ICLR 2026 论文列表

Articulated object manipulation is essential for various real-world robotic tasks, yet generalizing across diverse objects remains a major challenge. A key to generalization lies in understanding functional parts (e.g., door handles and knobs), which indicate where and how to manipulate across diverse object categories and shapes. Previous works attempted to achieve generalization by introducing foundation features, while these features are mostly 2D-based and do not specifically consider functional parts. When lifting these 2D features to geometry-profound 3D space, challenges arise, such as long runtimes, multi-view inconsistencies, and low spatial resolution with insufficient geometric information. To address these issues, we propose \textbf{Part-Aware 3D Feature Field (PA3FF)}, a novel dense 3D feature with part awareness for generalizable articulated object manipulation. PA3FF is trained by 3D part proposals from a large-scale labeled datasets, via a contrastive learning formulation. Given point clouds as input, PA3FF predicts a continuous 3D feature field in a feedforward manner, where the distance between point feature reflects the proximity of functional parts: points with similar features are more likely to belong to the same part. Building on this feature, we introduce the \textbf{Part-Aware Diffusion Policy (PADP)}, an imitation learning framework aimed at enhancing sample efficiency and generalization for robotic manipulation. We evaluate PADP on several simulated and real-world tasks, demonstrating that PA3FF consistently outperforms a range of 2D and 3D representations in manipulation scenarios, including CLIP, DINOv2, and Grounded-SAM, achieving state-of-the-art performance. Beyond imitation learning, PA3FF enables diverse downstream methods, including correspondence learning and segmentation task, making it a versatile foundation for robotic manipulation.

Calibration is a critical requirement for reliable probabilistic prediction, especially in high-risk applications. However, the theoretical understanding of which learning algorithms can simultaneously achieve high accuracy and good calibration remains limited, and many existing studies provide empirical validation or a theoretical guarantee in restrictive settings. To address this issue, in this work, we focus on the smooth calibration error (CE) and provide a uniform convergence bound, showing that the smooth CE is bounded by the sum of the smooth CE over the training dataset and a generalization gap. We further prove that the functional gradient of the loss function can effectively control the training smooth CE. Based on this framework, we analyze three representative algorithms: gradient boosting trees, kernel boosting, and two-layer neural networks. For each, we derive conditions under which both classification and calibration performances are simultaneously guaranteed. Our results offer new theoretical insights and practical guidance for designing reliable probabilistic models with provable calibration guarantees.

Reward models are key to language model post-training and inference pipelines. Conveniently, recent work showed that every language model defines an implicit reward model (IM-RM), without requiring any architectural changes. However, such IM-RMs tend to generalize worse, especially out-of-distribution, compared to explicit reward models (EX-RMs) that apply a dedicated linear head over the hidden representations of a language model. The existence of a generalization gap is puzzling, as EX-RMs and IM-RMs are nearly identical. They can be trained using the same data, loss function, and language model, and differ only in how the reward is computed. Toward a fundamental understanding of the implicit biases underlying different reward model types, we investigate the root cause of this gap. Our main finding, backed by theory and experiments, is that IM-RMs rely more heavily on superficial token-level cues. Consequently, they often generalize worse than EX-RMs under token-level distribution shifts, as well as in-distribution. Furthermore, we provide evidence against alternative hypotheses for the generalization gap. Most notably, we challenge the claim that IM-RMs struggle in tasks where generation is harder than verification because they can operate both as a verifier and a generator. Overall, our results highlight that seemingly minor design choices can substantially impact the generalization behavior of reward models.

Large Language Models (LLMs) have demonstrated impressive generalization ability by learning from extensive unlabeled text. However, they still exhibit reasoning mistakes, which can affect their trustworthiness and reliability. Although users can interact with LLMs and provide diverse and comprehensive queries to expose the flaws of LLMs, obtaining sufficient and effective feedback is demanding. Furthermore, comprehensively evaluating LLMs with limited labeled samples is difficult. These make it a challenge to diagnose and remedy the deficiencies in LLMs through rich label-free user queries. To tackle this challenge and considersing that LLMs' reasoning mistakes often stem from knowledge deficiencies, we propose label-free curricular meaningful learning (LaMer), which first employs relative entropy to diagnose and quantify knowledge deficiencies of LLMs in a label-free setting. Then, LaMer adaptively synthesizes augmentation data based on deficiency severity and progressively remedies them with a curricular remedy strategy. Experiments show that LaMer effectively diagnoses and remedies knowledge deficiencies in LLMs, improving various LLMs across seven out-of-distribution (OOD) reasoning benchmarks, achieving comparable results to baselines with only 40% training data. LaMer even surpasses methods that rely on labeled data for deficiency diagnosis. In application, LaMer offers a diagnostic tool for efficient LLM development.

With the rapid scaling of large language models, achieving efficient compression while maintaining model performance has become a critical challenge. To address the limitations of existing non-uniform quantization methods, which typically rely on fixed codebooks and require costly optimization, we propose a novel arbitrary bit-width non-uniform Quantization (NuBitQ). The framework enables flexible, layer-specific quantization strategies, significantly enhancing adaptability and efficiency. Notably, traditional outlier compensation methods used in uniform quantization are ill-suited for the anomalous distribution characteristics encountered in our context. To address this, we design a novel outlier evaluation metric that integrates weight perturbation, activation distribution, and perturbation propagation. Based on this metric, we further develop an Outlier Compensation Plugin (OCP) that implements multi-level, fine-grained outlier compensation strategies, effectively mitigating performance degradation caused by outliers. Our approach avoids direct complex Hessian computation and fine-tuning, offering strong applicability and scalability. Extensive experiments on multiple tasks and across various model series demonstrate the effectiveness of the proposed approach.

Counterfactual reasoning has emerged as a crucial technique for generalizing the reasoning capabilities of large language models (LLMs). By generating and analyzing counterfactual scenarios, researchers can assess the adaptability and reliability of model decision-making. Although prior work has shown that LLMs often struggle with counterfactual reasoning, it remains unclear which factors most significantly impede their performance across different tasks and modalities. In this paper, we propose a decompositional strategy that breaks down the counterfactual generation from causality construction to the reasoning over counterfactual interventions. To support decompositional analysis, we investigate 11 datasets spanning diverse tasks, including natural language understanding, mathematics, programming, and vision-language tasks. Through extensive evaluations, we characterize LLM behavior across each decompositional stage and identify how modality type and intermediate reasoning influence performance. By establishing a structured framework for analyzing counterfactual reasoning, this work contributes to the development of more reliable LLM-based reasoning systems and informs future elicitation strategies.

A fundamental challenge in multi-task reinforcement learning (MTRL) is achieving sample efficiency in visual domains where tasks exhibit significant heterogeneity in both observations and dynamics. Model-based RL (MBRL) offers a promising path to sample efficiency through world models, but standard monolithic architectures struggle to capture diverse task dynamics, leading to poor reconstruction and prediction accuracy. We introduce the Mixture-of-World Models (MoW), a scalable architecture that integrates three key components: i) modular VAEs for task-adaptive visual compression, ii) a hybrid Transformer-based dynamics model combining task-conditioned experts with a shared backbone, and, iii) a gradient-based task clustering strategy for efficient parameter allocation. On the Atari 100k benchmark, \textbf{a single MoW agent} (trained once over Atari $26$ games) achieves a mean human-normalized score of $\mathbf{110.4}$%, competitive with the $\mathbf{114.2}$% achieved by the recent STORM—an ensemble of $26$ task-specific models—while requiring $50$% fewer parameters. On Meta-World, MoW attains a $\mathbf{74.5}$% average success rate within 300k steps, establishing a new state-of-the-art. These results demonstrate that MoW provides a scalable and parameter-efficient foundation for generalist world models. Our code is available in the supplementary materials.

Learning PDE dynamics with neural solvers can significantly improve wall-clock efficiency and accuracy compared with classical numerical solvers. In recent years, foundation models for PDEs have largely adopted multi-scale windowed self-attention, with the scOT backbone in Poseidon serving as a representative example. However, because of their locality, truly globally consistent spectral coupling can only be propagated gradually through deep stacking and window shifting. This weakens global coupling and leads to error accumulation and drift during closed-loop rollouts. To address this, we propose DRIFT-Net. It employs a dual-branch design comprising a spectral branch and an image branch. The spectral branch is responsible for capturing global, large-scale low-frequency information, whereas the image branch focuses on local details and nonstationary structures. Specifically, we first perform controlled, lightweight mixing within the low-frequency range. Then we fuse the spectral and image paths at each layer via bandwise weighting, which avoids the width inflation and training instability caused by naive concatenation. The fused result is transformed back into the spatial domain and added to the image branch, thereby preserving both global structure and high-frequency details across scales. Compared with strong attention-based baselines, DRIFT-Net achieves lower error and higher throughput with fewer parameters under identical training settings and budget. On Navier--Stokes benchmarks, the relative $L_{1}$ error is reduced by 7%-54%, the parameter count decreases by about 15%, and the throughput remains higher than scOT. Ablation studies and theoretical analyses further demonstrate the stability and effectiveness of this design. The code is available at https://anonymous.4open.science/r/DRIFT-Net-26D6.

Spiking Neural Networks (SNNs) represent a promising paradigm for energy-efficient neuromorphic computing due to their bio-plausible and spike-driven characteristics. However, the robustness of SNNs in complex adversarial environments remains significantly constrained. In this study, we theoretically demonstrate that those threshold-neighboring spiking neurons are the key factors limiting the robustness of directly trained SNNs. We find that these neurons set the upper limits for the maximum potential strength of adversarial attacks and are prone to state-flipping under minor disturbances. To address this challenge, we propose a Threshold Guarding Optimization (TGO) method, which comprises two key aspects. First, we incorporate additional constraints into the loss function to move neurons' membrane potentials away from their thresholds. It increases SNNs' gradient sparsity, thereby reducing the theoretical upper bound of adversarial attacks. Second, we introduce noisy spiking neurons to transition the neuronal firing mechanism from deterministic to probabilistic, decreasing their state-flipping probability due to minor disturbances. Extensive experiments conducted in standard adversarial scenarios prove that our method significantly enhances the robustness of directly trained SNNs. These findings pave the way for advancing more reliable and secure neuromorphic computing in real-world applications.

Sharpness-Aware Minimization (SAM) enhances generalization by minimizing the maximum training loss within a predefined neighborhood around the parameters. However, its practical implementation approximates this as gradient ascent(s) followed by applying the gradient at the ascent point to update the current parameters. This practice can be justified as approximately optimizing the objective by neglecting the (full) derivative of the ascent point with respect to the current parameters. Nevertheless, a direct and intuitive understanding of why using the gradient at the ascent point to update the current parameters works superiorly, despite being computed at a shifted location, is still lacking. Our work bridges this gap by proposing a novel and intuitive interpretation. We show that the gradient at the single-step ascent point, when applied to the current parameters, provides a better approximation of the direction from the current parameters toward the maximum within the local neighborhood than the local gradient. This improved approximation thereby enables a more direct escape from the maximum within the local neighborhood. Nevertheless, our analysis further reveals two issues. First, the approximation by the gradient at the single-step ascent point is often inaccurate. Second, the approximation quality may degrade as the number of ascent steps increases. To address these limitations, we propose in this paper eXplicit Sharpness-Aware Minimization (XSAM). It tackles the first by explicitly estimating the direction of the maximum during training, and addresses the second by crafting a search space that effectively leverages the gradient information at the multi-step ascent point. XSAM features a unified formulation that applies to both single-step and multi-step settings and only incurs negligible computational overhead. Extensive experiments demonstrate the consistent superiority of XSAM against existing counterparts across various models, datasets, and settings.

Deploying vision-and-language navigation (VLN) agents requires adaptation across diverse scenes and environments, but fine-tuning on a specific scenario often causes catastrophic forgetting in others, which severely limits flexible long-term deployment. We formalize this challenge as the all-day multi-scenes lifelong VLN (AML-VLN) problem. Existing parameter-efficient adapters (e.g., LoRA and its variants) are limited by their two-dimensional matrix form, which fails to capture the multi-hierarchical navigation knowledge spanning multiple scenes and environments. To address this, we propose Tucker Adaptation (TuKA), which represents the multi-hierarchical navigation knowledge as a high-order tensor and leverages Tucker decomposition to decouple the knowledge into shared subspaces and scenario-specific experts. We further introduce a decoupled knowledge incremental learning strategy to consolidate shared subspaces while constraining specific experts for decoupled lifelong learning. Building on TuKA, we also develop a VLN agent named AlldayWalker, which continually learns across multiple navigation scenarios, achieving all-day multi-scenes navigation. Extensive experiments show that AlldayWalker consistently outperforms state-of-the-art baselines.

Humans often juggle multiple, sometimes conflicting objectives and shift their priorities as circumstances change, rather than following a fixed objective function. In contrast, most computational decision-making and multi-objective RL methods assume static preference weights or a known scalar reward. In this work, we study sequential decision-making problem when these preference weights are unobserved latent variables that drift with context. Specifically, we propose Dynamic Preference Inference (DPI), a cognitively inspired framework in which an agent maintains a probabilistic belief over preference weights, updates this belief from recent interaction, and conditions its policy on inferred preferences. We instantiate DPI as a variational preference inference module trained jointly with a preference-conditioned actor–critic, using vector-valued returns as evidence about latent trade-offs. In queueing, gridworld maze, and multi-objective continuous-control environments with event-driven changes in objectives, DPI adapts its inferred preferences to new regimes and achieves higher post-shift performance than fixed-weight and heuristic envelope baselines.

Noise Contrastive Estimation (NCE) has fueled major breakthroughs in representation learning and generative modeling. Yet a long-standing challenge remains: accurately estimating ratios between distributions that differ substantially, which significantly limits the applicability of NCE on modern high-dimensional and multimodal datasets. We revisit this problem from a less explored perspective: the magnitude of the noise distribution. Specifically, we show that with a virtually scaled (i.e., artificially increased) noise magnitude, the gradient of the NCE objective can closely align with that of Maximum Likelihood, enabling a trajectory-wise approximation from NCE to MLE, and faster convergence both theoretically and empirically. Building on this insight, we introduce "Noisier" NCE, a simple drop-in modification to vanilla NCE that incurs little to no extra computational cost, while effectively handling density-ratio estimation in challenging regimes where traditional MLE and NCE struggle. Beyond improving classical density-ratio learning, "Noisier" NCE proves broadly applicable: it achieves strong results across image modeling, anomaly detection, and offline black-box optimization. On CIFAR-10 and ImageNet64×64 datasets, it yields 10-step and even 1-step samplers that match or surpass state-of-the-art methods, while cutting training iterations by up to half.

Recent advances in deep-research agents have shown promise for autonomous knowledge construction through dynamic reasoning over external sources. However, existing approaches rely on a mono-contextual paradigm that accumulates all information in a single, expanding context window, leading to context suffocation and noise contamination that limit their effectiveness on long-horizon tasks. We introduce IterResearch, a novel iterative deep-research paradigm that reformulates long-horizon research as a Markov Decision Process with strategic workspace reconstruction. By maintaining an evolving report as memory and periodically synthesizing insights, our approach preserves consistent reasoning capacity across arbitrary exploration depths. We further develop Efficiency-Aware Policy Optimization (EAPO), a reinforcement learning framework that incentivizes efficient exploration through geometric reward discounting and enables stable distributed training via adaptive downsampling. Extensive experiments demonstrate that IterResearch achieves substantial improvements over existing open-source agents with average +14.5pp across six benchmarks and narrows the gap with frontier proprietary systems. Remarkably, our paradigm exhibits unprecedented interaction scaling, extending to 2048 interactions with dramatic performance gains (from 3.5\% to 42.5\%), and serves as an effective prompting strategy, improving frontier models by up to 19.2pp over ReAct on long-horizon tasks. These findings position IterResearch as a versatile solution for long-horizon reasoning, effective both as a trained agent and as a prompting paradigm for frontier models.

For overparameterized linear regression with isotropic Gaussian design and minimum-$\ell_p$ interpolator $p\in(1,2]$, we give a unified, high-probability characterization for the scaling of the family of parameter norms $ \\{ \lVert \widehat{w_p} \rVert_r \\}_{r \in [1,p]} $ with sample size. We solve this basic, but unresolved question through a simple dual-ray analysis, which reveals a competition between a signal *spike* and a *bulk* of null coordinates in $X^\top Y$, yielding closed-form predictions for (i) a data-dependent transition $n_\star$ (the "elbow"), and (ii) a universal threshold $r_\star=2(p-1)$ that separates $\lVert \widehat{w_p} \rVert_r$'s which plateau from those that continue to grow with an explicit exponent. This unified solution resolves the scaling of *all* $\ell_r$ norms within the family $r\in [1,p]$ under $\ell_p$-biased interpolation, and explains in one picture which norms saturate and which increase as $n$ grows. We then study diagonal linear networks (DLNs) trained by gradient descent. By calibrating the initialization scale $\alpha$ to an effective $p_{\mathrm{eff}}(\alpha)$ via the DLN separable potential, we show empirically that DLNs inherit the same elbow/threshold laws, providing a predictive bridge between explicit and implicit bias. Given that many generalization proxies depend on $\lVert \widehat {w_p} \rVert_r$, our results suggest that their predictive power will depend sensitively on which $l_r$ norm is used.

Modern language models define distributions over strings, but their outputs are not always suited to downstream task. For instance, a model generating byte-pair strings may not be suitable when word-level predictions are needed, and a DNA model may not fit applications requiring amino acids. In such cases, a deterministic string-to-string transformation can convert the model's output to the desired form. This is a familiar pattern in probability theory: applying a function $f$ to a random variable $X\sim p$ yields a transformed random variable $f(X)$ with an induced distribution. While such transformations are occasionally used in language modeling, they are not treated as yielding new, fully functional language models. We formalize this perspective and introduce a general framework for language models derived from deterministic string-to-string transformations. Focusing on transformations representable as finite-state transducers---a commonly used state-machine abstraction for efficient string-to-string mappings---we develop algorithms that compose a language model with an FST to *marginalize* over source strings mapping to a given target. This allows us to propagate probabilities through the transducer without altering model parameters and to *condition* on transformed outputs. We present an exact algorithm, an efficient approximation, and a theoretical analysis. We conduct experiments in three domains: converting token-level language models to character-level language models, token-level language models to word-level models, and deriving amino-acid models from DNA models. This demonstrates inference-time adaptation of pretrained language models to match application-specific output requirements.

Assessing Large Language Models (LLMs)' underlying value differences enables comprehensive comparison of their misalignment, cultural adaptability, and biases. Nevertheless, current value measurement methods face the informativeness challenge: with often outdated, contaminated, or generic test questions, they can only capture the orientations on comment safety values, e.g., HHH, shared among different LLMs, leading to indistinguishable and uninformative results. To address this problem, we introduce AdAEM, a novel, self-extensible evaluation algorithm for revealing LLMs' inclinations. Distinct from static benchmarks, AdAEM automatically and adaptively generates and extends its test questions. This is achieved by probing the internal value boundaries of a diverse set of LLMs developed across cultures and time periods in an in-context optimization manner. Such a process theoretically maximizes an information-theoretic objective to extract diverse controversial topics that can provide more distinguishable and informative insights about models' value differences. In this way, AdAEM is able to co-evolve with the development of LLMs, consistently tracking their value dynamics. We use AdAEM to generate novel questions and conduct an extensive analysis, demonstrating our method's validity and effectiveness, laying the groundwork for better interdisciplinary research on LLMs' values and alignment.

Language models obtain extensive capabilities through pre-training. However, the pre-training dynamics remains a black box. In this work, we track linear interpretable feature evolution across pre-training snapshots using a sparse dictionary learning method called crosscoders. We find that most features begin to form around a specific point, while more complex patterns emerge in later training stages. Feature attribution analyses reveal causal connections between feature evolution and downstream performance. Our feature-level observations are highly consistent with previous findings on Transformer's two-stage learning process, which we term a statistical learning phase and a feature learning phase. Our work opens up the possibility to track fine-grained representation progress during language model learning dynamics. Our code is available at https://github.com/OpenMOSS/Language-Model-SAEs.

Reinforcement learning (RL) with group relative policy optimization (GRPO) has become a widely adopted approach for enhancing the reasoning capabilities of multimodal large language models (MLLMs). While GRPO enables long-chain reasoning without a traditional critic model, it often suffers from sparse rewards, arising from the scarcity of positive feedback on difficult problems, and from advantage vanishing, which occurs when group-level rewards exhibit high consistency for problems that are too easy or too hard. Existing solutions fall into three categories: sample enhancement and expansion, which may aggravate vanishing advantage due to poor control of difficulty distribution; selective sample utilization, which fails to fully leverage the value of all data; and indirect reward design, which may introduce biased optimization directions due to misalignment between reasoning and the final outcome. However, these approaches overlook a fundamental question: for a given problem, how can we ensure that the within-group reward distribution of responses exhibits enough variance to yield clear optimization signals for each response? To address these issues, we propose DIVA-GRPO, a difficulty-adaptive variant augmentation advantage method that dynamically adjusts the difficulty distribution of variants for each problem from a global perspective. Our method dynamically assesses problem difficulty, samples variants with appropriate difficulty levels, and advantages are computed within both local and global(a problem and its variants) groups using difficulty-weighted and normalized scaling. This design alleviates reward sparsity and advantage vanishing, minimizes data waste, and improves training stability. Extensive experiments on six reasoning benchmarks demonstrate that DIVA-GRPO outperforms existing approaches in both training efficiency and reasoning performance.