NeurIPS 2025 Papers

Clinical decision‑making routinely demands reasoning over heterogeneous data, yet existing multimodal language models (MLLMs) remain largely vision‑centric and fail to generalize across clinical specialties. To bridge this gap, we introduce QoQ-Med-7B/32B, the first open generalist clinical foundation model that jointly reasons across medical images, time‑series signals, and text reports. QoQ-Med is trained with Domain‑aware Relative Policy Optimization (DRPO), a novel reinforcement‑learning objective that hierarchically scales normalized rewards according to domain rarity and modality difficulty, mitigating performance imbalance caused by skewed clinical data distributions. Trained on 2.61 million instruction tuning pairs spanning 9 clinical domains, we show that DRPO training boosts diagnostic performance by 43% in macro‑F1 on average across all visual domains as compared to other critic-free training methods like GRPO. Furthermore, with QoQ-Med trained on intensive segmentation data, it is able to highlight salient regions related to the diagnosis, with an IoU 10x higher than open models while reaching the performance of OpenAI o4-mini. To foster reproducibility and downstream research, we release (i) the full model weights, (ii) the modular training pipeline, and (iii) all intermediate reasoning traces.

Multimodal large language models (MLLMs) extend LLMs to handle images, videos, and audio by incorporating feature extractors and projection modules. However, these additional components—combined with complex inference pipelines and heterogeneous workloads—introduce significant inference overhead. Therefore, efficiently serving MLLMs remains a major challenge. Current tightly coupled serving architectures struggle to distinguish between mixed request types or adapt parallelism strategies to different inference stages, leading to increased time-to-first-token (TTFT) and poor resource utilization. To address this, we introduce Elastic Multimodal Parallelism (EMP), a new serving paradigm that elastically adapts to resource heterogeneity across request types and inference stages. Building upon EMP, we develop ElasticMM, an MLLM serving system that (1) separates requests into independent modality groups with dynamic resource allocation via a modality-aware load balancer; (2) decouples inference stages and enables parallelism adjustment and adaptive scaling via elastic partition scheduling; and (3) improves inference efficiency through unified multimodal prefix caching and non-blocking encoding. Experiments on diverse real-world datasets show that ElasticMM outperforms state-of-the-art (SOTA) serving systems, reducing TTFT by up to 4.2$\times$ and achieving 3.2–4.5$\times$ higher throughput while meeting service-level objectives (SLOs).

In domain adaption (DA), joint maximum mean discrepancy (JMMD), as a famous distribution-distance metric, aims to measure joint probability distribution difference between the source domain and target domain, while it is still not fully explored and especially hard to be applied into a subspace-learning framework as its empirical estimation involves a tensor-product operator whose partial derivative is difficult to obtain. To solve this issue, we deduce a concise JMMD based on the Representer theorem that avoids the tensor-product operator and obtains two essential findings. First, we reveal the uniformity of JMMD by proving that previous marginal, class conditional, and weighted class conditional probability distribution distances are three special cases of JMMD with different label reproducing kernels. Second, inspired by graph embedding, we observe that the similarity weights, which strengthen the intra-class compactness in the graph of Hilbert Schmidt independence criterion (HSIC), take opposite signs in the graph of JMMD, revealing why JMMD degrades the feature discrimination. This motivates us to propose a novel loss JMMD-HSIC by jointly considering JMMD and HSIC to promote discrimination of JMMD. Extensive experiments on several cross-domain datasets could demonstrate the validity of our revealed theoretical results and the effectiveness of our proposed JMMD-HSIC.

Vehicle GPS trajectories provide valuable movement information that supports various downstream tasks and applications. A desirable trajectory learning model should be able to transfer across regions and tasks without retraining, avoiding the need to maintain multiple specialized models and subpar performance with limited training data. However, each region has its unique spatial features and contexts, which are reflected in vehicle movement patterns and are difficult to generalize. Additionally, transferring across different tasks faces technical challenges due to the varying input-output structures required for each task. Existing efforts towards transferability primarily involve learning embedding vectors for trajectories, which perform poorly in region transfer and require retraining of prediction modules for task transfer. To address these challenges, we propose $\textit{TransferTraj}$, a vehicle GPS trajectory learning model that excels in both region and task transferability. For region transferability, we introduce RTTE as the main learnable module within TransferTraj. It integrates spatial, temporal, POI, and road network modalities of trajectories to effectively manage variations in spatial context distribution across regions. It also introduces a TRIE module for incorporating relative information of spatial features and a spatial context MoE module for handling movement patterns in diverse contexts. For task transferability, we propose a task-transferable input-output scheme that unifies the input-output structure of different tasks into the masking and recovery of modalities and trajectory points. This approach allows TransferTraj to be pre-trained once and transferred to different tasks without retraining. We conduct extensive experiments on three real-world vehicle trajectory datasets under various transfer settings, including task transfer, zero-shot region transfer, and few-shot region transfer. Experimental results demonstrate that TransferTraj significantly outperforms state-of-the-art baselines in different scenarios, validating its effectiveness in region and task transfer. Code is available at https://github.com/wtl52656/TransferTraj.

We sharply characterize the optimal first-order query complexity of agnostic active learning for all concept classes, and propose a new general active learning algorithm which achieves it. Remarkably, the optimal query complexity admits a leading term which is always strictly smaller than the sample complexity of passive supervised learning (by a factor proportional to the best-in-class error rate). This was not previously known to be possible in the agnostic setting. For comparison, in all previous general analyses, the leading term exhibits an additional factor, such as the disagreement coefficient or related complexity measure, and therefore only provides improvements over passive learning in restricted cases. The present work completely removes such factors from the leading term, implying that $\textit{every}$ concept class benefits from active learning in the non-realizable case. The results established in this work resolve an important long-standing open question central to the past two decades of research on the theory of agnostic active learning.

We contribute towards resolving the open question of how many hidden layers are required in ReLU networks for exactly representing all continuous and piecewise linear functions on $\mathbb{R}^d$. While the question has been resolved in special cases, the best known lower bound in general is still 2. We focus on neural networks that are compatible with certain polyhedral complexes, more precisely with the braid fan. For such neural networks, we prove a non-constant lower bound of $\Omega(\log\log d)$ hidden layers required to exactly represent the maximum of $d$ numbers. Additionally, we provide a combinatorial proof that neural networks satisfying this assumption require three hidden layers to compute the maximum of 5 numbers; this had only been verified with an excessive computation so far. Finally, we show that a natural generalization of the best known upper bound to maxout networks is not tight, by demonstrating that a rank-3 maxout layer followed by a rank-2 maxout layer is sufficient to represent the maximum of 7 numbers.

As the economic and environmental costs of training and deploying large vision or language models increase dramatically, analog in-memory computing (AIMC) emerges as a promising energy-efficient solution. However, the training perspective, especially its training dynamic, is underexplored. In AIMC hardware, the trainable weights are represented by the conductance of resistive elements and updated using consecutive electrical pulses. While the conductance changes by a constant in response to each pulse, in reality, the change is scaled by asymmetric and non-linear response functions, leading to a non-ideal training dynamic. This paper provides a theoretical foundation for gradient-based training on AIMC hardware with non-ideal response functions. We demonstrate that asymmetric response functions negatively impact Analog SGD by imposing an implicit penalty on the objective. To overcome the issue, we propose residual learning algorithm, which provably converges exactly to a critical point by solving a bilevel optimization problem. We show that the proposed method can be extended to deal with other hardware imperfections like limited response granularity. As far as we know, it is the first paper to investigate the impact of a class of generic non-ideal response functions. The conclusion is supported by simulations validating our theoretical insights.

Accurate, high-resolution ocean forecasting is crucial for maritime operations and environmental monitoring. While traditional numerical models are capable of producing sub-daily, eddy-resolving forecasts, they are computationally intensive and face challenges in maintaining accuracy at fine spatial and temporal scales. In contrast, recent data-driven approaches offer improved computational efficiency and emerging potential, yet typically operate at daily resolution and struggle with sub-daily predictions due to error accumulation over time. We introduce FuXi-Ocean, the first data-driven global ocean forecasting model achieving six-hourly predictions at eddy-resolving 1/12° spatial resolution, reaching depths of up to 1500 meters. The model architecture integrates a context-aware feature extraction module with a predictive network employing stacked attention blocks. The core innovation is the Mixture-of-Time (MoT) module, which adaptively integrates predictions from multiple temporal contexts by learning variable-specific reliability , mitigating cumulative errors in sequential forecasting. Through comprehensive experimental evaluation, FuXi-Ocean demonstrates superior skill in predicting key variables, including temperature, salinity, and currents, across multiple depths.

In this paper, we study the evolution of tokens through the depth of encoder-only transformer models at inference time by modeling them as a system of particles interacting in a mean-field way and studying the corresponding dynamics. More specifically, we consider this problem in the moderate interaction regime, where the number $N$ of tokens is large and the inverse temperature parameter $\beta$ of the model scales together with $N$. In this regime, the dynamics of the system displays a multiscale behavior: a fast phase, where the token empirical measure collapses on a low-dimensional space, an intermediate phase, where the measure further collapses into clusters, and a slow one, where such clusters sequentially merge into a single one. We provide a rigorous characterization of the limiting dynamics in each of these phases and prove convergence in the above mentioned limit, exemplifying our results with some simulations.

Accurate visual localization from aerial views is a fundamental problem with applications in mapping, large-area inspection, and search-and-rescue operations. In many scenarios, these systems require high-precision localization while operating with limited resources (e.g., no internet connection or GNSS/GPS support), making large image databases or heavy 3D models impractical. Surprisingly, little attention has been given to leveraging orthographic geodata as an alternative paradigm, which is lightweight and increasingly available through free releases by governmental authorities (e.g., the European Union). To fill this gap, we propose OrthoLoC, the first large-scale dataset comprising 16,425 UAV images from Germany and the United States with multiple modalities. The dataset addresses domain shifts between UAV imagery and geospatial data. Its paired structure enables fair benchmarking of existing solutions by decoupling image retrieval from feature matching, allowing isolated evaluation of localization and calibration performance. Through comprehensive evaluation, we examine the impact of domain shifts, data resolutions, and covisibility on localization accuracy. Finally, we introduce a refinement technique called AdHoP, which can be integrated with any feature matcher, improving matching by up to 95% and reducing translation error by up to 63%. The dataset and code are available at: https://deepscenario.github.io/OrthoLoC .

In this paper, we leverage stochastic projection and lossy compression to establish new conditional mutual information (CMI) bounds on the generalization error of statistical learning algorithms. It is shown that these bounds are generally tighter than the existing ones. In particular, we prove that for certain problem instances for which existing MI and CMI bounds were recently shown in Attias et al. [2024] and Livni [2023] to become vacuous or fail to describe the right generalization behavior, our bounds yield suitable generalization guarantees of the order of $\mathcal{O}(1/\sqrt{n})$, where $n$ is the size of the training dataset. Furthermore, we use our bounds to investigate the problem of data "memorization" raised in those works, and which asserts that there are learning problem instances for which any learning algorithm that has good prediction there exist distributions under which the algorithm must "memorize'' a big fraction of the training dataset. We show that for every learning algorithm, there exists an auxiliary algorithm that does not memorize and which yields comparable generalization error for any data distribution. In part, this shows that memorization is not necessary for good generalization.

We study the online calibration of multi-dimensional forecasts over an arbitrary convex set $\mathcal{P} \subset \mathbb{R}^d$ relative to an arbitrary norm $\Vert\cdot\Vert$. We connect this with the problem of external regret minimization for online linear optimization, showing that if it is possible to guarantee $O(\sqrt{\rho T})$ worst-case regret after $T$ rounds when actions are drawn from $\mathcal{P}$ and losses are drawn from the dual $\Vert \cdot \Vert_*$ unit norm ball, then it is also possible to obtain $\epsilon$-calibrated forecasts after $T = \exp(O(\rho /\epsilon^2))$ rounds. When $\mathcal{P}$ is the $d$-dimensional simplex and $\Vert \cdot \Vert$ is the $\ell_1$-norm, the existence of $O(\sqrt{T\log d})$ algorithms for learning with experts implies that it is possible to obtain $\epsilon$-calibrated forecasts after $T = \exp(O(\log{d}/\epsilon^2)) = d^{O(1/\epsilon^2)}$ rounds, recovering a recent result of Peng 2025. Interestingly, our algorithm obtains this guarantee without requiring access to any online linear optimization subroutine or knowledge of the optimal rate $\rho$ -- in fact, our algorithm is identical for every setting of $\mathcal{P}$ and $\Vert \cdot \Vert$. Instead, we show that the optimal regularizer for the above OLO problem can be used to upper bound the above calibration error by a swap regret, which we then minimize by running the recent TreeSwap algorithm with Follow-The-Leader as a subroutine. The resulting algorithm is highly efficient and plays a distribution over simple averages of past observations in each round. Finally, we prove that any online calibration algorithm that guarantees $\epsilon T$ $\ell_1$-calibration error over the $d$-dimensional simplex requires $T \geq \exp(\mathrm{poly}(1/\epsilon))$ (assuming $d \geq \mathrm{poly}(1/\epsilon)$). This strengthens the corresponding $d^{\Omega(\log{1/\epsilon})}$ lower bound of Peng 2025, and shows that an exponential dependence on $1/\epsilon$ is necessary.

Bandeira et al. (2022) introduced the Franz-Parisi (FP) criterion for characterizing the computational hard phases in statistical detection problems. The FP criterion, based on an annealed version of the celebrated Franz-Parisi potential from statistical physics, was shown to be equivalent to low-degree polynomial (LDP) lower bounds for Gaussian additive models, thereby connecting two distinct approaches to understanding the computational hardness in statistical inference. In this paper, we propose a refined FP criterion that aims to better capture the geometric ``overlap" structure of statistical models. Our main result establishes that this optimized FP criterion is equivalent to Statistical Query (SQ) lower bounds---another foundational framework in computational complexity of statistical inference. Crucially, this equivalence holds under a mild, verifiable assumption satisfied by a broad class of statistical models, including Gaussian additive models, planted sparse models, as well as non-Gaussian component analysis (NGCA), single-index (SI) models, and convex truncation detection settings. For instance, in the case of convex truncation tasks, the assumption is equivalent with the Gaussian correlation inequality (Royen, 2014) from convex geometry. In addition to the above, our equivalence not only unifies and simplifies the derivation of several known SQ lower bounds—such as for the NGCA model (Diakonikolas et al., 2017) and the SI model (Damian et al., 2024)—but also yields new SQ lower bounds of independent interest, including for the computational gaps in mixed sparse linear regression (Arpino et al., 2023) and convex truncation (De et al., 2023).

Large Multi-modality Models (LMMs) have made significant progress in visual understanding and generation, but they still face challenges in General Visual Editing, particularly in following complex instructions, preserving appearance consistency, and supporting flexible input formats. To study this gap, we introduce RISEBench, the first benchmark for evaluating Reasoning-Informed viSual Editing (RISE). RISEBench focuses on four key reasoning categories: Temporal, Causal, Spatial, and Logical Reasoning. We curate high-quality test cases for each category and propose an robust evaluation framework that assesses Instruction Reasoning, Appearance Consistency, and Visual Plausibility with both human judges and the LMM-as-a-judge approach. We conducted experiments evaluating nine prominent visual editing models, comprising both open-source and proprietary models. The evaluation results demonstrate that current models face significant challenges in reasoning-based editing tasks. Even the most powerful model evaluated, GPT-image-1, achieves an accuracy of merely 28.8%. RISEBench effectively highlights the limitations of contemporary editing models, provides valuable insights, and indicates potential future directions for the field of reasoning-aware visual editing. Our code and data have been released at https://github.com/PhoenixZ810/RISEBench.

Reinforcement learning (RL) systems have countless applications, from energy-grid management to protein design. However, such real-world scenarios are often extremely difficult, combinatorial in nature, and require complex coordination between multiple agents. This level of complexity can cause even state-of-the-art RL systems, trained until convergence, to hit a performance ceiling which they are unable to break out of with zero-shot inference. Meanwhile, many digital or simulation-based applications allow for an inference phase that utilises a specific time and compute budget to explore multiple attempts before outputting a final solution. In this work, we show that such an inference phase employed at execution time, and the choice of a corresponding inference strategy, are key to breaking the performance ceiling observed in complex multi-agent RL problems. Our main result is striking: we can obtain up to a 126% and, on average, a 45% improvement over the previous state-of-the-art across 17 tasks, using only a couple seconds of extra wall-clock time during execution. We also demonstrate promising compute scaling properties, supported by over 60k experiments, making it the largest study on inference strategies for complex RL to date. We make all of our experimental data and code available.

Linear Mode Connectivity (LMC) is a notable phenomenon in the loss landscapes of neural networks, wherein independently trained models have been observed to be connected—up to permutation symmetries—by linear paths in parameter space along which the loss remains consistently low. This observation challenges classical views of non-convex optimization and has implications for model ensembling, generalization, and our understanding of neural loss geometry. Inspired by recent studies on LMC in standard neural networks, we systematically investigate this phenomenon within Mixture-of-Experts (MoE) architectures—a class of models known for their scalability and computational efficiency, which combine traditional neural networks—referred to as experts—through a learnable gating mechanism. We begin by conducting a comprehensive analysis of both dense and sparse gating regimes, demonstrating that the symmetries inherent to MoE architectures are fully characterized by permutations acting on both the expert components and the gating function. Building on these foundational findings, we propose a matching algorithm that enables alignment between independently trained MoEs, thereby facilitating the discovery of LMC. Finally, we empirically validate the presence of LMC using our proposed algorithm across diverse MoE configurations—including dense, sparse, and shared-expert variants—under a wide range of model settings and datasets of varying scales and modalities. Our results confirm the existence of LMC in MoE architectures and offer fundamental insights into the functional landscape and optimization dynamics of deep learning models.

Sparse Autoencoders (SAEs) aim to decompose the activation space of large language models (LLMs) into human-interpretable latent directions or features. As we increase the number of features in the SAE, hierarchical features tend to split into finer features (“math” may split into “algebra”, “geometry”, etc.), a phenomenon referred to as feature splitting. However, we show that sparse decomposition and splitting of hierarchical features is not robust. Specifically, we show that seemingly monosemantic features fail to fire where they should, and instead get “absorbed” into their children features. We coin this phenomenon feature absorption, and show that it is caused by optimizing for sparsity in SAEs whenever the underlying features form a hierarchy. We introduce a metric to detect absorption in SAEs, and validate our findings empirically on hundreds of LLM SAEs. Our investigation suggests that varying SAE sizes or sparsity is insufficient to solve this issue. We discuss the implications of feature absorption in SAEs and some potential approaches to solve the fundamental theoretical issues before SAEs can be used for interpreting LLMs robustly and at scale.

Survey papers are foundational to the scholarly progress of research communities, offering structured overviews that guide both novices and experts across disciplines. However, the recent surge of AI-generated surveys, especially enabled by large language models (LLMs), has transformed this traditionally labor-intensive genre into a low-effort, high-volume output. While such automation lowers entry barriers, it also introduces a critical threat: the phenomenon we term the "survey paper DDoS attack" to the research community. This refers to the unchecked proliferation of superficially comprehensive but often redundant, low-quality, or even hallucinated survey manuscripts, which floods preprint platforms, overwhelms researchers, and erodes trust in the scientific record. In this position paper, we argue that we must stop uploading massive amounts of AI-generated survey papers (i.e., survey paper DDoS attack) to the research community, by instituting strong norms for AI-assisted review writing. We call for restoring expert oversight and transparency in AI usage and, moreover, developing new infrastructures such as Dynamic Live Surveys, community-maintained, version-controlled repositories that blend automated updates with human curation. Through quantitative trend analysis, quality audits, and cultural impact discussion, we show that safeguarding the integrity of surveys is no longer optional but imperative to the research community.

Previous research has explored the expressivity of Transformer models in simulating Boolean circuits or Turing machines. However, the efficient learnability of Transformers from data has remained an open question. Our study addresses this gap by providing the first polynomial-time learnability results (specifically strong, agnostic PAC learning) for single-layer Transformers with linear attention. We show that learning the optimal multi head linear attention can be recast as finding the optimal kernel predictor in a suitably defined RKHS. Moving to generalization, we construct an algorithm that, given a dataset, checks in polynomial time whether the set of best fit multi head linear attention networks on this data all perform an identical computation--a powerful notion for out of distribution generalization. We empirically validate our theoretical findings on several canonical tasks: learning random linear attention networks, key--value associations, and learning to execute finite automata. Our findings bridge a critical gap between theoretical expressivity and learnability of Transformer models.