NeurIPS 2025 Papers

Partial observability is a notorious challenge in reinforcement learning (RL), due to the need to learn complex, history-dependent policies. Recent empirical successes have used *privileged expert distillation* -- which leverages availability of latent state information during training (e.g., from a simulator) to learn and imitate the optimal latent, Markovian policy -- to disentangle the task of ''learning to see'' from ''learning to act''. While expert distillation is more computationally efficient than RL without latent state information, it also has well-documented failure modes. In this paper -- through a simple but instructive theoretical model called the *perturbed Block MDP*, and controlled experiments on challenging simulated locomotion tasks -- we investigate the algorithmic trade-off between privileged expert distillation and standard RL without privileged information. Our main findings are: **(1)** The trade-off empirically hinges on the *stochasticity* of the latent dynamics, as theoretically predicted by contrasting *approximate decodability* with *belief contraction* in the perturbed Block MDP; and **(2)** The optimal latent policy is not always the best latent policy to distill. Our results suggest new guidelines for effectively exploiting privileged information, potentially advancing the efficiency of policy learning across many practical partially observable domains.

We study the problem of contextual combinatorial semi-bandits, where input contexts are mapped into subsets of size $m$ of a collection of $K$ possible actions. In each round of the interaction, the learner observes feedback consisting of the realized reward of the predicted actions. Motivated by prototypical applications of contextual bandits, we focus on the $s$-sparse regime where we assume that the sum of rewards is bounded by some value $s \ll K$. For example, in recommendation systems the number of products purchased by any customer is significantly smaller than the total number of available products. Our main result is for the $(\varepsilon,\delta)$-PAC variant of the problem for which we design an algorithm that returns an $\varepsilon$-optimal policy with high probability using a sample complexity of $\widetilde{O}\big( (\mathrm{poly}(K/m) + sm / \varepsilon^2) \log (|\Pi|/\delta) \big)$ where $\Pi$ is the underlying (finite) class and $s$ is the sparsity parameter. This bound improves upon known bounds for combinatorial semi-bandits whenever $s \ll K$, and in the regime where $s = O(1)$, the leading terms in our bound match the corresponding full-information rates, implying that bandit feedback essentially comes at no cost. Our PAC learning algorithm is also computationally efficient given access to an ERM oracle for $\Pi$. Our framework generalizes the list multiclass classification problem with bandit feedback, which can be seen as a special case with binary reward vectors. In the special case of single-label classification corresponding to $s=m=1$, we prove an $O \big((K^7 + 1/\varepsilon^2)\log (|\mathcal{H}|/\delta)\big)$ sample complexity bound for a finite hypothesis class $\mathcal{H}$, which improves upon recent results in this scenario. Additionally, we consider the regret minimization setting where data can be generated adversarially, and establish a regret bound of $\widetilde O(|\Pi| + \sqrt{smT \log |\Pi|})$, extending the result of Erez et al. ('24) who consider the simpler single label classification setting.

Recent advances in 3D Gaussian diffusion models suffer from time-intensive denoising and post-denoising processing due to the massive number of Gaussian primitives, resulting in slow generation and limited scalability along sampling trajectories. To improve the efficiency of 3D diffusion models, we propose $\textbf{TRIM}$ ($\textbf{T}$rajectory $\textbf{R}$eduction and $\textbf{I}$nstance $\textbf{M}$ask denoising), a post-training approach that incorporates both temporal and spatial trimming strategies, to accelerate inference without compromising output quality while supporting the inference-time scaling for Gaussian diffusion models. Instead of scaling denoising trajectories in a costly end-to-end manner, we develop a lightweight selector model to evaluate latent Gaussian primitives derived from multiple sampled noises, enabling early trajectory reduction by selecting candidates with high-quality potential. Furthermore, we introduce instance mask denoising to prune learnable Gaussian primitives by filtering out redundant background regions, reducing inference computation at each denoising step. Extensive experiments and analysis demonstrate that TRIM significantly improves both the efficiency and quality of 3D generation.

Unsupervised Graph Domain Adaptation has become a promising paradigm for transferring knowledge from a fully labeled source graph to an unlabeled target graph. Existing graph domain adaptation models primarily focus on the closed-set setting, where the source and target domains share the same label spaces. However, this assumption might not be practical in the real-world scenarios, as the target domain might include classes that are not present in the source domain. In this paper, we investigate the problem of unsupervised open-set graph domain adaptation, where the goal is to not only correctly classify target nodes into the known classes, but also recognize previously unseen node types into the unknown class. Towards this end, we propose a novel framework called GraphRTA, which conducts reprogramming on both the graph and model sides. Specifically, we reprogram the graph by modifying target graph structure and node features, which facilitates better separation of known and unknown classes. Meanwhile, we also perform model reprogramming by pruning domain-specific parameters to reduce bias towards the source graph while preserving parameters that capture transferable patterns across graphs. Additionally, we extend the classifier with an extra dimension for the unknown class, thus eliminating the need of manually specified threshold in open-set recognition. Comprehensive experiments on several public datasets demonstrate that our proposed model can achieve satisfied performance compared with recent state-of-the-art baselines. Our source codes and datasets are publicly available at https://github.com/cszhangzhen/GraphRTA.

Differentiable economics—the use of deep learning for auction design—has driven progress in multi-item auction design with additive and unit-demand valuations. However, there has been little progress for combinatorial auctions (CAs), even in the simplest and yet important single bidder case, due to exponential growth of the bundle space with the number of items. We address this challenge by introducing a deep network architecture for a menu-based CA, which supports the first dominant-strategy incentive compatible (DSIC), revenue-optimizing single-bidder CA. Our idea is to generate a bundle distribution through an ordinary differential equation (ODE) applied to a tractable initial distribution. The BundleFlow method learns suitable ODE-based transforms, one for each menu element, to optimize expected revenue. BundleFlow achieves up to 2.23$\times$ higher revenue than baselines on standard CA testbeds and scales up to 500 items. Compared with other menu-learning baselines, BundleFlow also reduces training iterations by 3.6-9.5$\times$ and cuts training time by about 80% in settings with 50 and 100 items.

Recent approaches for vision-language models (VLMs) have shown remarkable success in achieving fast downstream adaptation. When applied to real-world downstream tasks, VLMs inevitably encounter both the in-distribution (ID) data and out-of-distribution (OOD) data. The OOD datasets often include both covariate shifts (e.g., known classes with changes in image styles) and semantic shifts (e.g., test-time unseen classes). This highlights the importance of improving VLMs' generalization ability to covariate-shifted OOD data, while effectively detecting open-set semantic-shifted OOD classes. In this paper, inspired by the substantial energy change observed in closed-set data when re-aligning vision-language modalities—specifically by directly reducing the maximum cosine similarity to a low value—we introduce a novel OOD score, named $\Delta\mathrm{Energy}$. $\Delta\mathrm{Energy}$ significantly outperforms the vanilla energy-based OOD score and provides a more reliable approach for OOD detection. Furthermore, $\Delta\mathrm{Energy}$ can simultaneously improve OOD generalization under covariate shifts, which is achieved by lower-bound maximization for $\Delta\mathrm{Energy}$ (termed EBM). EBM is theoretically proven to not only enhance OOD detection but also yields a domain-consistent Hessian, which serves as a strong indicator for OOD generalization. Based on this finding, we developed a unified fine-tuning framework that allows for improving VLMs' robustness in both OOD generalization and OOD detection. Extensive experiments on challenging OOD detection and generalization benchmarks demonstrate the superiority of our method, outperforming recent approaches by 10\%–25\% in AUROC.

Large vision-language models (LVLMs) have emerged as foundational tools for real-world AI applications. Despite their remarkable capabilities, current LVLMs process entire images at the token level, leading to significant inefficiencies compared to human cognition, which selectively focuses on high-level vision concepts. This token-level redundancy becomes increasingly problematic for high-resolution images and long video sequences, resulting in large computational costs and limited scalability in practical applications. To address this limitation, we introduce the concept of a vision concept model, a novel paradigm that enables LVLMs to dynamically extract the most relevant vision concepts from complex inputs, based on task-specific instructions. To optimize this vision concept modeling process, we propose VCM, a self-supervised framework that leverages vision-language correlations across diverse instances. VCM is designed to learn meaningful vision concepts without the need for expensive concept-level annotations. At its core, it employs a forward-backward optimization algorithm that supports LVLMs to adjust concept granularity and spatial alignment dynamically. Experiments demonstrate that VCM remarkably reduces computational costs (e.g., achieving up to 85\% fewer FLOPs for LLaVA-1.5-7B), while maintaining strong performance across a series of vision-language tasks. The codebase is available at https://github.com/RainBowLuoCS/VCM.

Large multimodal models (LMMs) have gained impressive performance due to their outstanding capability in various understanding tasks. However, these models still suffer from some fundamental limitations related to robustness and generalization due to the alignment and correlation between visual and textual features. In this paper, we introduce a simple but efficient learning mechanism for improving the robust alignment between visual and textual modalities by solving shuffling problems. In particular, the proposed approach can improve reasoning capability, visual understanding, and cross-modality alignment by introducing two new tasks: reconstructing the image order and the text order into the LMM's pre-training and fine-tuning phases. In addition, we propose a new directed-token approach to capture visual and textual knowledge, enabling the capability to reconstruct the correct order of visual inputs. Then, we introduce a new Image-to-Response Guided loss to further improve the visual understanding of the LMM in its responses. The proposed approach consistently achieves state-of-the-art (SoTA) performance compared with prior LMMs on academic task-oriented and instruction-following LMM benchmarks.

Vision-language models (VLMs) as foundation models have significantly enhanced performance across a wide range of visual and textual tasks, without requiring large-scale training from scratch for downstream tasks. However, these deterministic VLMs fail to capture the inherent ambiguity and uncertainty in natural language and visual data. Recent probabilistic post-hoc adaptation methods address this by mapping deterministic embeddings onto probability distributions; however, existing approaches do not account for the asymmetric uncertainty between modalities, and the constraint that meaningful deterministic embeddings reside on a unit hypersphere, potentially leading to suboptimal performance. In this paper, we address the asymmetric uncertainty structure inherent in textual and visual data, and propose AsymVLM to build probabilistic embeddings from pre-trained VLMs on the unit hypersphere, enabling uncertainty quantification. We validate the effectiveness of the probabilistic embeddings on established benchmarks, and present comprehensive ablation studies demonstrating the inherent nature of asymmetry in the uncertainty structure of textual and visual data.

Generating realistic time series samples is crucial for stress-testing models and protecting user privacy by using synthetic data. In engineering and safety-critical applications, these samples must meet certain hard constraints that are domain-specific or naturally imposed by physics or nature. Consider, for example, generating electricity demand patterns with constraints on peak demand times. This can be used to stress-test the functioning of power grids during adverse weather conditions. Existing approaches for generating constrained time series are either not scalable or degrade sample quality. To address these challenges, we introduce Constrained Posterior Sampling (CPS), a diffusion-based sampling algorithm that aims to project the posterior mean estimate into the constraint set after each denoising update. Notably, CPS scales to a large number of constraints ($\sim100$) without requiring additional training. We provide theoretical justifications highlighting the impact of our projection step on sampling. Empirically, CPS outperforms state-of-the-art methods in sample quality and similarity to real time series by around 70\% and 22\%, respectively, on real-world stocks, traffic, and air quality datasets.

Large Language Models (LLMs) have demonstrated outstanding performance in mathematical reasoning capabilities. However, we argue that current large-scale reasoning models primarily rely on scaling up training datasets with diverse mathematical problems and long thinking chains, which raises questions about whether LLMs genuinely acquire mathematical concepts and reasoning principles or merely remember the training data. In contrast, humans tend to break down complex problems into multiple fundamental atomic capabilities. Inspired by this, we propose a new paradigm for evaluating mathematical atomic capabilities. Our work categorizes atomic abilities into two dimensions: (1) field-specific abilities across four major mathematical fields, algebra, geometry, analysis, and topology, and (2) logical abilities at different levels, including conceptual understanding, forward multi-step reasoning with formal math language, and counterexample-driven backward reasoning. We propose corresponding training and evaluation datasets for each atomic capability unit, and conduct extensive experiments about how different atomic capabilities influence others, to explore the strategies to elicit the required specific atomic capability. Evaluation and experimental results on advanced models show many interesting discoveries and inspirations about the different performances of models on various atomic capabilities and the interactions between atomic capabilities. Our findings highlight the importance of decoupling mathematical intelligence into atomic components, providing new insights into model cognition and guiding the development of training strategies toward a more efficient, transferable, and cognitively grounded paradigm of "atomic thinking".

Algorithmic fairness has emerged as a central issue in ML, and it has become standard practice to adjust ML algorithms so that they will satisfy fairness requirements such as Equal Opportunity. In this paper we consider the effects of adopting such fair classifiers on the overall level of _ecosystem fairness_. Specifically, we introduce the study of fairness with competing firms, and demonstrate the failure of fair classifiers in yielding fair ecosystems. Our results quantify the loss of fairness in systems, under a variety of conditions, based on classifiers' correlation and the level of their data overlap. We show that even if competing classifiers are individually fair, the ecosystem's outcome may be unfair; and that adjusting biased algorithms to improve their individual fairness may lead to an overall decline in ecosystem fairness. In addition to these theoretical results, we also provide supporting experimental evidence. Together, our model and results provide a novel and essential call for action.

This paper investigates the dynamical properties of tokens in pre-trained transformer models and explores their application to improving Transformers. To this end, we analyze the dynamical system governing the continuous-time limit of the pre-trained model and characterize the asymptotic behavior of its solutions. Specifically, we characterize when tokens move closer to or farther from one another over time, depending on the model parameters. We provide sufficient conditions, based on these parameters, to identify scenarios where tokens either converge to zero or diverge to infinity. Unlike prior works, our conditions are broader in scope and more applicable to real-world models. Furthermore, we investigate how different forms of positional encoding - specifically absolute and rotary - affect these dynamical regimes. Empirical evidence reveals that the convergence scenario adversely impacts model performance. Motivated by these insights, we propose simple refinements to Transformer architectures that mitigate convergence behavior in models with absolute or rotary positional encoding. These findings support theoretical foundations and design principles for improving Transformer models.

Large Language Models (LLMs) are expected to possess comprehensive medical knowledge to support real-world clinical applications. While domain-specific fine-tuning effectively injects medical knowledge into LLMs, it often causes catastrophic forgetting of previously acquired knowledge and instruction-following capabilities. In this paper, we investigate this issue and reveal a pattern of proximity-dependent forgetting: knowledge that is semantically or topically close to the injected content is more likely to be forgotten, while unrelated knowledge shows minimal degradation. Moreover, we observe that existing mitigation techniques fail to address this type of forgetting effectively. Motivated by this observation and inspired by human learning mechanisms, we proposeInternAL (\Internal Knowledge Augmentation Learning), a novel approach that leverages LLMs' own internal knowledge to mitigate forgetting. InternAL first probes internal knowledge closely related to the injection by prompting the model with questions derived from the injected knowledge. This knowledge is then used to augment the original injection dataset, guiding the model to retain related prior knowledge during training. Experimental results on multiple LLMs (LLaMA, Qwen) demonstrate that InternAL significantly mitigates proximity-related forgetting while maintaining strong knowledge injection performance. Our findings provide new insights into the nature of catastrophic forgetting in medical knowledge injection and highlight a promising direction for robust domain adaptation in LLMs. Code and datasets are available at https://github.com/THUMLP/InternAL.

Urban warming differs markedly from regional background trends, highlighting the unique behavior of urban climates and the challenges they present. Accurately predicting local urban climate necessitates modeling the interactions between urban surfaces and atmospheric forcing. Although off-the-shelf machine learning (ML) algorithms offer considerable accuracy for climate prediction, they often function as black boxes, learning data mappings rather than capturing physical evolution. As a result, they struggle to capture key land-atmosphere interactions and may produce physically inconsistent predictions. To address these limitations, we propose UCformer, a novel multi-task, physics-guided Transformer architecture designed to emulate nonlinear urban climate processes. UCformer jointly estimates 2-m air temperature $\(T\)$, specific humidity $\(q\)$, and dew point temperature $\(t\)$ in urban areas, while embedding domain and physical priors into its learning structure. Experimental results demonstrate that incorporating domain and physical knowledge leads to significant improvements in emulation accuracy and generalizability under future urban climate scenarios. Further analysis reveals that learning shared correlations across cities enables the model to capture transferable urban surface–atmosphere interaction patterns, resulting in improved accuracy in urban climate emulation. Finally, UCformer shows strong potential to fit real-world data: when fine-tuned with limited observational data, it achieves competitive performance in estimating urban heat fluxes compared to a physics-based model.

On-device large models (LMs) reduce cloud dependency but expose proprietary model weights to the end-user, making them vulnerable to white-box model stealing (MS) attacks. A common defense is TEE-Shielded DNN Partition (TSDP), which places all trainable LoRA adapters (fine tuned on private data) inside a trusted execution environment (TEE). However, this design suffers from excessive host-to-TEE communication latency. We propose AegisGuard, a fine tuning and deployment framework that selectively shields the MS sensitive adapters while offloading the rest to the GPU, balancing security and efficiency. AegisGuard integrates two key components: i) RL-based Sensitivity Measurement (RSM), which injects Gaussian noise during training and applies a lightweight reinforcement learning to rank adapters based on their impact on model stealing; and (ii) Shielded-Adapter Compression (SAC), which structurally prunes the selected adapters to reduce both parameter size and intermediate feature maps, further lowering TEE computation and data transfer costs. Extensive experiments demonstrate that AegisGuard achieves black-box level MS resilience (surrogate accuracy around 39%, matching fully shielded baselines), while reducing end-to-end inference latency by 2–3× and cutting TEE memory usage by 4× compared to state-of-the-art TSDP methods.

The text produced by language models (LMs) can exhibit specific `behaviors,' such as a failure to follow alignment training, that we hope to detect and react to during deployment. Identifying these behaviors can often only be done post facto, i.e., after the entire text of the output has been generated. We provide evidence that there are times when we can predict how an LM will behave early in computation, before even a single token is generated. We show that probes trained on the internal representation of input tokens alone can predict a wide range of eventual behaviors over the entire output sequence. Using methods from conformal prediction, we provide provable bounds on the estimation error of our probes, creating precise early warning systems for these behaviors. The conformal probes can identify instances that will trigger alignment failures (jailbreaking) and instruction-following failures, without requiring a single token to be generated. An early warning system built on the probes reduces jailbreaking by 91%. Our probes also show promise in pre-emptively estimating how confident the model will be in its response, a behavior that cannot be detected using the output text alone. Conformal probes can preemptively estimate the final prediction of an LM that uses Chain-of-Thought (CoT) prompting, hence accelerating inference. When applied to an LM that uses CoT to perform text classification, the probes drastically reduce inference costs (65% on average across 27 datasets), with negligible accuracy loss. Encouragingly, probes generalize to unseen datasets and perform better on larger models, suggesting applicability to the largest of models in real-world settings.

Unsigned Distance Fields (UDFs) are a natural implicit representation for open surfaces but, unlike Signed Distance Fields (SDFs), are challenging to triangulate into explicit meshes. This is especially true at high resolutions where neural UDFs exhibit higher noise levels, which makes it hard to capture fine details. Most current techniques perform within single voxels without reference to their neighborhood, resulting in missing surface and holes where the UDF is ambiguous or noisy. We show that this can be remedied by performing several passes and by reasoning on previously extracted surface elements to incorporate neighborhood information. Our key contribution is an iterative neural network that does this and progressively improves surface recovery within each voxel by spatially propagating information from increasingly distant neighbors. Unlike single-pass methods, our approach integrates newly detected surfaces, distance values, and gradients across multiple iterations, effectively correcting errors and stabilizing extraction in challenging regions. Experiments on diverse 3D models demonstrate that our method produces significantly more accurate and complete meshes than existing approaches, particularly for complex geometries, enabling UDF surface extraction at higher resolutions where traditional methods fail.

Neural networks offer a promising data-driven approach to tackle computationally challenging optimization problems. In this work, we introduce neural approximation frameworks for a family of geometric "extent measure" problems, including shape-fitting descriptors (e.g. minimum enclosing ball or annulus). Central to our approach is the \textit{alignment} of our neural model with a new variant of the classical $\varepsilon$-kernel technique from computational geometry. In particular, we develop a new relaxed-$\varepsilon$-kernel theory that maintains the approximation guarantees of the classical $\varepsilon$-kernels but with the crucial benefit that it can be easily implemented with \textit{bounded model complexity} (i.e, constant number of parameters) by the simple SumFormer neural network. This leads to a simple neural model to approximate objects such as the directional width of any input point set, and empirically shows excellent out-of-distribution generalization. Many geometric extent measures, such as the minimum enclosing spherical shell, cannot be directly captured by $\varepsilon$-kernels. To this end, we show that an encode-process-decode framework with our kernel approximating NN used as the ``process'' module can approximate such extent measures, again, with bounded model complexity where parameters scale only with the approximation error $\varepsilon$ and not the size of the input set. Empirical results on diverse point‐cloud datasets demonstrate the practical performance of our models.