ICLR 2026 论文列表

3D object detection from LiDAR point cloud data is important for autonomous driving systems. Recent two-stage 3D object detectors struggle to achieve satisfactory performance due to limitations in proposal quality, stemming from the degradation of geometric detail information in the generated proposal features caused by high sparsity and uneven distribution of point clouds, as well as a lack of effective exploitation of surrounding contextual cues in the independent proposal refinement stage. To this end, we propose a Proposal-centric Transformer Network (PTN), which includes a Hierarchical Attentive Feature Alignment (HAFA) module and a Collaborative Proposal Refinement Module (CPRM). More concretely, to obtain multi-granularity proposal representations, HAFA employs a dual-stream architecture that extracts both coarse-grained voxel features and fine-grained point features to enhance proposal features, then harmo- nizes them through a feature alignment network in a unified space. The CPRM first generates object queries for all objects and then establishes contextual-aware interactions to extract complementary information from semantically similar and spatially relevant proposals. PTN achieves promising performance on large-scale Waymo and KITTI benchmark, demonstrating the superiority of PTN.

Zero-shot time-series forecasting holds great promise, but is still in its infancy, hindered by limited and biased data corpora, leakage-prone evaluation, and privacy and licensing constraints. We propose the first practical univariate time-series simulation pipeline, which is simultaneously fast enough for on-the-fly data generation and enables notable zero-shot forecasting performance on M-Series and GiftEval benchmarks that capture trend/seasonality/intermittency patterns, typical of industrial forecasting applications across a variety of domains. Our simulator is based off of a seasonal autoregressive integrated moving average (SARIMA) as its core data source. Due to instability in the autoregressive component, naive SARIMA parameter sampling often leads to unusable paths. Instead, our simulator follows a three-step procedure: (1) we sample well-behaved trajectories from the system's characteristic polynomial stability region; (2) we introduce a superposition scheme that combines multiple paths into rich multi-seasonality traces; and (3) we add rate-based heavy-tailed noise models to capture burstiness and intermittency alongside seasonalities and trends. Orders of magnitude faster than kernel-based generators, our simulator enables training on circa 1B unique purely simulated series generated on the fly, after which well-established neural backbones exhibit strong zero-shot generalization, surpassing strong statistical forecasters and recent foundation baselines, while operating under strict zero-shot protocol. Notably, on GiftEval we observe a "student-beats-teacher" effect: models trained on our simulations exceed the forecasting accuracy of the generating processes (i.e. AutoARIMA), suggesting emergent generalization beyond the simulator's components.

We investigate a principal-agent problem modeled within a Markov Decision Process, where the principal and the agent have their own rewards. The principal can provide subsidies to influence the agent’s action choices, and the agent’s resulting action policy determines the rewards accrued to the principal. Our focus is on designing a robust subsidy scheme that maximizes the principal’s cumulative expected return, even when the agent displays bounded rationality and may deviate from the optimal action policy after receiving subsidies. As a baseline, we first analyze the case of a perfectly rational agent and show that the principal’s optimal subsidy coincides with the policy that maximizes social welfare, the sum of the utilities of both the principal and the agent. We then introduce a bounded-rationality model: the globally $\epsilon$-incentive-compatible agent, who accepts any policy whose expected cumulative utility lies within $\epsilon$ of the personal optimum. In this setting, we prove that the optimal robust subsidy scheme problem simplifies to a one-dimensional concave optimization. This reduction not only yields a clean analytical solution but also highlights a key structural insight: optimal subsidies are concentrated along the social-welfare-maximizing trajectories. We further characterize the loss in social welfare—the degradation under the robust subsidy scheme compared to the maximum achievable—and provide an upper bound on this loss. Finally, we investigate a finer-grained, state-wise $\epsilon$-incentive-compatible model. In this setting, we show that under two natural definitions of state-wise incentive-compatibility, the problem becomes intractable: one definition results in non-Markovian agent action policy, while the other renders the search for an optimal subsidy scheme NP-hard.

Diffusion language models (DLMs) generate tokens in parallel through iterative denoising, which can reduce latency and enable bidirectional conditioning. However, the safety risks posed by jailbreak attacks that exploit this inference mechanism are not well understood. In this paper, we reveal that DLMs have a critical vulnerability stemming from their iterative denoising process and propose a countermeasure. Specifically, our investigation identifies that if an affirmative token for a harmful query appears at an intermediate step, subsequent denoising can be steered toward a harmful response even in aligned models. Furthermore, we demonstrate that the vulnerability enables existing optimization-based jailbreak attacks to be applied to MDLMs. Building on this analysis, we propose a novel safety alignment method tailored to DLMs that trains models to generate safe responses from contaminated intermediate denoising steps containing affirmative tokens. Our experiments indicate that the proposed method significantly mitigates the vulnerability with minimal impact on task performance. Furthermore, our method also improves robustness against conventional jailbreak attacks. Our work underscores the need for DLM-specific safety research. Our code is available at [https://github.com/mdl-lab/dlm-priming-vulnerability](https://github.com/mdl-lab/dlm-priming-vulnerability).

A geometric graph associated with a set of points $P= \{x_1, x_2, \cdots, x_n \} \subset \mathbb{R}^d$ and a fixed kernel function $\mathsf{K}:\mathbb{R}^d\times \mathbb{R}^d\to\mathbb{R}_{\geq 0}$ is a complete graph on $P$ such that the weight of edge $(x_i, x_j)$ is $\mathsf{K}(x_i, x_j)$. We present a fully-dynamic data structure that maintains a spectral sparsifier of a geometric graph under updates that change the locations of points in $P$ one at a time. The update time of our data structure is $n^{o(1)}$ with high probability, and the initialization time is $n^{1+o(1)}$. Under certain assumption, our data structure can be made robust against adaptive adversaries, which makes our sparsifier applicable in iterative optimization algorithms. We further show that the Laplacian matrices corresponding to geometric graphs admit a randomized sketch for maintaining matrix-vector multiplication and projection in $n^{o(1)}$ time, under sparse updates to the query vectors, or under modification of points in $P$.

Maximizing conditional mutual information (CMI) has recently been shown to enhance the effectiveness of teacher networks in knowledge distillation (KD). Prior work achieves this by fine-tuning a pretrained teacher to maximize a proxy of its CMI. However, fine-tuning large-scale teachers is often impractical, and proxy-based optimization introduces inaccuracies. To overcome these limitations, we propose Differentiable JPEG-based Input Perturbation (DJIP), a plug-and-play framework that improves teacher–student knowledge transfer without modifying the teacher. DJIP employs a trainable differentiable JPEG layer inserted before the teacher to perturb teacher inputs in a way that directly increases CMI. We further introduce a novel alternating optimization algorithm to efficiently learn the coding parameters of the JPEG layer to maximize the perturbed CMI. Extensive experiments on CIFAR-100 and ImageNet, across diverse distillers and architectures, demonstrate that DJIP consistently improves student accuracy-achieving up to 4.11% gains-while remaining computationally lightweight and fully compatible with standard KD pipelines.

Using intelligent systems to perceive psychological and social behaviors, that is, the underlying affective, cognitive, and pathological states that are manifested through observable behaviors and social interactions, remains a challenge due to their complex, multifaceted, and personalized nature. Existing work tackling these dimensions through specialized datasets and single-task systems often miss opportunities for scalability, cross-task transfer, and broader generalization. To address this gap, we curate Human Behavior Atlas, a unified benchmark of diverse behavioral tasks designed to support the development of foundation models for understanding psychological and social behaviors. Human Behavior Atlas comprises over 100,000 samples spanning text, audio, and visual modalities, covering tasks on *affective states*, *cognitive states*, *pathologies*, and *social processes*. Our unification efforts can reduce redundancy and cost, enable training to scale efficiently across tasks, and enhance generalization of behavioral features across domains. On Human Behavior Atlas, we train three models: Omnisapiens-7B SFT, Omnisapiens-7B BAM, and Omnisapiens-7B RL. We show that training on Human Behavior Atlas enables models to consistently outperform existing multimodal LLMs across diverse behavioral tasks. Pretraining on Human Behavior Atlas also improves transfer to novel behavioral datasets; with the targeted use of behavioral descriptors yielding meaningful performance gains. The benchmark, models, and codes can be found at: https://github.com/MIT-MI/human_behavior_atlas.

The Linear Representation Hypothesis asserts that the embeddings learned by neural networks can be understood as linear combinations of features corresponding to high-level concepts. Based on this ansatz, sparse autoencoders (SAEs) have recently become a popular method for decomposing embeddings into a sparse combination of linear directions, which have been shown empirically to often correspond to human-interpretable semantics. However, recent attempts to apply SAEs to multimodal embedding spaces (such as the popular CLIP embeddings for image/text data) have found that SAEs often learn ``split dictionaries,'' where most of the learned sparse features are essentially unimodal, active only for data of a single modality. In this work, we study how to effectively adapt SAEs for the setting of multimodal embeddings while ensuring multimodal alignment. We first argue that the existence of a split dictionary decomposition on an aligned embedding space implies the existence of a non-split dictionary with improved modality alignment. Then, we propose a new SAE-based approach to multimodal embedding decomposition using cross-modal random masking and group-sparse regularization. We apply our method to popular embeddings for image/text (CLIP) and audio/text (CLAP) data and show that, compared to standard SAEs, our approach learns a more multimodal dictionary while reducing the number of dead neurons and improving feature semanticity. We finally demonstrate how this improvement in alignment of concepts between modalities can enable improvements in the interpretability and control of cross-modal tasks.

Artificial Neural Networks (ANNs), including fully-connected networks and transformers, are highly flexible and powerful function approximators, widely applied in fields like computer vision and natural language processing. However, their inability to inherently respect causal structures can limit their robustness, making them vulnerable to covariate shift and difficult to interpret/explain. This poses significant challenges for their reliability in real-world applications. In this paper, we introduce Causal Transformers (CaTs), a general model class designed to operate under predefined causal constraints, as specified by a Directed Acyclic Graph (DAG). CaTs retain the powerful function approximation abilities of traditional neural networks while adhering to the underlying structural constraints, improving robustness, reliability, and interpretability at inference time. This approach opens new avenues for deploying neural networks in more demanding, real-world scenarios where robustness and explainability is critical.

Diffusion models have achieved impressive results in generating high-quality images. Yet, they often struggle to faithfully align the generated images with the input prompts. This limitation is associated with synchronous denoising, where all pixels simultaneously evolve from random noise to clear images. As a result, during generation, the prompt-related regions can only reference the unrelated regions at the same noise level, failing to obtain clear context and ultimately impairing text-to-image alignment. To address this issue, we propose asynchronous diffusion models, a novel framework that allocates distinct timesteps to different pixels and reformulates the pixel-wise denoising process. By dynamically modulating the timestep schedules of individual pixels, prompt-related regions are denoised more gradually than unrelated regions, thereby allowing them to leverage clearer inter-pixel context. Consequently, these prompt-related regions achieve better alignment in the final images. Extensive experiments demonstrate that our asynchronous diffusion models can significantly improve text-to-image alignment across diverse prompts.

Explaining online time series monitoring models is crucial across sensitive domains such as healthcare and finance, where temporal and contextual prediction dynamics underpin critical decisions. While recent XAI methods have improved the explainability of time series models, they mostly analyze each time step independently, overlooking temporal dependencies. This results in further challenges: explaining prediction changes is non-trivial, methods fail to leverage online dynamics, and evaluation remains difficult. To address these challenges, we propose Delta-XAI, which adapts 14 existing XAI methods through a wrapper function and introduces a principled evaluation suite for the online setting, assessing diverse aspects, such as faithfulness, sufficiency, and coherence. Experiments reveal that classical gradient-based methods, such as Integrated Gradients (IG), can outperform recent approaches when adapted for temporal analysis. Building on this, we propose Shifted Window Integrated Gradients (SWING), which incorporates past observations in the integration path to systematically capture temporal dependencies and mitigate out-of-distribution effects. Extensive experiments consistently demonstrate the effectiveness of SWING across diverse settings with respect to diverse metrics. Our code is publicly available at https://anonymous.4open.science/r/Delta-XAI.

Transferability estimation metrics are used to find a high-performing pre-trained model for a given target task without fine-tuning models and without access to the source dataset. Despite the growing interest in developing such metrics, the benchmarks used to measure their progress have gone largely unexamined. In this work, we empirically show the shortcomings of widely used benchmark setups to evaluate transferability estimation metrics. We argue that the benchmarks on which these metrics are evaluated are fundamentally flawed. We empirically demonstrate that their unrealistic model spaces and static performance hierarchies artificially inflate the perceived performance of existing metrics, to the point where simple, dataset-agnostic heuristics can outperform sophisticated methods. Our analysis reveals a critical disconnect between current evaluation protocols and the complexities of real-world model selection. To address this, we provide concrete recommendations for constructing more robust and realistic benchmarks to guide future research in a more meaningful direction.

Multimodal Large Language Models (MLLMs) achieve remarkable cross-modal understanding by training on vast web-scale datasets, but inadvertently internalize sensitive personal and proprietary information. Existing machine unlearning methods address this by irreversibly altering model parameters to permanently erase knowledge. This destructive paradigm conflicts with modern privacy regulations that mandate auditable, reversible, and user-controllable data management. To address these challenges, we propose Knowledge Externalization, a novel framework for reversible and modular knowledge management in MLLMs. We first propose Dual-Stream Memory Tuning, a method that transfers targeted knowledge from a model's internal parameters into external memory tokens. To mitigate gradient interference when externalizing multiple concepts, we further introduce Soft Orthogonal Weighting, a technique that preserves the independence of each token. Our resulting framework demonstrates three key capabilities: (i) It achieves effective forgetting of target concepts within the base model, while enabling high-fidelity knowledge restoration using the corresponding memory token. (ii) It supports continuous knowledge editing, allowing the information stored within an external token to be dynamically updated post-externalization. (iii) It displays a remarkable emergent ability for compositionality, where multiple memory tokens (including edited ones) can be freely combined to simultaneously recover knowledge corresponding to each concept. Our source code will be released in the near future.

Pretrained vision-language models (VLMs), such as CLIP, have shown remarkable potential in few-shot image classification and led to numerous effective transfer learning strategies. These methods leverage the pretrained knowledge of VLMs to enable effective domain adaptation while mitigating overfitting through parameter-efficient tuning or instance-based consistency constraints. However, such regularizations often neglect the geometric structure of data distribution, which may lead to distortion of the overall semantic representation. To overcome this limitation, we propose a novel fine-tuning method, Manifold-Preserving and Sculpting Tuning (MPS-Tuning). Regarding the data distribution in feature space as a semantic manifold, MPS-Tuning explicitly constrains the intrinsic geometry of this manifold while further sculpting it to enhance class separability. Specifically, MPS-Tuning preserves both macroscopic and microscopic topological structures of the original manifold by aligning Gram matrices of features before and after fine-tuning. Theoretically, this constraint is shown to approximate an upper bound of the Gromov-Wasserstein distance. Furthermore, features from the image and text modalities are paired, and pairwise similarities are optimized to enhance the manifold’s class discriminability. Extensive experiments demonstrate that MPS-Tuning significantly improves model performance while effectively preserving the structure of the semantic manifold. The code will be released.

A systematic, comparative investigation into the effects of low-quality data reveals a stark spectrum of robustness across modern probabilistic models. We find that autoregressive language models, from token prediction to sequence-to-sequence tasks, are remarkably resilient (for GPT-2, test NLL increases modestly from 2.87 to 3.59 despite 50\% token corruption). By contrast, under the same levels of data corruption, class-conditional diffusion models degrade catastrophically (image-label consistency plummets by 56.81\% relative to baseline), while classifiers show a moderate impact that diminishes with dataset scale. To explain these discrepancies, we analyze the results through a multi-perspective lens, integrating information theory, PAC learning, and gradient dynamics. \textcolor{blue}{These analyses suggest that robustness is heavily influenced by two key principles}: the \textbf{richness of conditioning information}, which constrains the learning problem, and the \textbf{absolute information content} of the training data, which allows the signal from correct information to dominate statistical noise.

Intelligent systems must form internal representations that support both broad generalization and precise identification. Here, we show that these two goals are fundamentally in tension with one another. We derive closed-form expressions proving that any model whose representations have a finite semantic resolution, impairing long-range similarity computations, must lie on a universal Pareto front linking its probability of correct generalization $p_S$ and identification $p_I$. We extend this analysis to general input spaces and to parallel processing scenarios, predicting a sharp $1/n$ collapse in the capacity of processing multiple inputs at the same time. A minimal ReLU network reproduces these laws: a resolution boundary emerges during learning, and empirical $(p_S,p_I)$ trajectories closely match the theory for linearly decaying similarity. Finally, we show that the same limits appear in far more complex systems, including a convolutional neural network and state-of-the-art vision–language models, indicating that learned finite-resolution similarity are broad and foundational informational constraints rather than toy-model artifacts. Together, these results provide a precise theory of the generalization–identification tradeoff and clarify how semantic resolution shapes the representational capacity of deep networks and brains alike.

Recent advancements in algorithms for local intrinsic dimension (LID) estimation have been closely tied to progress in neural networks (NN). However, NN architectures are often tailored to specific domains, such as audio or image data, incorporating inductive biases that limit their transferability across domains. Moreover, existing LID estimation methods leveraging these architectures are typically evaluated on either overly simplistic benchmarks or domain datasets where the true LID is unknown, resulting in potentially erroneous evaluations. To close this research gap, we first isolate problematic aspects of LID estimation and leverage them to analyze the limitations of state-of-the-art methods. Our approach employs several techniques to create LID benchmarks for arbitrary domains, including the introduction of a method to transform any manifold into the domain while preserving the manifold structure, thereby addressing challenges posed by biases in neural network-based methods. Our comparative analysis reveals critical limitations and identifies new directions for future development in LID estimation methods. Code will be available on github when published.

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.

Accurate recall from large-scale memories remains a core challenge for memory-augmented AI assistants performing question answering (QA), especially in similarity-dense scenarios where existing methods mainly rely on semantic distance to the query for retrieval. Inspired by how humans link information associatively, we propose AssoMem, a novel framework constructing an associative memory graph that anchors dialogue utterances to automatically extracted clues. This structure provides a rich organizational view of the conversational context and facilitates importance-aware ranking. Further, AssoMem integrates multi-dimensional retrieval signals—relevance, importance, and temporal alignment—using an adaptive mutual information (MI)-driven fusion strategy. Extensive experiments across three benchmarks and a newly introduced dataset, MeetingQA, demonstrate that AssoMem consistently outperforms state-of-the-art baselines, verifying its superiority in context-aware memory recall.