulate token deletion and insertion as discrete diffusion processes, replacing the masking and unmasking processes in current MDLMs. DID improves training and inference efficiency by eliminating two major sources of computational overhead in MDLMs: the computations on non-informative 1) $\texttt{\<MASK\>}$ tokens inherent to its paradigm, and 2) $\texttt{\<PAD\>}$ tokens introduced in variable-length settings. Furthermore, DID offers greater flexibility by: 1) natively supporting variable-length sequences without requiring fixed-length padding, and 2) an intrinsic self-correction mechanism during generation due to insertion that dynamically adjusts token positions. To train DID, we design a score-based approach that assigns scores to token insertion operations and derive appropriate training objectives. The objectives involve subsequence counting problems, which we efficiently solve via a parallelized dynamic programming algorithm. Our experiments across fixed and variable-length settings demonstrate the advantage of DID over baselines of MDLMs and existing insertion-based LMs, in terms of modeling performance, sampling quality, and training/inference speed, without any hyperparameter tuning.

A meaningful text can be hidden inside another, completely different yet still coherent and plausible, text of the same length. For example, a tweet containing a harsh political critique could be embedded in a tweet that celebrates the same political leader, or an ordinary product review could conceal a secret manuscript. This uncanny state of affairs is now possible thanks to Large Language Models, and in this paper we present *Calgacus*, a simple and efficient protocol to achieve it. We show that even modest 8‑billion‑parameter open‑source LLMs are sufficient to obtain high‑quality results, and a message as long as this abstract can be encoded and decoded locally on a laptop in seconds. The existence of such a protocol demonstrates a radical decoupling of text from authorial intent, further eroding trust in written communication, already shaken by the rise of LLM chatbots. We illustrate this with a concrete scenario: a company could covertly deploy an unfiltered LLM by encoding its answers within the compliant responses of a safe model. This possibility raises urgent questions for AI safety and challenges our understanding of what it means for a Large Language Model to know something.

Recent advances in text-to-video generation have sparked interest in generative video editing tasks. Previous methods often rely on task-specific architectures (e.g., additional adapter modules) or dedicated customizations (e.g., DDIM inversion), which limit the integration of versatile editing conditions and the unification of various editing tasks. In this paper, we introduce UNified In-Context Video Editing (UNIC), a simple yet effective framework that unifies diverse video editing tasks within a single model in an in-context manner. To achieve this unification, we represent the inputs of various video editing tasks as three types of tokens: the source video tokens, the noisy video latent, and the multi-modal conditioning tokens that vary according to the specific editing task. Based on this formulation, our key insight is to integrate these three types into a single consecutive token sequence and jointly model them using the native attention operations of DiT, thereby eliminating the need for task-specific adapter designs. Nevertheless, direct task unification under this framework is challenging, leading to severe token collisions and task confusion due to the varying video lengths and diverse condition modalities across tasks. To address these, we introduce task-aware RoPE to facilitate consistent temporal positional encoding, and condition bias that enables the model to clearly differentiate different editing tasks. This allows our approach to adaptively perform different video editing tasks by referring the source video and varying condition tokens "in context", and support flexible task composition. To validate our method, we construct a unified video editing benchmark containing six representative video editing tasks. Results demonstrate that our unified approach achieves comparable performance with task specialists and exhibits emergent task composition abilities.

Multi-view evidential learning aims to integrate information from multiple views to improve prediction performance and provide trustworthy uncertainty estimation. Most previous methods assume that view-specific evidence learning is naturally reliable. However, in practice, the evidence learning process tends to be biased. Through empirical analysis on real-world data, we reveal that samples tend to be assigned more evidence to support data-rich classes, thereby leading to unreliable uncertainty estimation in predictions. This motivates us to delve into a new Biased Evidential Multi-view Learning (BEML) problem. To this end, we propose Fairness-Aware Multi-view Evidential Learning (FAML). FAML first introduces an adaptive prior based on training trajectories, which acts as a regularization strategy to flexibly calibrate the biased evidence learning process. Furthermore, we explicitly incorporate a fairness constraint based on class-wise evidence variance to promote balanced evidence allocation. In the multi-view fusion stage, we propose an opinion alignment mechanism to mitigate view-specific bias across views, thereby encouraging the integration of consistent and mutually supportive evidence. Theoretical analysis shows that FAML enhances fairness in the evidence learning process. Extensive experiments on six real-world multi-view datasets demonstrate that FAML achieves more balanced evidence allocation and improves both prediction performance and the reliability of uncertainty estimation compared to state-of-the-art methods.

This paper revisits the topic of rotation estimation through the lens of special unitary matrices. We begin by reformulating Wahba’s problem using $SU(2)$ to derive multiple solutions that yield linear constraints on corresponding quaternion parameters. We then explore applications of these constraints by formulating efficient methods for related problems. Finally, from this theoretical foundation, we propose two novel continuous representations for learning rotations in neural networks. Extensive experiments validate the effectiveness of the proposed methods.

Synthetically augmenting training datasets with diffusion models has been an effective strategy for improving generalization of image classifiers. However, existing techniques struggle to ensure the diversity of generation and increase the size of the data by up to 10-30x to improve the in-distribution performance. In this work, we show that synthetically augmenting part of the data that is not learned early in training with faithful images—containing same features but different noise—outperforms augmenting the entire dataset. By analyzing a two-layer CNN, we prove that this strategy improves generalization by promoting homogeneity in feature learning speed without amplifying noise. Our extensive experiments show that by augmenting only 30%-40% of the data, our method boosts generalization by up to 2.8% in a variety of scenarios, including training ResNet, ViT, ConvNeXt, and Swin Transformer on CIFAR-10/100, and TinyImageNet, with various optimizers including SGD and SAM. Notably, our method applied with SGD outperforms the SOTA optimizer, SAM, on CIFAR-100 and TinyImageNet.

Document-level knowledge graph (KG) construction faces a fundamental scaling challenge: existing methods either rely on expensive large language models (LLMs), making them economically unviable for large-scale corpora, or employ smaller models that produce incomplete and inconsistent graphs. We identify that this limitation stems not from model capabilities but from the absence of high-quality training data for document-level KG construction. To address this gap, we introduce SynthKG, a multi-step data synthesis pipeline that generates high-quality document-KG pairs through systematic chunking, decontextualization, and structured extraction using LLMs. By further fine-tuning a smaller LLM on synthesized document-KG pairs, we streamline the multi-step process into a single-step KG generation approach called Distill-SynthKG. Furthermore, we re-purpose existing question-answering datasets to establish KG evaluation datasets and introduce new evaluation metrics. Using KGs produced by Distill-SynthKG, we also design a novel graph-based retrieval framework for RAG. Experimental results demonstrate that Distill-SynthKG not only surpasses all baseline models in KG quality (including models up to eight times larger) but also consistently excels in retrieval and question-answering tasks. Additionally, our proposed graph retrieval framework outperforms all KG-retrieval methods across multiple benchmark datasets. We make SynthKG and Distill-SynthKG publicly available.

Transformers have achieved state-of-the-art performance across diverse language and vision tasks. This success drives the imperative to interpret their internal mechanisms with the dual goals of enhancing performance and improving behavioral control. Attribution methods help advance interpretability by assigning model outputs associated with a target concept to specific model components. Current attribution research primarily studies multi-layer perceptron (MLP) neurons and addresses relatively simple concepts such as factual associations (e.g., Paris is located in France). This focus tends to overlook the impact of the attention mechanism and lacks a unified approach for analyzing more complex concepts. To fill these gaps, we introduce Scalable Attention Module Discovery (SAMD), a concept-agnostic method for mapping arbitrary, complex concepts to specific attention heads of general transformer models. We accomplish this by representing each concept as a vector, calculating its cosine similarity with each attention head, and selecting the TopK-scoring heads to construct the concept-associated attention module. We then propose Scalar Attention Module Intervention (SAMI), a simple strategy to diminish or amplify the effects of a concept by adjusting the attention module using only a single scalar parameter. Empirically, we demonstrate SAMD on concepts of varying complexity, and visualize the locations of their corresponding modules. Our results demonstrate that module locations remain stable before and after LLM post-training, and confirm prior work on the mechanics of LLM multi-lingualism. Through SAMI, we facilitate jailbreaking on HarmBench (+72.7%) by diminishing “safety” and improve performance on the GSM8K benchmark (+1.6%) by amplifying “reasoning”. Lastly, we highlight the domain-agnostic nature of our approach by suppressing the image classification accuracy of vision transformers on ImageNet.

Vision–language models (VLMs) align images and text with remarkable success, yet the geometry of their shared embedding space remains poorly understood. To probe this geometry, we begin from the Iso-Energy Assumption, which exploits cross-modal redundancy: a concept that is truly shared should exhibit the same average energy across modalities. We operationalize this assumption with an Aligned Sparse Autoencoder (SAE) that encourages energy consistency during training while preserving reconstruction. We find that this inductive bias changes the SAE solution without harming reconstruction, giving us a representation that serves as a tool for geometric analysis. Sanity checks on controlled data with known ground truth confirm that alignment improves when Iso-Energy holds and remains neutral when it does not. Applied to foundational VLMs, our framework reveals a clear structure with practical consequences: **(*i*)** sparse *bimodal* atoms carry the entire *cross-modal* alignment signal; **(*ii*)** *unimodal* atoms act as *modality-specific* biases and fully explain the modality gap; **(*iii*)** removing unimodal atoms collapses the gap without harming performance; **(*iv*)** restricting vector arithmetic to the bimodal subspace yields in-distribution edits and improved retrieval. These findings suggest that the right inductive bias can both preserve model fidelity and render the latent geometry interpretable and actionable.

Computation methods for solving entropy-regularized reward optimization—a class of problems widely used for fine-tuning generative models—have advanced rapidly. Among those, Adjoint Matching (AM, Domingo-Enrich et al., 2025) has proven highly effective in continuous state spaces with differentiable rewards. Transferring these practical successes to _discrete_ generative modeling, however, remains particularly challenging and largely unexplored, mainly due to the drastic shift in generative model classes to discrete state spaces, which are nowhere differentiable. In this work, we propose Discrete Adjoint Matching (DAM)—a discrete variant of AM for fine-tuning discrete generative models characterized by Continuous-Time Markov Chains, such as diffusion-based large language models. The core of DAM is the introduction of _discrete adjoint_—an estimator of the optimal solution to the original problem but formulated on discrete domains—from which standard matching frameworks can be applied. This is derived via a purely statistical standpoint, in contrast to the control-theoretic viewpoint in AM, thereby opening up new algorithmic opportunities for general adjoint-based estimators. We showcase DAM’s effectiveness on synthetic and mathematical reasoning tasks.

Low-rank architectures have become increasingly important for efficient large language model (LLM) pre-training, providing substantial reductions in both parameter complexity and memory/computational demands. Despite these advantages, current low-rank methods face three critical shortcomings: (1) compromised model performance, (2) considerable computational overhead, and (3) limited activation memory savings. To address these limitations, we propose **C**ross-layer Low-**R**ank residual **Net**work (**CR-Net**), an innovative parameter-efficient framework inspired by our discovery that inter-layer activation residuals possess low-rank properties. CR-Net implements this insight through a dual-path architecture that efficiently reconstructs layer activations by combining previous-layer outputs with their low-rank differences, thereby maintaining high-rank information with minimal parameters. We further develop a specialized activation recomputation strategy tailored for CR-Net that dramatically reduces memory requirements. Extensive pre-training experiments across model scales from 60M to 7B parameters demonstrate that CR-Net consistently outperforms state-of-the-art low-rank frameworks while requiring fewer computational resources and less memory.

Cross-domain generalization is very important in Time Series Forecasting because similar historical information may lead to distinct future trends due to the domain-specific characteristics. Recent works focus on building unimodal time series foundation models and end-to-end multimodal supervised models. Since domain-specific knowledge is often contained in modalities like texts, the former lacks the explicit utilization of them, thus hindering the performance. The latter is tailored for end-to-end scenarios and does not support zero-shot inference for cross-domain scenarios. In this work, we introduce Aurora, a Multimodal Time Series Foundation Model, which supports multimodal inputs and zero-shot inference. Pretrained on Corss-domain Multimodal Time Series Corpus, Aurora can adaptively extract and focus on key domain knowledge contained in corrsponding text or image modalities, thus possessing strong Cross-domain generalization capability. Through tokenization, encoding, and distillation, Aurora can extract multimodal domain knowledge as guidance and then utilizes a Modality-Guided Multi-head Self-Attention to inject them into the modeling of temporal representations. In the decoding phase, the multimodal representations are used to generate the conditions and prototypes of future tokens, contributing to a novel Prototype-Guided Flow Matching for generative probabilistic forecasting. Comprehensive experiments on well-recognized benchmarks, including TimeMMD, TSFM-Bench and ProbTS, demonstrate the consistent state-of-the-art performance of Aurora on both unimodal and multimodal scenarios.

Optimization algorithms that seek flatter minima, such as Sharpness-Aware Minimization (SAM), are credited with improved generalization and robustness to noise. We ask whether such gains impact membership privacy. Surprisingly, we find that SAM is more prone to Membership Inference Attacks (MIA) than classical SGD across multiple datasets and attack methods, despite achieving lower test error. This suggests that the very geometric mechanism of SAM that improves generalization simultaneously exacerbates membership leakage. We investigate this phenomenon through extensive analysis of memorization and influence scores. Our results reveal that SAM is more capable of capturing atypical subpatterns, leading to higher memorization scores of samples. Conversely, SGD depends more heavily on majority features, exhibiting worse generalization on atypical subgroups and lower memorization. Crucially, this characteristic of SAM can be linked to lower variance in the output confidence for unseen samples, thereby amplifying membership signals. Finally, we model SAM under a perfectly interpolating linear regime and theoretically prove that geometric mechanism of SAM inherently reduces variance, guaranteeing a higher MIA advantage for confidence and likelihood ratio attacks.

We introduce jqBench, a new benchmark for evaluating language models on JSON querying and transformation tasks, where the intent can be given specified using natural language and/or examples. Whereas jqBench is mainly aimed at using the `jq` tool, it can be used to evaluate other programming languages that query and/or transform JSON. Benchmarks are automatically created from two rich sources of data: Stack Overflow discussions (751 instances with instructions and examples, called jqStack) and the Spider dataset for SQL generation from natural language (893 instances with instructions and JSON Schema, called jqSpider). We describe and analyze the automated pipeline for benchmark creation, and perform extensive baseline experiments on different models to analyze the complexity and failure modes. Using implicit feedback, the best model (Claude Opus 4.1) scores 77% on the jqStack benchmarks and 81\% on the jqSpider benchmarks. Additionally, we show (1) that access to the documentation surprisingly does not help, (2) `jq` performs comparable to Python, and (3) that automatic feedback (and therefore examples) is crucial. Besides the final benchmarks, we release the intermediate artifacts from each generation step (including failed or invalid conversions) as well as an LLM-friendly version of the documentation, to facilitate further research on JSON querying and transformation.

Diffusion-based policies have recently shown strong results in robot manipulation, but their extension to multi-task scenarios is hindered by the high cost of scaling model size and demonstrations. We introduce Skill Mixture-of-Experts Policy (SMP), a diffusion-based mixture-of-experts policy that learns a compact orthogonal skill basis and uses sticky routing to compose actions from a small, task-relevant subset of experts at each step. A variational training objective supports this design, and adaptive expert activation at inference yields fast sampling without oversized backbones. We validate SMP in simulation and on a real dual-arm platform with multi-task learning and transfer learning tasks, where SMP achieves higher success rates and markedly lower inference cost than large diffusion baselines. These results indicate a practical path toward scalable, transferable multi-task manipulation: learn reusable skills once, activate only what is needed, and adapt quickly when tasks change.

With the rise of deep neural networks, especially in safety-critical applications, robustness and interpretability are crucial to ensure their trustworthiness. Recent advances in 3D-aware classifiers that map image features to volumetric representation of objects, rather than relying solely on 2D appearance, have greatly improved robustness on out-of-distribution (OOD) data. Such classifiers have not yet been studied from the perspective of interpretability. Meanwhile, current concept-based XAI methods often neglect OOD robustness. We aim to address both aspects with CAVE - Concept Aware Volumes for Explanations - a new direction that unifies interpretability and robustness in image classification. We design CAVE as a robust and inherently interpretable classifier that learns sparse concepts from 3D object representation. We further propose 3D Consistency (3D-C), a metric to measure spatial consistency of concepts. Unlike existing metrics that rely on human-annotated parts on images, 3D-C leverages ground-truth object meshes as a common surface to project and compare explanations across concept-based methods. CAVE achieves competitive classification performance while discovering consistent and meaningful concepts across images in various OOD settings.

Model merging in weight space is an efficient way to reuse pretrained models, but existing methods typically assume matching architectures or sizes, making heterogeneous merges brittle or infeasible. We address this limitation by encoding model weights into a smooth latent space, enabling cross-architecture operations, and performing the merge in the latent space before decoding back to weights. This approach faces two major challenges. First, LLMs contain billions of parameters, which makes latent encoding computationally demanding. Second, using high compression ratios often hinders the encoder’s ability to generalize to unseen weights. We tackle these issues with a transformer-based variational autoencoder (VAE) trained in a two-stage compression curriculum with structured layer-aware chunking: the model first learns a high-capacity latent representation and then distills to a compact code, improving both stability and out-of-distribution generalization. To align heterogeneous models, we introduce a dimensionality-matching projection that allows interpolation between models of different sizes. Empirically, latent-space interpolation is consistently more robust than direct weight-space averaging and yields stronger downstream performance when merging models of different sizes. Together, these components provide a scalable, architecture-agnostic recipe for model merging.

Recent advances in Large Language Model (LLM) compression, such as quantization and pruning, have achieved notable success. However, as these techniques gradually approach their limits, relying on a single method for further compression has become increasingly challenging. In this work, we explore an alternative solution by combining quantization and sparsity. This joint approach, though promising, introduces new difficulties due to the inherently conflicting requirements on weight distributions: quantization favors compact ranges, while pruning benefits from high variance. To attack this problem, we propose Optimal Brain Restoration (OBR), a general and training-free framework that aligns pruning and quantization by error compensation between both. OBR minimizes performance degradation on downstream tasks by building on a second-order Hessian objective, which is then reformulated into a tractable problem through surrogate approximation and ultimately reaches a closed-form solution via group error compensation. Experiments show that OBR incurs only a 1.4 perplexity degradation on Llama2-7B to enable aggressive W4A4KV4 quantization with 50% sparsity, delivering up to 4.72x speedup and 6.4x memory reduction compared to the FP16-dense baseline.

Layer-wise normalization (LN) is an essential component of virtually all transformer-based large language models. While its effects on training stability are well documented, its role at inference time is poorly understood. Additionally, LN layers hinder mechanistic interpretability by introducing additional nonlinearities and increasing the interconnectedness of individual model components. Here we show that all LN layers can be removed from every GPT-2 model with only a small increase in validation loss (e.g. +0.03 cross-entropy loss for GPT-2 XL). Thus LN cannot play a substantial role in language modeling. We find that the amount of fine-tuning data needed for LN removal grows sublinearly with model parameters, suggesting scaling to larger models is feasible. We release a suite of LN-free GPT-2 models on Hugging Face. Furthermore, we test interpretability techniques on LN-free models. Direct logit attribution now gives the exact direct effect of individual components, while the accuracy of attribution patching does not significantly improve. We also confirm that GPT-2's "confidence neurons" are inactive in the LN-free models. Our work clarifies the role of LN layers in language modeling, showing that GPT-2-class models can function without LN layers. We hope that our LN-free analogues of the GPT-2 family of models will enable more precise interpretability research and improve our understanding of language models.