ICLR 2026 Papers

Existing reinforcement learning (RL) methods struggle with complex dynamical systems that demand interactions at high frequencies or irregular time intervals. Continuous-time RL (CTRL) has emerged as a promising alternative by replacing discrete-time Bellman recursion with differentiable value functions defined as viscosity solutions of the Hamilton–Jacobi–Bellman (HJB) equation. While CTRL has shown promise, its applications have been largely limited to the single-agent domain. This limitation stems from two key challenges: (i) conventional methods for solving HJB equations suffer from the curse of dimensionality (CoD), making them intractable in high-dimensional systems; and (ii) even with learning-based approaches to alleviate the CoD, accurately approximating centralized value functions in multi-agent settings remains difficult, which in turn destabilizes policy training. In this paper, we propose a CT-MARL framework that uses physics-informed neural networks (PINNs) to approximate HJB-based value functions at scale. To ensure the value is consistent with its differential structure, we align value learning with value-gradient learning by introducing a Value Gradient Iteration (VGI) module that iteratively refines value gradients along trajectories. This improves gradient accuracy, in turn yielding more precise value approximations and stronger policy learning. We evaluate our method using continuous‑time variants of standard benchmarks, including multi‑agent particle environment (MPE) and multi‑agent MuJoCo. Our results demonstrate that our approach consistently outperforms existing continuous‑time RL baselines and scales to complex cooperative multi-agent dynamics.

Generative AI excels in various tasks through advanced language modeling techniques, with its performance heavily influenced by input prompts. This has driven significant research into prompt optimization, particularly in commercial generative AI platforms, where prompt optimization is treated as a black-box optimization problem. Most existing research on black-box prompt optimization primarily focuses on offline learning and overlooks the randomness in outputs. However, in real-world applications, black-box prompt optimization typically operates in an online learning setting, which remains largely unexplored, especially given the noisy outputs. To address these challenges, we propose an \textbf{A}daptive \textbf{O}nline \textbf{Z}eroth-order \textbf{P}rompt \textbf{T}uning (AOZPT) approach which integrates zeroth-order optimization with online learning in the non-convex setting. Specifically, we developed an uncertainty-scale-adjustment mechanism to mitigate the noise inherent in generative AI and the high variance associated with zeroth-order estimates. We conducted a comprehensive regret analysis of the AOZPT approach, and the results indicate that sublinear regret convergence is achievable. Extensive generative experiments demonstrate that AOZPT outperforms existing black-box prompt tuning methods, particularly in terms of stability in online scenarios.

Dynamic feature transformation (the rich regime) does not always align with predictive performance (better representation), yet accuracy is often used as a proxy for richness, limiting analysis of their relationship. We propose a computationally efficient, performance-independent metric of richness grounded in the low-rank bias of rich dynamics, which recovers neural collapse as a special case. The metric is empirically more stable than existing alternatives and captures known lazy-to-rich transitions (e.g., grokking) without relying on accuracy. We further use it to examine how training factors (e.g., learning rate) relate to richness, confirming recognized assumptions and highlighting new observations (e.g., batch normalization promote rich dynamics). An eigendecomposition-based visualization is also introduced to support interpretability, together providing a diagnostic tool for studying the relationship between training factors, dynamics, and representations.

Vision-Language-Action models (VLAs) are emerging as powerful tools for learning generalizable visuomotor control policies. However, current VLAs are mostly trained on large-scale image–text–action data and remain limited in two key ways: (i) they struggle with pixel-level scene understanding, and (ii) they rely heavily on textual prompts, which reduces their flexibility in real-world settings. To address these challenges, we introduce PixelVLA, the first VLA model designed to support both pixel-level reasoning and multimodal prompting with text and visual inputs. Our approach is built on a new visuomotor instruction tuning framework that integrates a multiscale pixel-aware encoder with a visual prompting encoder. To train PixelVLA effectively, we further propose a two-stage automated annotation pipeline that generates Pixel-160K, a large-scale dataset with pixel-level annotations derived from existing robot data. Experiments on three standard VLA benchmarks and two VLA model variants show that PixelVLA improves manipulation success rates by $10.1\%\sim28.7\%$ over OpenVLA, while requiring only $1.5\%$ of its pretraining cost. These results demonstrate that PixelVLA can be integrated into existing VLAs to enable more accurate, efficient, and versatile robot control in complex environments. The dataset and code will be released as open source.

Protein representation learning has achieved major advances using large sequence and structure datasets, yet current models primarily operate at the level of individual residues or entire proteins. This overlooks a critical aspect of protein biology: proteins are composed of recurrent, evolutionarily conserved substructures that mediate core molecular functions. Despite decades of curated biological knowledge, these substructures remain largely unexploited in modern protein models. We introduce Magneton, an integrated environment for developing substructure-aware protein models. Magneton provides (1) a large-scale dataset of 530,601 proteins annotated with over 1.7 million substructures spanning 13,075 types, (2) a training framework for incorporating substructures into existing models, and (3) a benchmark suite of 13 tasks probing residue-, substructure-, and protein-level representations. Using Magneton, we develop substructure-tuning, a supervised fine-tuning method that distills substructural knowledge into pretrained protein models. Across state-of-the-art sequence- and structure-based models, substructure-tuning improves function-related tasks while revealing that substructural signals are complementary to global structural information. The Magneton environment, datasets, and substructure-tuned models are all openly available.

Unsupervised Multimodal anomaly detection (MAD) — identifying defects by jointly analyzing RGB images and 3D data — is crucial for quality control in manufacturing. However, existing MAD methods struggle when only a few normal samples are available. Cross-modal alignment models fail to learn stable correspondences with scarce training data, and memory-based approaches misclassify any unseen normal variation as anomalous.To addresses the few-shot challenge, we introduce PIRN (Prototypical-based Intra-modal Reconstruction with Normality Communication for Multi-modal Anomaly Detection.), a prototype-based intra-modal reconstruction framework with explicit cross-modal knowledge transfer. PRINC features three key innovations: (1) Balanced Prototype Assignment (BPA) formulates token‑to‑prototype routing as a balanced optimal‑transport problem, guaranteeing uniform utilisation of all prototypes and preventing codebook collapse.(2) Adaptive Prototype Refinement (APR) treats prototypes as adaptive memory and updates them on‑the‑fly with gated GRU cells driven by optimally‑matched image context, expanding coverage to unseen yet normal variations while suppressing anomalies. (3) Multi‑modal Normality Communication (MNC) exchanges complementary normal cues across modalities via gated cross‑attention. MNC enables one modality to reconstruct its feature map not only from its own prototypes, but also with high-level normal patterns provided by the other modality.Extensive experiments on standard benchmarks demonstrate that PIRN significantly outperforms prior methods, achieving new state-of-the-art results, especially in challenging few-shot scenarios.

Missing modalities present a fundamental challenge in multimodal models, often causing catastrophic performance degradation. Our observations suggest that this fragility stems from an imbalanced learning process, where the model develops an implicit preference for certain modalities, leading to the under-optimization of others. We propose a simple yet efficiency method to address this challenge. The central insight of our work is that the dominance relationship between modalities can be effectively discerned and quantified in the frequency domain. To leverage this principle, we first introduce a **F**requency **R**atio **M**etric (FRM) to quantify modality preference by analyzing features in the frequency domain. Guided by FRM, we then propose a **M**ultimodal **W**eight **A**llocation **M**odule, a plug-and-play component that dynamically re-balances the contribution of each branch during training, promoting a more holistic learning paradigm. Extensive experiments demonstrate that MWAM can be seamlessly integrated into diverse architectural backbones, such as those based on CNNs and ViTs. Furthermore, MWAM delivers consistent performance gains across a wide range of tasks and modality combinations. This advancement extends beyond merely optimizing the performance of the base model; it also manifests as further performance improvements to state-of-the-art methods addressing the missing modality problem.

Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model. We validate these findings through experiments with large language models and robotic foundation models and further provide theoretical justification for why on-policy RL updates lead to a smaller KL change. We term this principle $\textit{RL’s Razor}$: among all ways to solve a new task, RL prefers those closest in KL to the original model.

We study the problem of posterior sampling in the context of score based generative models. We have a trained score network for a prior $p(x)$, a measurement model $p(y|x)$, and are tasked with sampling from the posterior $p(x|y)$. Prior work has shown this to be intractable in KL (in the worst case) under well-accepted computational hardness assumptions. Despite this, popular algorithms for tasks such as image super-resolution, stylization, and reconstruction enjoy empirical success. Rather than establishing distributional assumptions or restricted settings under which exact posterior sampling is tractable, we view this as a more general "tilting" problem of biasing a distribution towards a measurement. Under minimal assumptions, we show that one can tractably sample from a distribution that is simultaneously close to the posterior of a noised prior in KL divergence and the true posterior in Fisher divergence. Intuitively, this combination ensures that the resulting sample is consistent with both the measurement and the prior. To the best of our knowledge these are the first formal results for (approximate) posterior sampling in polynomial time.

We study the convergence rate of learning pairwise interactions in single-layer attention-style models, where tokens interact through a weight matrix and a non-linear activation function. We prove that the minimax rate is $M^{-\frac{2\beta}{2\beta+1}}$ with $M$ being the sample size, depending only on the smoothness $\beta$ of the activation, and crucially independent of token count, ambient dimension, or rank of the weight matrix. These results highlight a fundamental dimension-free statistical efficiency of attention-style nonlocal models, even when the weight matrix and activation are not separately identifiable and provide a theoretical understanding of the attention mechanism and its training.

We present a novel differentiable grid-based representation for efficiently solving differential equations (DEs). Widely used architectures for neural solvers, such as sinusoidal neural networks, are coordinate-based MLPs that are, both, computationally intensive and slow to train. Although grid-based alternatives for implicit representations (e.g., Instant-NGP and K-Planes) train faster by exploiting signal structure, their reliance on linear interpolation restricts their ability to compute higher-order derivatives, rendering them unsuitable for solving DEs. In contrast, our approach overcomes these limitations by combining the efficiency of feature grids with radial basis function interpolation, which is infinitely often differentiable. To effectively capture high-frequency solutions and enable stable and faster computation of global gradients, we introduce a multi-resolution decomposition with co-located grids. Our proposed representation, $\boldsymbol{\partial^\infty}$-Grid, is trained implicitly using the differential equations as loss functions, enabling accurate modeling of physical fields. We validate $\boldsymbol{\partial^\infty}$-Grid on a variety of tasks, including Poisson equation for image reconstruction, the Helmholtz equation for wave fields, and the Kirchhoff-Love boundary value problem for cloth simulation. Our results demonstrate a 5–20× speed-up over coordinate-based MLP-based methods, solving differential equations in seconds or minutes while maintaining comparable accuracy and compactness.

Accurate registration of brain MRI scans is fundamental for cross-subject analysis in neuroscientific studies. This involves aligning both the cortical surface of the brain and the interior volume. Traditional methods treat volumetric and surface-based registration separately, which often leads to inconsistencies that limit downstream analyses. We propose a deep learning framework, UCS, that registers 3D brain MRI images by jointly aligning both cortical and subcortical regions, through a unified volume-and-surface-based representation. Our approach leverages an intermediate spherical coordinate space to bridge anatomical surface topology with volumetric anatomy, enabling consistent and anatomically accurate alignment. By integrating spherical registration into the learning, our method ensures geometric coherence between volume and surface domains. In a series of experiments on both in-domain and out-of-domain datasets, our method consistently outperforms both classical and machine learning-based registration methods--improving the Dice score by up to 7 points while maintaining regular deformation fields. Additionally, it is orders of magnitude faster than the standard method for this task, and is simpler to use because it requires no additional inputs beyond an MRI scan. Its superior accuracy, fast inference, and ease of use sets a new standard for joint cortical and subcortical registration.

Accurate Subseasonal-to-Seasonal (S2S) forecasting is vital for decision-making in agriculture, energy production, and emergency management. However, it remains a challenging and underexplored problem due to the chaotic nature of the weather system. Recent data-driven studies have shown promising results, but their performance is limited by the inadequate incorporation of climate states and a model tendency to degrade, progressively losing fine-scale details and yielding over-smoothed forecasts. To overcome these limitations, we propose TianQuan-S2S, a global S2S forecasting model that integrates initial weather states with climatological means via incorporating climatology into patch embedding and enhancing variability capture through an uncertainty-augmented Transformer. Extensive experiments on the Earth Reanalysis 5 (ERA5) reanalysis dataset demonstrate that our model yields a significant improvement in both deterministic and ensemble forecasting over the climatology mean, traditional numerical methods, and data-driven models. Ablation studies empirically show the effectiveness of our model designs. Remarkably, our model outperforms skillful numerical ECMWF-S2S and advanced data-driven Fuxi-S2S in key meteorological variables. The code implementation can be found in https://github.com/zhangminglang42/TianQuan.

Any-to-any generation seeks to translate between arbitrary subsets of modalities, enabling flexible cross-modal synthesis. Despite recent success, existing flow-based approaches are challenged by its inefficiency, as they require large-scale datasets often with restrictive pairing constraints, incur high computation cost from modeling joint distribution, and multi-stage training pipeline. We propose \textbf{FlowBind}, an efficient framework for any-to-any generation. Our approach is distinguished by its simplicity: it learns a shared latent space capturing cross-modal information, with modality-specific invertible flows bridging this latent to each modality. Both components are optimized jointly under a single flow-matching objective, and at inference the invertible flows act as encoders and decoders for direct translation across modalities. By factorizing interactions through the shared latent, FlowBind naturally leverages arbitrary subsets of modalities for training, and achieves competitive generation quality while substantially reducing data requirements and computational cost. Experiments on text, image, and audio demonstrate that FlowBind attains comparable quality while requiring up to 6× fewer parameters and training 10× faster than prior methods.

Active Prompt Learning (APL) using vision-language models (\textit{e.g.}, CLIP) has attracted considerable attention for mitigating the dependence on fully labeled dataset in downstream task adaptation. However, existing methods fail to explicitly leverage prompt to guide sample selection, resulting in the selected samples being ineffective in facilitating the prompt template's downstream task adaptation, while also overlooking valuable complementary information in the unselected samples. To fill this gap, we propose a novel Prompt-Guided Self-Training Sampling Policy (PSP) for APL, which integrates Soft Actor-Critic with a customized real-pseudo hybrid reward and vectorized critics to incorporate prompts in guiding sample selection toward those that facilitate the optimization of prompt template, by jointly considering both selected and unselected samples. Specifically, PSP comprises two prominent components: Vectorized Soft Actor-Critic Sampling Policy (VSSP) and Uncertainty Augmented Self-Training (UST) mechanism. VSSP customizes a real-pseudo hybrid reward based on learned prompts and image features, which is fed into vectorized critics to estimate Q-value for each sample and compute gradients that optimize the actor, allowing it to refine its sampling policy in an End-to-End manner to identify the most informative samples for prompt learning. Moreover, UST leverages the CLIP from the previous round to generate reliable pseudo-labeled data based on uncertainty and confidence of average predictions, thereby deepening the understanding of the overall data. Extensive experiments conducted on diverse real-world datasets validate the effectiveness of our PSP.

Off-policy reinforcement learning (RL) for large language models (LLMs) is attracting growing interest, driven by practical constraints in real-world applications, the complexity of LLM-RL infrastructure, and the need for further innovations of RL methodologies. While classic REINFORCE and its modern variants like Group Relative Policy Optimization (GRPO) are typically regarded as on-policy algorithms with limited tolerance of off-policyness, we present in this work a first-principles derivation for *group-relative REINFORCE* without assuming a specific training data distribution, showing that it admits a *native off-policy interpretation*. This perspective yields two general principles for adapting REINFORCE to truly off-policy settings: regularizing policy updates, and actively shaping the data distribution. Our analysis demystifies some myths about the roles of importance sampling and clipping in GRPO, unifies and reinterprets two recent algorithms --- Online Policy Mirror Descent and Asymmetric REINFORCE --- as regularized forms of the REINFORCE loss, and offers theoretical justification for seemingly heuristic data-weighting strategies. Our findings lead to actionable insights that are validated with extensive empirical studies, and open up new opportunities for principled algorithm design in off-policy RL for LLMs.

Spiking Transformer has sparked growing interest, with the Spiking Self-Attention merging spikes with self-attention to deliver both energy efficiency and competitive performance. However, existing work primarily focuses on 2D visual tasks, and in the domain of 3D point clouds, the disorder and complexity of spatial information, along with the scale of the point clouds, present significant challenges. For point clouds, we introduce spiking discrepancy, measuring differences in spike features to highlight key information, and then construct the Spiking Discrepancy Attention Mechanism (SDAM). SDAM contains two variants: the Spiking Element Discrepancy Attention captures local geometric correlations between central points and neighboring points, while the Spiking Intensity Discrepancy Attention characterizes structural patterns of point clouds based on macroscopic spike statistics. Moreover, we propose a Spatially-Aware Spiking Neuron. Based on these, we construct a hierarchical Spiking Discrepancy Transformer. Experimental results demonstrate that our method achieves state-of-the-art performance within the Spiking Neural Networks and exhibits impressive performance compared to Artificial Neural Networks along with a few parameters and significantly lower theoretical energy consumption.

Scalable and realistic simulation of multi-agent traffic behavior is critical for advancing autonomous driving technologies. Although existing data-driven simulators have made significant strides in this domain, they predominantly rely on supervised learning to align simulated distributions with real-world driving scenarios. A persistent challenge, however, lies in the distributional shift that arises between training and testing, which often undermines model generalization in unseen environments. To address this limitation, we propose SMART-R1, a novel R1-style reinforcement fine-tuning paradigm tailored for next-token prediction models to better align agent behavior with human preferences and evaluation metrics. Our approach introduces a metric-oriented policy optimization algorithm to improve distribution alignment and an iterative "SFT-RFT-SFT" training strategy that alternates between Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) to maximize performance gains. Extensive experiments on the large-scale Waymo Open Motion Dataset (WOMD) validate the effectiveness of this simple yet powerful R1-style training framework in enhancing foundation models. The results on the Waymo Open Sim Agents Challenge (WOSAC) showcase that SMART-R1 achieves state-of-the-art performance with an overall realism meta score of 0.7858, ranking first on the leaderboard at the time of submission.

Cell Painting combines multiplexed fluorescent staining, high‑content imaging, and quantitative analysis to generate high-dimensional phenotypic readouts to support diverse downstream tasks such as mechanism-of-action (MoA) inference, toxicity prediction, and construction of drug–disease atlases. However, existing workflows are slow, costly and difficult to interpret. Approaches for drug screening modeling predominantly focus on molecular representation learning, while neglecting actual experimental context (e.g., cell line, dosing schedule, etc.), limiting generalization and MoA resolution. We introduce CP-Agent, an agentic multimodal large language model (MLLM) capable of generating mechanism-relevant, human-interpretable rationales for cell morphological changes under drug perturbations. At its core, CP-Agent leverages a context-aware alignment module, CP-CLIP, that jointly embeds high-content images and experimental metadata to enable robust treatment and MoA discrimination (achieving a maximum F1-score of 0.896). By integrating CP-CLIP outputs with agentic tool usage and reasoning, CP‑Agent compiles rationales into a structured report to guide experimental design and hypothesis refinement. These capabilities highlight CP-Agent’s potential to accelerate drug discovery by enabling more interpretable, scalable, and context-aware phenotypic screening---streamlining iterative cycles of hypothesis generation in drug discovery.