ICLR 2026 论文列表

Current approaches to LLM safety fundamentally rely on a brittle cat-and-mouse game of identifying and blocking known threats via guardrails. We argue for a fresh approach: robust safety comes not from enumerating what is harmful, but from \emph{deeply understanding what is safe}. We introduce \textbf{T}rust \textbf{T}he \textbf{T}ypical \textbf{(T3)}, a framework that operationalizes this principle by treating safety as an out-of-distribution (OOD) detection problem. T3 learns the distribution of acceptable prompts in a semantic space and flags any significant deviation as a potential threat. Unlike prior methods, it requires no training on harmful examples, yet achieves state-of-the-art performance across 18 benchmarks spanning toxicity, hate speech, jailbreaking, multilingual harms, and over-refusal, reducing false positive rates by up to 40x relative to specialized safety models. A single model trained only on safe English text transfers effectively to diverse domains and over 14 languages without retraining. Finally, we demonstrate production readiness by integrating a GPU-optimized version into vLLM, enabling continuous guardrailing during token generation with less than 6\% overhead even under dense evaluation intervals on large-scale workloads.

Score-based methods are powerful across machine learning, but they face a paradox: theoretically path-independent, yet practically path-dependent. We resolve this by proving that practical training objectives differ from the ideal, ground-truth objective by a crucial, overlooked term: the path variance of the score function. We propose the MinPV (**Min**imum **P**ath **V**ariance) Principle to minimize this path variance. Our key contribution is deriving a closed-form expression for the variance, making optimization tractable. By parameterizing the path with a flexible Kumaraswamy Mixture Model, our method learns data-adaptive, low-variance paths without heuristic manual selection. This principled optimization of the complete objective yields more accurate and stable estimators, establishing new state-of-the-art results on challenging benchmarks and providing a general framework for optimizing score-based interpolation.

Mixture of Experts (MoE) models have achieved great success by significantly improving performance while maintaining computational efficiency through sparse expert activation. However, their enormous parameter sizes and memory demands pose significant challenges for deployment in resource-constrained environments. Vector Quantization (VQ) offers a promising approach for ultra-low-bit compression in Large Language Models (LLMs) by constructing and leveraging a codebook—where weight vectors are mapped to the most similar discrete codewords within the codebook. However, its direct application to MoEs suffers from significant performance degradation caused by two critical obstacles: (1) redundant representation among experts leads to VQ repeatedly quantizing similar representations for each expert, resulting in inefficient utilization of the limited codebook capacity; and (2) cumulative outputs bias, amplified by expert aggregation, leads to distributional shifts in the quantized outputs, resulting in degraded model accuracy. To this end, we propose KBVQ-MoE, a novel VQ framework to enhance extremely low-bit quantization for MoE-based LLMs. KBVQ-MoE introduces two lightweight and offline techniques that introduce negligible runtime computational and memory overhead: (1) Input-driven redundancy elimination, where a Karhunen–Loève Transform (KLT) guided singular value decomposition (SVD) extracts and shares dominant weight components across experts. (2) Bias-corrected output stabilization, where vector quantization is applied to expert-specific (i.e., non-redundant) representations and the quantized outputs are corrected with channel-wise affine compensation. Experiments on various MoE LLMs demonstrate that KBVQ-MoE preserves accuracy substantially better than existing quantization methods. For instance, 3-bit quantization of Qwen1.5-MoE-A2.7B achieves an average accuracy of 67.99, nearly identical to the FP16 baseline of 68.07, underscoring the potential of KBVQ-MoE for efficient deployment on edge devices and other resource-constrained platforms.

Federated Learning (FL) is a popular distributed learning paradigm to break down data silo. Traditional FL approaches largely rely on gradient-based updates, facing significant issues about heterogeneity, scalability, convergence, and overhead, etc. Recently, some analytic-learning-based work has attempted to handle these issues by eliminating gradient-based updates via analytical (i.e., closed-form) solutions. Despite achieving superior invariance to data heterogeneity, these approaches are fundamentally limited by their single-layer linear model with a frozen pre-trained backbone. As a result, they can only achieve suboptimal performance due to their lack of representation learning capabilities. In this paper, to enable representable analytic models while preserving the ideal invariance to data heterogeneity for FL, we propose our Deep Analytic Federated Learning approach, named DeepAFL. Drawing inspiration from the great success of ResNet in gradient-based learning, we design gradient-free residual blocks in our DeepAFL with analytical solutions. We further introduce an efficient layer-wise protocol for training our deep analytic models layer by layer in FL through least squares. Both theoretical analyses and empirical evaluations validate our DeepAFL's superior performance with its dual advantages in heterogeneity invariance and representation learning, outperforming state-of-the-art baselines by up to 5.68%-8.42% across three benchmark datasets. The related codes will be made open-sourced upon the acceptance of this paper.

We propose Q-learning with Adjoint Matching (QAM), a novel TD-based reinforcement learning (RL) algorithm that tackles a long-standing challenge in continuous-action RL: efficient optimization of an expressive diffusion/flow-matching based policy with respect to a parameterized value function (i.e., the critic $Q_\phi(s, a)$). Effective optimization requires exploiting the first-order information of the critic (i.e., the action gradient, $\nabla_a Q_\phi(s, a)$), but it is especially challenging to do so for flow/diffusion policy because direct gradient-based optimization via backpropagation through their multi-step denoising process is unstable. Existing methods work around either by only using the value and discarding the gradient information, or by relying on approximations that sacrifice policy expressivity or bias the learned policy. QAM sidesteps both of these challenges by leveraging adjoint matching, a recently proposed technique in generative modeling, which transforms the critic's action gradient to form a step-wise objective function that is free from unstable backpropagation, while providing an unbiased, expressive policy at the optimum. Combined with temporal-difference (TD) backup for critic learning, QAM consistently outperforms prior approaches across challenging, sparse reward tasks in both offline and offline-to-online RL settings.

Conventional continual pretraining (CPT) for large language model (LLM) domain adaptation often suffers from catastrophic forgetting and limited domain capacity. Existing strategies adopt layer expansion, introducing additional trainable parameters to accommodate new knowledge. However, the uniform expansion and updates still entangle general and domain learning, undermining its effectiveness. Our pilot studies reveal that LLMs exhibit functional specialization, where layers and units differentially encode general-critical capabilities, suggesting that parameter expansion and optimization should be function-aware. We then propose ADEPT, Adaptive Expansion and Dynamic Decoupled Tuning for continual pretraining, a two-stage framework for domain-adaptive CPT. ADEPT first performs General-Competence Guided Selective Layer Expansion, duplicating layers least critical for the general domain to increase representational capacity while minimizing interference with general knowledge. It then applies Adaptive Unit-Wise Decoupled Tuning, disentangling parameter units within expanded layers according to their general-domain importance and assigning asymmetric learning rates to balance knowledge injection and retention. Experiments on mathematical and medical domains show that ADEPT outperforms full-parameter CPT by up to 5.76% on the general benchmarks and 5.58% on the target domain benchmarks with only 15% of parameters tuned and less than 50% training time. Ablation studies, theoretical analysis, and extended investigations further demonstrate the necessity of targeted expansion and decoupled optimization, providing new principles for efficient and robust domain-adaptive CPT. Our code is open-sourced at https://anonymous.4open.science/r/ADEPT-F2E3

We introduce Stackelberg Learning from Human Feedback (SLHF), a new framework for preference optimization. SLHF frames the alignment problem as a sequential-move game between two policies: a Leader, which commits to an action, and a Follower, which responds conditionally on the Leader's action. This approach decomposes preference optimization into a refinement problem for the Follower and an optimization problem against an adversary for the Leader. Unlike Reinforcement Learning from Human Feedback (RLHF), which assigns scalar rewards to actions, or Nash Learning from Human Feedback (NLHF), which seeks a simultaneous-move equilibrium, SLHF leverages the asymmetry of sequential play to capture richer preference structures. The sequential design of SLHF naturally enables inference-time refinement, as the Follower learns to improve the Leader’s actions, and these refinements can be leveraged through iterative sampling. We compare the solution concepts of SLHF, RLHF, and NLHF, and lay out key advantages in consistency, data sensitivity, and robustness to intransitive preferences. Experiments on large language models demonstrate that SLHF achieves strong alignment across diverse preference datasets, scales from 0.5B to 8B parameters, and yields inference-time refinements that transfer across model families without further fine-tuning.

High-dimensional data visualization (HDV) plays an important role in data science and engineering applications. Traditional HDV methods, such as Autoencoder and t-SNE, require hyperparameter tuning and iterative optimization on every dataset and cannot effectively utilize the knowledge from historical datasets, which lowers the efficiency, convenience, and accuracy in real applications. In this paper, we present AutoDV, an end-to-end deep learning model, for high-dimensional data visualization. AutoDV is built upon a graph transformer network and an invariant loss function and is trained on a number of diverse datasets converted into multi-weight graphs. Given a new dataset, AutoDV outputs the 2D or 3D embeddings of all data points directly. AutoDV has the following merits: 1) There is no hyperparameter selection during the data visualization stage; 2) The end-to-end model avoids re-training or iterative optimization when visualizing data; 3) The input dataset can have any number of features and can be from any domain. Our experiments show that AutoDV can successfully generalize to unseen datasets without retraining with 89.37\% precision of t-SNE and 91.05\% precision of UMAP on the unseen CIFAR10 datasets. Compared with existing parametric data visualization deep models, our method obtains significant improvement with 86.65% precision gain. AutoDV can perform even better than t-SNE and UMAP on gene and UCI tabular datasets.

Recent empirical evidence shows that LLM representations encode human-interpretable concepts. Nevertheless, the mechanisms by which these representations emerge remain largely unexplored. To shed further light on this, we introduce a novel generative model that generates tokens on the basis of such concepts formulated as latent discrete variables. Under mild conditions, even when the mapping from the latent space to the observed space is non-invertible, we establish rigorous identifiability result: the representations learned by LLMs through next-token prediction can be approximately modeled as the logarithm of the posterior probabilities of these latent discrete concepts given input context, up to an invertible linear transformation. This theoretical finding: 1) provides evidence that LLMs capture essential underlying generative factors, 2) offers a unified and principled perspective for understanding the linear representation hypothesis, and 3) motivates a theoretically grounded approach for evaluating sparse autoencoders. Empirically, we validate our theoretical results through evaluations on both simulation data and the Pythia, Llama, and DeepSeek model families.

As Large Language Models (LLMs) are increasingly applied in high-stakes domains, their ability to reason strategically under uncertainty becomes critical. Poker provides a rigorous testbed, requiring not only strong actions but also principled, game-theoretic reasoning. In this paper, we conduct a systematic study of LLMs in multiple realistic poker tasks, evaluating both gameplay outcomes and reasoning traces. Our analysis reveals LLMs fail to compete against traditional algorithms and identifies three recurring flaws: reliance on heuristics, factual misunderstandings, and a “knowing–doing” gap where actions diverge from reasoning. An initial attempt with behavior cloning and step-level reinforcement learning improves reasoning style but remains insufficient for accurate game-theoretic play. Motivated by these limitations, we propose ToolPoker, a tool-integrated reasoning framework that combines external solvers for GTO-consistent actions with more precise professional-style explanations. Experiments demonstrate that ToolPoker achieves state-of-the-art gameplay while producing reasoning traces that closely reflect game-theoretic principles.

AI agents have been developed for complex real-world tasks from coding to customer service. But AI agent evaluations suffer from many challenges that undermine our understanding of how well agents really work (Figure 1). We introduce the Holistic Agent Leaderboard (HAL) to address these challenges. We make three main contributions. First, we provide a standardized evaluation harness that orchestrates parallel evaluations across hundreds of VMs, reducing evaluation time from weeks to hours while eliminating common implementation bugs. Second, we conduct three-dimensional analysis spanning models, scaffolds, and benchmarks. We validate the harness by conducting 21,730 agent rollouts across 9 models and 9 benchmarks in coding, web navigation, science, and customer service with a total cost of about $40,000. Our analysis reveals surprising insights, such as higher reasoning effort reducing accuracy in the majority of runs. Third, we use LLM-aided log inspection to uncover previously unreported behaviors, such as searching for the benchmark on HuggingFace instead of solving a task, or misusing credit cards in flight booking tasks. We share all agent logs, comprising 2.5B tokens of language model calls, to incentivize further research into agent behavior. By standardizing how the field evaluates agents and addressing common pitfalls in agent evaluation, we hope to shift the focus from agents that ace benchmarks to agents that work reliably in the real world.

Data augmentation is a fundamental technique in deep learning, widely applied in both representation learning and automated data augmentation (AutoDA). In representation learning, augmentations are used to construct contrastive views for learning task-agnostic embeddings, while in AutoDA the augmentations are directly optimized to improve downstream task performance. However, existing paradigms face critical limitations: representation learning relies on a two-stage scheme with limited adaptability, and current AutoDA frameworks are largely designed for image data, rendering them ineffective for capturing time series–specific features. To address these issues, we introduce **AutoDA-Timeseries**, the first general-purpose automated data augmentation framework tailored for time series. AutoDA-Timeseries incorporates time series features into augmentation policy design and adaptively optimizes both augmentation probability and intensity in a single-stage, end-to-end manner. We conduct extensive experiments on five mainstream tasks, including classification, long-term forecasting, short-term forecasting, regression, and anomaly detection, showing that AutoDA-Timeseries consistently outperforms strong baselines across diverse models and datasets.

For a two-player imperfect-information extensive-form game (IIEFG) with $K$ time steps and a player action space of size $U$, the game tree complexity is $U^{2K}$, causing existing IIEFG solvers to struggle with large or infinite $(U,K)$, e.g., differential games with continuous action spaces. To partially address this scalability challenge, we focus on an important class of 2p0s games where the informed player (P1) knows the payoff while the uninformed player (P2) only has a belief over the set of $I$ possible payoffs. Such games encompass a wide range of scenarios in sports, defense, cybersecurity, and finance. We prove that under mild conditions, P1's (resp. P2's) equilibrium strategy at any infostate concentrates on at most $I$ (resp. $I+1$) action prototypes. When $I\ll U$, this equilibrium structure causes the game tree complexity to collapse to $I^K$ for P1 when P2 plays pure best responses, and $(I+1)^K$ for P2 in a dual game where P1 plays pure best responses. We then show that exploiting this structure in standard learning modes, i.e., model-free multiagent reinforcement learning and model predictive control, is straightforward, leading to significant improvements in learning accuracy and efficiency from SOTA IIEFG solvers. Our demonstration solves a 22-player football game ($K=10$, $U=\infty$) where the attacking team has to strategically conceal their intention until a critical moment in order to exploit information advantage. Code is available [here](https://anonymous.4open.science/r/iclr_2026).

Recently, breakthroughs in the video diffusion transformer have shown remarkable capabilities in diverse motion generations. As for the motion-transfer task, current methods mainly use two-stage Low-Rank Adaptations (LoRAs) finetuning to obtain better performance. However, existing adaptation-based motion transfer still suffers from **motion inconsistency** and **tuning inefficiency** when applied to large video diffusion transformers. Naive two-stage LoRA tuning struggles to maintain motion consistency between generated and input videos due to the inherent spatial-temporal coupling in the 3D attention operator. In addition, they require time-consuming fine-tuning processes in both stages. To tackle these issues, we propose EffiVMT, an efficient **three-stage** video motion transfer framework that finetunes a powerful video diffusion transformer to synthesize complex motion. In **stage 1**, we propose a spatial-temporal head classification technique to decouple the heads of 3D attention to distinct groups for spatial-appearance and temporal motion processing. We then finetune the spatial heads in the **stage 2**. In the **stage 3** of temporal head tuning, we design the sparse motion sampling and adaptive RoPE to accelerate the tuning speed. To address the lack of a benchmark for this field, we introduce MotionBench, a comprehensive benchmark comprising diverse motion, including creative camera motion, single object motion, multiple object motion, and complex human motion. We show extensive evaluations on MotionBench to verify the superiority of EffiVMT.

3D Gaussian Splatting (3DGS) has emerged as a powerful representation for high-quality rendering across a wide range of applications. However, its high computational demands and large storage costs pose significant challenges for deployment on mobile devices. In this work, we propose a mobile-tailored real-time Gaussian Splatting method, dubbed Mobile-GS, enabling efficient inference of Gaussian Splatting on edge devices. Specifically, we first identify alpha blending as the primary computational bottleneck, since it relies on the time-consuming Gaussian depth sorting process. To solve this issue, we propose a depth-aware order-independent rendering scheme that eliminates the need for sorting, thereby substantially accelerating rendering. Although this order-independent rendering improves rendering speed, it may introduce transparency artifacts in regions with overlapping geometry due to the scarcity of rendering order. To address this problem, we propose a neural view-dependent enhancement strategy, enabling more accurate modeling of view-dependent effects conditioned on viewing direction, 3D Gaussian geometry, and appearance attributes. In this way, Mobile-GS can achieve both high-quality and real-time rendering. Furthermore, to facilitate deployment on memory-constrained mobile platforms, we propose first-degree spherical harmonics distillation, a neural vector quantization technique, and a contribution-based pruning strategy to reduce the number of Gaussian primitives and compress the 3D Gaussian representation with the assistance of neural networks. Extensive experiments demonstrate that our proposed Mobile-GS achieves real-time rendering and compact model size while preserving high visual quality, making it well-suited for mobile applications.

Temporal reasoning over long, multi-session dialogues is a critical capability for conversational agents. As dialogue histories grow in length and accumulate noise, existing long-context models struggle to accurately identify temporally pertinent information, significantly impairing reasoning performance. To address this, we introduce **Memory-T1**, a framework that learns a time-aware memory selection policy using reinforcement learning (RL). It employs a coarse-to-fine strategy, first pruning the dialogue history into a candidate set with temporal and retriever filters, followed by an RL agent that selects the precise evidence. The RL training is guided by a multi-level reward function optimizing (i) accuracy, (ii) evidence grounding, and (iii) temporal consistency. This temporal consistency reward provides a dense signal by evaluating alignment at both the session-level (range proximity) and the utterance-level (evidence density), enabling the agent to resolve subtle chronological ambiguities. On the Time-Dialog benchmark, Memory-T1 boosts a 7B model to an overall score of 67.0\%, establishing a new state-of-the-art performance for open-source models and outperforming a 14B baseline by 10.2\%. Ablation studies show temporal consistency and evidence grounding rewards jointly contributing to a 15.0\% performance gain.Moreover, Memory-T1 maintains robustness up to 128k tokens, where baseline models collapse, proving effectiveness against noise in extensive dialogue histories.

Continual learning in large language models (LLMs) is prone to catastrophic forgetting, where adapting to new tasks significantly degrades performance on previously learned ones. Existing parameter-efficient methods often limit model expressivity or introduce new parameters per task, creating scalability issues. To address these limitations, we introduce **Orthogonal Subspace Fine-Tuning (OSFT)**, a novel parameter-efficient approach for continual learning. OSFT leverages adaptive singular value decomposition (SVD) to dynamically identify and preserve critical, high-rank parameter subspaces that encode prior knowledge. All updates for new tasks are constrained to be strictly orthogonal to these preserved subspaces, which minimizes interference while maintaining a fixed parameter count and avoiding the need to store task-specific gradients. We extensively evaluate OSFT on standard continual learning benchmarks using both encoder-decoder (T5-Large) and decoder-only (LLaMA-2 7B, Mistral-7B) models across diverse tasks. Empirically, our method achieves a state-of-the-art trade-off between learnability and knowledge retention, dominating the Pareto frontier, with **up to 7\% higher** average accuracy than recent baselines like O-LoRA, and **reduces forgetting to near-negligible levels**. It notably maintains the model's general linguistic capabilities, instruction-following, and safety throughout the learning process. OSFT provides a practical, theoretically grounded, and scalable solution that effectively balances model plasticity and knowledge retention for continual learning in LLMs.

Unlearning in Large Language Models (LLMs) is crucial for protecting private data and removing harmful knowledge. Most existing approaches rely on fine-tuning to balance unlearning efficiency with general language capabilities. However, these methods typically require training or access to retain data, which is often unavailable in real world scenarios. Although these methods can perform well when both forget and retain data are available, few works have demonstrated equivalent capability in more practical, data-limited scenarios. To overcome these limitations, we propose Detect-Reasoning Augmented GeneratiON (DRAGON), a systematic, reasoning-based framework that utilizes in-context chain-of-thought (CoT) instructions to guard deployed LLMs before inference. Instead of modifying the base model, DRAGON leverages the inherent instruction-following ability of LLMs and introduces a lightweight detection module to identify forget-worthy prompts without any retain data. These are then routed through a dedicated CoT guard model to enforce safe and accurate in-context intervention. To robustly evaluate unlearning performance, we introduce novel metrics for unlearning performance and the continual unlearning setting. Extensive experiments across three representative unlearning tasks validate the effectiveness of DRAGON, demonstrating its strong unlearning capability, scalability, and applicability in practical scenarios.

The professionalism of a human doctor in outpatient service depends on two core abilities: the ability to make accurate medical decisions and the medical consultation skill to conduct strategic, empathetic patient inquiry. Existing Large Language Models (LLMs) have achieved remarkable accuracy on medical decision-making benchmarks. However, they often lack the ability to conduct the strategic and empathetic consultation, which is essential for real-world clinical scenarios. To address this gap, we propose Doctor-R1, an AI doctor agent trained to master both of the capabilities by ask high-yield questions and conduct strategic multi-turn inquiry to guide decision-making. Our framework introduces three key components: a multi-agent interactive environment, a two-tiered reward architecture that separately optimizes clinical decision-making and communicative inquiry skills, and an experience repository to ground policy learning in high-quality prior trajectories. We evaluate Doctor-R1 on OpenAI's HealthBench and MAQuE, assessed across multi-facet metrics, such as communication quality, user experience, and task accuracy. Remarkably, Doctor-R1 surpasses state-of-the-art open-source specialized LLMs by a substantial margin with higher parameter efficiency and outperforms powerful proprietary models. Furthermore, the human evaluations show a strong preference for Doctor-R1 to generate human-preferred clinical dialogue, demonstrating the effectiveness of the framework.