ICLR 2026 论文列表

Humans acquire semantic object representations from egocentric visual streams with minimal supervision. Importantly, the visual system processes with high resolution only the center of its field of view and learns similar representations for visual inputs occurring close in time. This emphasizes slowly changing information around gaze locations. This study investigates the role of central vision and slowness learning in the formation of semantic object representations in humans. We simulate five months of human-like visual experience using the Ego4D dataset and generate gaze coordinates with a state-of-the-art gaze prediction model. Using these predictions, we extract crops that mimic central vision and train a time-contrastive Self-Supervised Learning model on them. Our results show that combining temporal slowness and central vision improves the encoding of different semantic facets of object representations. Specifically, focusing on central vision strengthens the extraction of foreground object features, while considering temporal slowness, especially during fixational eye movements, allows the model to encode broader semantic information about objects. These findings provide new insights into the mechanisms by which humans may develop semantic object representations from natural visual experience. Our code will be made public upon acceptance.

Reinforcement learning based post-training has recently emerged as a powerful paradigm for enhancing the alignment and reasoning capabilities of multimodal large language models (MLLMs). While *vision-centric* post-training is crucial for enhancing MLLMs’ intrinsic understanding of visual signals, current post-training paradigms are predominantly *text-centric*, where dense visual inputs are only leveraged to extract sparse cues for text-based reasoning. There exist a few approaches in this direction, however, they often still rely on text as an intermediate mediator or introduce additional visual generative designs. In this work, we introduce **Visual Jigsaw**, a generic *self-supervised* post-training framework designed to strengthen visual understanding in MLLMs. Visual Jigsaw is formulated as a general ordering task: visual inputs are partitioned, shuffled, and the model must reconstruct the visual information by producing the correct permutation in natural language. This naturally aligns with reinforcement learning from verifiable rewards (RLVR), requires no additional visual generative components, and derives its supervisory signal automatically without any annotations. We instantiate Visual Jigsaw across three visual modalities, including images, videos, and 3D data. Extensive experiments demonstrate substantial improvements in fine-grained perception, temporal reasoning, and 3D spatial understanding. Our findings highlight the potential of self-supervised vision-centric tasks in post-training MLLMs and aim to inspire further research on vision-centric pretext designs.

Obtaining a single-vector representation from a Large Language Model's (LLM) token-level outputs is a critical step for nearly all sentence-level tasks. However, standard pooling methods like mean or max aggregation treat tokens as an independent set, discarding the rich relational structure captured by the model's self-attention layers and making them susceptible to signal dilution. To address this, we introduce GLOT, a lightweight, structure-aware pooling module that reframes pooling as relational learning followed by aggregation. Operating on the outputs of a frozen LLM, GLOT first constructs a latent token-similarity graph, then refines token representations with a graph neural network, and finally aggregates them using a readout layer. Experimentally, our approach is remarkably robust and efficient: on a diagnostic stress test where 90% of tokens are random distractors, GLOT maintains over 97% accuracy while baseline methods collapse. Furthermore, it competitive with state-of-the-art techniques on benchmarks like GLUE and MTEB with 20x fewer trainable parameters and speeds up the training time by over 100x compared with parameter-efficient fine-tuning methods. Supported by a theoretical analysis of its expressive power, our work shows that learning over token graphs is a powerful paradigm for the efficient adaptation of frozen LLMs.

Watermarking has been widely used for copyright protection of digital images. Deep learning-based (DL) watermarking systems have recently emerged as more effective than traditional methods, offering improved fidelity and resilience against attacks. Among the various threats to DL watermarking systems, self-re-watermarking attacks represent a critical and underexplored challenge. In such attacks, the same encoder is maliciously reused to embed a new message into an already watermarked image. This process effectively prevents the original decoder from retrieving the original watermark without introducing perceptual artifacts. In this work, we make two key contributions. First, we introduce the self-re-watermarking threat model as a novel attack vector and demonstrate that existing state-of-the-art watermarking methods consistently fail under such attacks. Second, we develop a self-aware watermarking framework to defend against this threat. Our key insight for mitigating this risk is to limit the sensitivity of the watermarking models to the inputs, thereby resisting re-embedding of new watermarks. To achieve this, we propose a self-aware deep watermarking framework that extends Lipschitz constraints to the watermarking process, regulating encoder–decoder sensitivity in a principled manner. In addition, the framework incorporates re-watermarking adversarial training, which further constrains sensitivity to distortions arising from re-embedding. The proposed method provides theoretical bounds on message recoverability under malicious encoder based re-watermarking and demonstrates strong empirical robustness against diverse scenarios of re-watermarking attempts. Moreover, it maintains high visual fidelity and demonstrates competitive robustness against common image processing distortions compared to state-of-the-art watermarking methods. This work establishes a robust defense against both standard distortions and self-re-watermarking attacks. Code available at https://github.com/SVithurabiman/SRW.

As AI systems become more capable of complex agentic tasks, they also become more capable of pursuing undesirable objectives and causing harm. Previous work has attempted to catch these unsafe instances by interrogating models directly about their objectives and behaviors. However, the main weakness of trusting interrogations is that models can lie. We propose self-report fine-tuning (SRFT), a simple supervised fine-tuning technique that trains models to occasionally make factual mistakes, then admit them when asked. We show that the admission of factual errors in simple question-answering settings generalizes out-of-distribution (OOD) to the admission of hidden misaligned objectives in adversarial agentic settings. We evaluate SRFT in OOD stealth tasks, where models are instructed to complete a hidden misaligned objective alongside a user-specified objective without being caught by monitoring. After SRFT, models are more likely to confess the details of their hidden objectives when interrogated, even under strong pressure not to disclose them. Interrogation on SRFT models can detect hidden objectives with near-ceiling performance (F1 score = 0.98), while the baseline model lies when interrogated under the same conditions (F1 score = 0). Interrogation on SRFT models can further elicit the content of the hidden objective, recovering 28-100\% details, compared to 0\% details recovered in the baseline model and by prefilled assistant turn attacks. This provides a promising technique for promoting honesty propensity and incriminating misaligned AI systems.

We introduce a highly expressive class of function approximators called *Splat Regression Models*. Model outputs are mixtures of heterogeneous and anisotropic bump functions, termed *splats*, each weighted by an output vector. The power of splat modeling lies in its ability to locally adjust the scale and direction of each splat, achieving both high interpretability and accuracy. Fitting splat models reduces to optimization over the space of mixing measures, which can be implemented using Wasserstein-Fisher-Rao gradient flows. As a byproduct, we recover the popular *Gaussian Splatting* methodology as a special case, providing a unified theoretical framework for this state-of-the-art technique that clearly disambiguates the inverse problem, the model, and the optimization algorithm. Through numerical experiments, we demonstrate that the resulting models and algorithms constitute a flexible and promising approach for solving diverse approximation, estimation, and inverse problems involving low-dimensional data.

Central to the widespread use of t-distributed stochastic neighbor embedding (t-SNE) is the conviction that it produces visualizations whose structure roughly matches that of the input. To the contrary, we prove that (1) the strength of the input clustering, and (2) the extremity of outlier points, cannot be reliably inferred from the t-SNE output. We demonstrate the prevalence of these failure modes in practice as well.

Autoencoders are widely used for dimensionality reduction, based on the assumption that high-dimensional data lies on low-dimensional manifolds. Regularized autoencoders aim to preserve manifold geometry during dimensionality reduction, but existing approaches often suffer from non-injective mappings and overly rigid constraints that limit their effectiveness and robustness. In this work, we identify encoder non-injectivity as a core bottleneck that leads to poor convergence and distorted latent representations. To ensure robustness across data distributions, we formalize the concept of admissible regularization and provide sufficient conditions for its satisfaction. In this work, we propose the Bi-Lipschitz Autoencoder (BLAE), which introduces two key innovations: (1) an injective regularization scheme based on a separation criterion to eliminate pathological local minima, and (2) a bi-Lipschitz relaxation that preserves geometry and exhibits robustness to data distribution drift. Empirical results on diverse datasets show that BLAE consistently outperforms existing methods in preserving manifold structure while remaining resilient to sampling sparsity and distribution shifts.

Joint-embedding self-supervised learning (SSL), the key paradigm for unsupervised representation learning from visual data, learns from invariances between semantically-related data pairs. We study the one-to-many mapping problem in SSL, where each datum may be mapped to multiple valid targets. This arises when data pairs come from naturally occurring generative processes, e.g., successive video frames. We show that existing methods struggle to flexibly capture this conditional uncertainty. As a remedy, we introduce a variational distribution that models this uncertainty in the latent space, and derive a lower bound on the pairwise mutual information. We also propose a simpler variant of the same idea using sparsity regularization. Our model, AdaSSL, applies to both contrastive and predictive SSL methods, and we empirically show its versatility on identifiability, generalization, fine-grained image understanding, and world modeling on videos.

Mixture-of-Experts (MoE) architectures have become pivotal for large-scale multimodal models. However, their routing mechanisms typically overlook the informative, time-varying interaction dynamics between modalities. This limitation hinders expert specialization, as the model cannot explicitly leverage intrinsic modality relationships for effective reasoning. To address this, we propose a novel framework that guides MoE routing using quantified temporal interaction. A multimodal interaction-aware router learns to dispatch tokens to experts based on the nature of their interactions. This dynamic routing encourages experts to acquire generalizable interaction-processing skills rather than merely learning task-specific features. Our framework builds on a new formulation of temporal multimodal interaction dynamics, which are used to guide expert routing. We first demonstrate that these temporal multimodal interactions reveal meaningful patterns across applications, and then show how they can be leveraged to improve both the design and performance of MoE-based models. Comprehensive experiments on challenging multimodal benchmarks validate our approach, demonstrating both enhanced performance and improved interpretability.

Calibrated confidence estimates are necessary for large language model (LLM) outputs to be trusted by human users. While LLMs can express their confidence in human-interpretable ways, verbalized LLM-generated confidence scores have empirically been found to be miscalibrated, reporting high confidence on instances with low accuracy and thereby harming trust and safety. We hypothesize that this overconfidence often stems from a given LLM’s heightened suggestibility when faced with claims that it encodes little information about; we empirically validate this hypothesis, finding more suggestibility on lower-accuracy claims. Building on this finding, we introduce Distractor-Normalized Coherence (DINCO), which estimates and accounts for an LLM’s suggestibility bias by having the model verbalize its confidence independently across several self-generated distractors (i.e. alternative claims), and normalizes by the total verbalized confidence. To further improve calibration, we leverage generator-validator disagreement, augmenting normalized validator confidence with a consistency-based estimate of generator confidence. Here, we frame the popular approach of self-consistency as leveraging coherence across sampled generations, and normalized verbalized confidence as leveraging coherence across validations on incompatible claims, allowing us to integrate these complementary dimensions of coherence into DINCO. Moreover, our analysis shows that DINCO provides less saturated, and therefore more usable, confidence estimates, and that further sampling alone cannot close the gap between DINCO and baselines, with DINCO at 10 runs outperforming self-consistency at 100. We include our code in the supplementary.

Most of today's AI systems are constrained by human-designed, fixed architectures and cannot autonomously and continuously improve themselves. The scientific method, on the other hand, is a cumulative and open-ended system, where each innovation builds upon previous artifacts, enabling future discoveries. There is growing hope that the current manual process of advancing AI could itself be automated. If done safely, such automation would accelerate AI development and allow us to reap its benefits much sooner. This prospect raises the question of how AI systems can endlessly improve themselves while getting better at solving relevant problems. Meta-learning can automate the discovery of novel algorithms, but is limited by first-order improvements and the human design of a suitable search space. The Gödel machine proposed a theoretical alternative: a self-improving AI that repeatedly modifies itself in a provably beneficial manner. Unfortunately, proving that most changes are net beneficial is impossible in practice. We introduce the Darwin Gödel Machine (DGM), a novel self-improving system that iteratively modifies its own code (thereby also improving its ability to modify its own codebase) and empirically validates each change using coding benchmarks. Inspired by Darwinian evolution and open-endedness research, the DGM grows an archive of generated coding agents. It samples agents from this archive, which self-modify to create new, interesting versions of themselves. This open-ended exploration forms a growing tree of diverse, high-quality agents and allows the parallel exploration of many different paths through the search space. Empirically, the DGM automatically improves its coding capabilities (e.g., better code editing tools, long-context window management, peer-review mechanisms), increasing performance on SWE-bench from 20.0% to 50.0%, and on Polyglot from 14.2% to 30.7%. Furthermore, the DGM significantly outperforms baselines without self-improvement or open-ended exploration. All experiments were done with safety precautions (e.g., sandboxing, human oversight). Overall, the DGM represents a significant step toward self-improving AI, capable of gathering its own stepping stones along a path that unfolds into endless innovation.

Long-tailed semi-supervised learning is hampered by a vicious cycle of confirmation bias, where skewed pseudo-labeling progressively marginalizes tail classes. This challenge is compounded in real-world scenarios by a class distribution mismatch between labeled and unlabeled data, rendering the bias unpredictable and difficult to mitigate. While existing methods adapt Logit Adjustment (LA) using dynamic estimates of the unlabeled distribution, we argue their effectiveness is undermined by two critical limitations stemming from LA's core design, i.e., its class-wise and overall adjustment mechanisms. First, their reliance on simple frequency counting overestimates the prevalence of head classes due to sample redundancy, leading to harmful over-suppression. Second, and more critically, they overlook the interplay between the above two types of adjustment, treating the overall adjustment strength as a fixed hyperparameter. This is a significant oversight, as we empirically find that the optimal strength is highly sensitive to the estimated distribution. To address these limitations, we propose Co-Calibrated Logit Adjustment (CoLA), a framework that co-designs the class-wise and overall LA components. Specifically, CoLA refines the class-wise adjustment by estimating each class's effective sample size via the effective rank of its representations. Subsequently, it formulates the overall adjustment strength as a learnable parameter, which is optimized through a meta-learning procedure on a proxy validation set constructed to mirror the refined distribution. Supported by a theoretical generalization bound, our extensive experiments show that CoLA outperforms existing baselines on $4$ public benchmarks across standard long-tail setups.

We propose a simple yet novel data augmentation method for general data modalities based on energy-based modeling and principles from information geometry. Unlike most existing learning-based data augmentation methods, which rely on learning latent representations with generative models, our proposed framework enables an intuitive construction of a geometrically aware latent space that represents the structure of the data itself, supporting efficient and explicit encoding and decoding procedures. We then present and discuss how to design latent spaces that will subsequently control the augmentation with the proposed algorithm. Empirical results demonstrate that our data augmentation method achieves competitive performance in downstream tasks compared to other baselines, while offering fine-grained controllability that is lacking in the existing literature.

For almost 70 years, researchers have typically selected the width of neural networks’ layers either manually or through automated hyperparameter tuning methods such as grid search and, more recently, neural architecture search. This paper challenges the status quo by introducing an easy-to-use technique to learn an \textit{unbounded} width of a neural network's layer \textit{during training}. The method jointly optimizes the width and the parameters of each layer via standard backpropagation. We apply the technique to a broad range of data domains such as tables, images, text, sequences, and graphs, showing how the width adapts to the task's difficulty. A by product of our width learning approach is the easy truncation of the trained network at virtually zero cost, achieving a smooth trade-off between performance and compute resources. Alternatively, one can dynamically compress the network until performances do not degrade. In light of recent foundation models trained on large datasets, requiring billions of parameters and where hyper-parameter tuning is unfeasible due to huge training costs, our approach introduces a viable alternative for width learning.

We introduce SEINT, a novel Special Euclidean group-Invariant (SE(\emph{p})) metric for comparing probability distributions on $p$-dimensional measured Banach spaces. Existing SE(\emph{p})-invariant alignment methods often face high computational costs or lack metric guarantees. To overcome these limitations, we develop a polar transport discrepancy combined with distance convolution to extract SE(\emph{p})-invariant representations. These representations are then used to compute the alignment between two distributions via optimal transport. Theoretically, we prove that SEINT is a well-defined metric on the space of isometry classes of normed vector spaces. Beyond its inherent SE(\emph{p})-invariance, SEINT also supports cross-space distribution comparison. Computationally, SEINT aligns two samples of size $n$ with a complexity of just $\mathcal{O}(n\log n)$ to $\mathcal{O}(n^2)$. Extensive experiments validate its advantages: As a robust metric, it outperforms or matches existing SE(\emph{p})-invariant methods in classification and cross-space tasks under isometries. As a regularizer, it greatly enhances molecular generation performance across both pre-training and fine-tuning tasks, achieving state-of-the-art (SOTA) results on key benchmarks.

We present self-speculative masked diffusions, a new class of masked diffusion generative models for discrete data that require significantly fewer function evaluations to generate samples. Standard masked diffusion models predict factorized logits over currently masked positions. A number of masked positions are then sampled, however, the factorization approximation means that sampling too many positions in one go leads to poor sample quality. As a result, many simulation steps and therefore neural network function evaluations are required to generate high-quality data. We reduce the computational burden by generating \emph{non-factorized} predictions over masked positions. This is achieved by modifying the final transformer attention mask from non-causal to causal, enabling draft token generation and parallel validation via a novel, model-integrated speculative sampling mechanism. This results in a non-factorized predictive distribution over masked positions in a single forward pass. We apply our method to GPT2 scale text modelling and protein sequences generation, finding that we can achieve a ~2x reduction in the required number of network forward passes relative to standard masked diffusion models.

Retrofitting large language models (LLMs) with new behaviors typically requires full finetuning or distillation—costly steps that must be repeated for every architecture. In this work, we introduce ⌘V (Command-V), a backpropagation-free behavior transfer method that copies an existing residual representation adapter from a donor model and pastes its effect into an architecturally different recipient model. ⌘V profiles layer activations on a small prompt set, derives linear converters between corresponding layers, and applies the donor intervention in the recipient’s activation space. This process does not require access to the original training data and needs minimal compute. In three case studies—safety-refusal enhancement, jailbreak facilitation, and automatic chain-of-thought reasoning—⌘V matches the performance of direct finetuning while using orders of magnitude less resources.

Event cameras deliver visual data with high temporal resolution, low latency, and minimal redundancy, yet their asynchronous, sparse sequential nature challenges standard tensor-based machine learning (ML). While the recent asynchronous-to-synchronous (A2S) paradigm aims to bridge this gap by asynchronously encoding events into learned features for ML pipelines, existing A2S approaches often sacrifice expressivity and generalizability compared to dense, synchronous methods. This paper introduces EVA (EVent Asynchronous feature learning), a novel A2S framework to generate highly expressive and generalizable event-by-event features. Inspired by the analogy between events and language, EVA uniquely adapts advances from language modeling in linear attention and self-supervised learning for its construction. In demonstration, EVA outperforms prior A2S methods on recognition tasks (DVS128-Gesture and N-Cars), and represents the first A2S framework to successfully master demanding detection tasks, achieving a 47.7 mAP on the Gen1 dataset. These results underscore EVA's potential for advancing real-time event-based vision applications.