NeurIPS 2025 论文列表

Comprehending natural language and following human instructions are critical capabilities for intelligent agents. However, the flexibility of linguistic instructions induces substantial ambiguity across language-conditioned tasks, severely degrading algorithmic performance. To address these limitations, we present a novel method named DAIL (Distributional Aligned Learning), featuring two key components: distributional policy and semantic alignment. Specifically, we provide theoretical results that the value distribution estimation mechanism enhances task differentiability. Meanwhile, the semantic alignment module captures the correspondence between trajectories and linguistic instructions. Extensive experimental results on both structured and visual observation benchmarks demonstrate that DAIL effectively resolves instruction ambiguities, achieving superior performance to baseline methods. Our implementation is available at https://github.com/RunpengXie/Distributional-Aligned-Learning.

Large Language Models (LLMs) have yet to effectively leverage the vast amounts of edge-device data, and Federated Learning (FL) offers a promising paradigm to collaboratively fine-tune LLMs without transferring private edge data to the cloud. To operate within the computational and communication constraints of edge devices, recent literature on federated fine-tuning of LLMs proposes the use of low-rank adaptation (LoRA) and similar parameter-efficient methods. However, LoRA-based methods suffer from accuracy degradation in FL settings, primarily because of data and computational heterogeneity across clients. We propose Ravan, an adaptive multi-head LoRA method that balances parameter efficiency and model expressivity by reparameterizing the weight updates as the sum of multiple LoRA heads, $s_i\textbf{B}_i\textbf{H}_i\textbf{A}_i$, in which only the $\textbf{H}_i$ parameters and their lightweight scaling factors $s_i$ are trained. These trainable scaling factors let the optimization focus on the most useful heads, recovering a higher-rank approximation of the full update without increasing the number of communicated parameters since clients upload $s_i\textbf{H}_i$ directly. Experiments on vision and language benchmarks show that Ravan improves test accuracy by 2–8\% over prior parameter-efficient baselines, making it a robust and scalable solution for federated fine-tuning of LLMs.

Federated learning enables collaborative training while preserving the privacy of all participants. However, the heterogeneity in data distribution across multiple training nodes poses significant challenges to the construction of federated models. Prior studies were dedicated to mitigating the effects of data heterogeneity by using global information as a blueprint and restricting the local update of the model for reaching a "hard consensus". But this practice makes it difficult to balance local and global information, and it neglects to negotiate amicably between local and global models to reach mutually agreeable results, called ``soft consensus". In this paper, a multiple-path solving method is proposed to balance global and local features and combine these two feature preference paths to reach a soft consensus. Rather than relying on global information as the sole criterion, a negotiation process is employed to address the same objective by accommodating diverse feature preferences, thereby facilitating the discovery of a more plausible solution through multiple distinct pathways. Considering the overwhelming power of local features during local training, a swapping strategy is applied to weaken them to balance the solution paths. Moreover, to minimize the additional communication cost caused by the introduction of multiple paths, the solution of the task network is converted into data adaptation to reduce the amount of parameter transmission. Extensive experiments are conducted to demonstrate the advantages of the proposed method.

We present Orientation-anchored Gaussian Splatting (OriGS), a novel framework for high-quality 4D reconstruction from casually captured monocular videos. While recent advances extend 3D Gaussian Splatting to dynamic scenes via various motion anchors, such as graph nodes or spline control points, they often rely on low-rank assumptions and fall short in modeling complex, region-specific deformations inherent to unconstrained dynamics. OriGS addresses this by introducing a hyperdimensional representation grounded in scene orientation. We first estimate a Global Orientation Field that propagates principal forward directions across space and time, serving as stable structural guidance for dynamic modeling. Built upon this, we propose Orientation-aware Hyper-Gaussian, a unified formulation that embeds time, space, geometry, and orientation into a coherent probabilistic state. This enables inferring region-specific deformation through principled conditioned slicing, adaptively capturing diverse local dynamics in alignment with global motion intent. Experiments demonstrate the superior reconstruction fidelity of OriGS over mainstream methods in challenging real-world dynamic scenes.

This paper addresses the Bayesian optimization problem (also referred to as the Bayesian setting of the Gaussian process bandit), where the learner seeks to minimize the regret under a function drawn from a known Gaussian process (GP). Under a Mat\'ern kernel with some extent of smoothness, we show that the Gaussian process upper confidence bound (GP-UCB) algorithm achieves $\tilde{O}(\sqrt{T})$ cumulative regret with high probability. Furthermore, our analysis yields $O(\sqrt{T \ln^2 T})$ regret under a squared exponential kernel. These results fill the gap between the existing regret upper bound of GP-UCB and the current best upper bound provided by Scarlett [2018]. The key idea in our proof is to capture the concentration behavior of the input sequence realized by GP-UCB, enabling us to handle GP's information gain in a refined manner.

In various scientific and engineering domains, where data acquisition is costly—such as in medical imaging, environmental monitoring, or remote sensing—strategic sampling from unobserved regions, guided by prior observations, is essential to maximize target discovery within a limited sampling budget. In this work, we introduce Diffusion-guided Active Target Discovery (DiffATD), a novel method that leverages diffusion dynamics for active target discovery. DiffATD maintains a belief distribution over each unobserved state in the environment, using this distribution to dynamically balance exploration-exploitation. Exploration reduces uncertainty by sampling regions with the highest expected entropy, while exploitation targets areas with the highest likelihood of discovering the target, indicated by the belief distribution and an incrementally trained reward model designed to learn the characteristics of the target. DiffATD enables efficient target discovery in a partially observable environment within a fixed sampling budget, all without relying on any prior supervised training. Furthermore, DiffATD offers interpretability, unlike existing black-box policies that require extensive supervised training. Through extensive experiments and ablation studies across diverse domains, including medical imaging, species discovery and remote sensing, we show that DiffATD performs significantly better than baselines and competitively with supervised methods that operate under full environmental observability.

Open-vocabulary part segmentation (OVPS) struggles with structurally connected boundaries due to the inherent conflict between continuous image features and discrete classification mechanism. To address this, we propose PBAPS, a novel training-free framework specifically designed for OVPS. PBAPS leverages structural knowledge of object-part relationships to guide a progressive segmentation from objects to fine-grained parts. To further improve accuracy at challenging boundaries, we introduce a Boundary-Aware Refinement (BAR) module that identifies ambiguous boundary regions by quantifying classification uncertainty, enhances the discriminative features of these ambiguous regions using high-confidence context, and adaptively refines part prototypes to better align with the specific image. Experiments on Pascal-Part-116, ADE20K-Part-234, PartImageNet demonstrate that PBAPS significantly outperforms state-of-the-art methods, achieving 46.35\% mIoU and 34.46\% bIoU on Pascal-Part-116. Our code is available at https://github.com/TJU-IDVLab/PBAPS.

The scaling of model and data sizes has reshaped the AI landscape, establishing finetuning pretrained models as the standard paradigm for solving downstream tasks. However, dominant finetuning methods typically rely on weight adaptation, often lack interpretability, and depend on heuristically chosen hyperparameters. In this paper, we take a different perspective and shift the focus from weights to activation functions, viewing them through the lens of spline operators. We propose Curvature Tuning (CT), an interpretable and principled steering method that modulates a model's decision boundary by injecting a single hyperparameter into its activation functions. We show that CT provably adjusts model decision boundary curvature and, more fundamentally, projects a model onto a space of smooth functions—thereby complementing current finetuning methods, whose effect lies primarily in feature adaptation. Making this hyperparameter trainable gives rise to a novel and highly parameter-efficient finetuning method. Empirically, CT improves both generalization and robustness. For example, it boosts downstream accuracy of ResNet-50/152 by 8.59\%/8.34\% over linear probing and 4.64\%/1.70\% over LoRA across 12 datasets, and improves robust accuracy on the $\ell_{\infty}$ benchmark from RobustBench by 1032.64\%/1494.46\%. Our code is available at https://github.com/Leon-Leyang/curvature-tuning.

Sparse-reward reinforcement learning (RL) can model a wide range of highly complex tasks. Solving sparse-reward tasks is RL's core premise — requiring efficient exploration coupled with long-horizon credit assignment — and overcoming these challenges is key for building self-improving agents with superhuman ability. We argue that solving complex and high-dimensional tasks requires solving simpler tasks that are *relevant* to the target task. In contrast, most prior work designs strategies for selecting exploratory tasks with the objective of solving *any* task, making exploration of challenging high-dimensional, long-horizon tasks intractable. We find that the sense of direction, necessary for effective exploration, can be extracted from existing reinforcement learning algorithms, without needing any prior information. Based on this finding, we propose a method for _**di**rected **s**parse-reward goal-**co**nditioned **ve**ry long-horizon **R**L_ (DISCOVER), which selects exploratory goals in the direction of the target task. We connect DISCOVER to principled exploration in bandits, formally bounding the time until the target task becomes achievable in terms of the agent's initial distance to the target, but independent of the volume of the space of all tasks. Empirically, we perform a thorough evaluation in high-dimensional simulated environments. We find that the directed goal selection of DISCOVER solves exploration problems that are beyond the reach of prior state-of-the-art exploration methods in RL.

Recent booming time series models have demonstrated remarkable forecasting performance. However, these methods often place greater focus on more effectively modelling the historical series, largely neglecting the forecasting phase, which generates long-term forecasts by separately predicting multiple time points. Given that real-world time series typically consist of various long short-term dynamics, independent predictions over individual time points may fail to express complex underlying patterns and can lead to a lack of global views. To address these issues, this work explores new perspectives from the forecasting phase and proposes a novel Implicit Forecaster (IF) as an additional decoding module. Inspired by decomposition forecasting, IF adopts a more nuanced approach by implicitly predicting constituent waves represented by their frequency, amplitude, and phase, thereby accurately forming the time series. Extensive experimental results from multiple real-world datasets show that IF can consistently boost mainstream time series models, achieving state-of-the-art forecasting performance. Code is available at this repository: [https://github.com/rakuyorain/Implicit-Forecaster](https://github.com/rakuyorain/Implicit-Forecaster).

Pan-sharpening aims to generate a spatially and spectrally enriched multi-spectral image by integrating complementary cross-modality information from low-resolution multi-spectral image and texture-rich panchromatic counterpart. In this work, we propose a WKV-sharing embraced random shuffle RWKV high-order modeling paradigm for pan-sharpening from Bayesian perspective, coupled with random weight manifold distribution training strategy derived from Functional theory to regularize the solution space adhering to the following principles: 1) Random-shuffle RWKV. Recently, the Vision RWKV model, with its inherent linear complexity in global modeling, has inspired us to explore its untapped potential in pan-sharpening tasks. However, its attention mechanism, relying on a recurrent bidirectional scanning strategy, suffers from biased effects and demands significant processing time. To address this, we propose a novel Bayesian-inspired scanning strategy called Random Shuffle, complemented by a theoretically-sound inverse shuffle to preserve information coordination invariance, effectively eliminating biases associated with fixed sequence scanning. The Random Shuffle approach mitigates preconceptions in global 2D dependencies in mathematical expectation, providing the model with an unbiased prior. In line with similar spirit of Dropout, we introduce a testing methodology based on Monte Carlo averaging to ensure the model’s output aligns more closely with expected results. 2) WKV-sharing high-order. Regarding KV’s attention score calculation in spatial mixer of RWKV, we leverage WKV-sharing mechanism to transfer KV activations across RWKV layers, achieving lower latency and improved trainability, and revisit the channel mixer in RWKV, originally a first-order weighting function, and redevelop its high-order potential by sharing the gate mechanism across RWKV layer. Comprehensive experiments across pan-sharpening benchmarks demonstrate our model’s effectiveness, consistently outperforming state-of-the-art alternatives

Large language models (LLMs) have shown impressive performance on reasoning benchmarks like math and logic. While many works have largely assumed well-defined tasks, real-world queries are often underspecified and only solvable by acquiring missing information. We formalize this information-gathering problem as a constraint satisfaction problem (CSP) with missing variable assignments. Using a special case where only one necessary variable assignment is missing, we can evaluate an LLM's ability to identify the minimal necessary question to ask. We present QuestBench, a set of underspecified reasoning tasks solvable by asking at most one question, which includes: (1) Logic-Q: logical reasoning tasks with one missing proposition, (2) Planning-Q: PDDL planning problems with partially-observed initial states, (3) GSM-Q: human-annotated grade school math problems with one unknown variable, and (4) GSME-Q: equation-based version of GSM-Q. The LLM must select the correct clarification question from multiple options. While current models excel at GSM-Q and GSME-Q, they achieve only 40-50% accuracy on Logic-Q and Planning-Q. Analysis shows that the ability to solve well-specified reasoning problems is not sufficient for success on our benchmark: models struggle to identify the right question even when they can solve the fully specified version. This highlights the need for specifically optimizing models' information acquisition capabilities.

The vulnerability of Vision Large Language Models (VLLMs) to jailbreak attacks appears as no surprise. However, recent defense mechanisms against these attacks have reached near-saturation performance on benchmark evaluations, often with minimal effort. This dual high performance in both attack and defense gives rise to a fundamental and perplexing paradox. To gain a deep understanding of this issue and thus further help strengthen the trustworthiness of VLLMs, this paper makes three key contributions: i) One tentative explanation for VLLMs being prone to jailbreak attacks--inclusion of vision inputs, as well as its in-depth analysis. ii) The recognition of a largely ignored problem in existing VLLM defense mechanisms--over-prudence. The problem causes these defense methods to exhibit unintended abstention, even in the presence of benign inputs, thereby undermining their reliability in faithfully defending against attacks. iii) A simple safety-aware method--LLM-Pipeline. Our method repurposes the more advanced guardrails of LLMs on the fly, serving as an effective alternative detector prior to VLLM response. Last but not least, we find that the two representative evaluation methods for jailbreak often exhibit chance agreement. This limitation makes it potentially misleading when evaluating attack strategies or defense mechanisms. We believe the findings from this paper offer useful insights to rethink the foundational development of VLLM safety with respect to benchmark datasets, defense strategies, and evaluation methods.

Given a task and a set of steps composing it, Video Step Grounding (VSG) aims to detect which steps are performed in a video. Standard approaches for this task require a labeled training set (e.g., with step-level annotations or narrations), which may be costly to collect. Moreover, they process the full video offline, limiting their applications for scenarios requiring online decisions. Thus, in this work, we explore how to perform VSG online and without training. We achieve this by exploiting the zero-shot capabilities of recent Large Multimodal Models (LMMs). In particular, we use LMMs to predict the step associated with a restricted set of frames, without access to the whole video. We show that this online strategy without task-specific tuning outperforms offline and training-based models. Motivated by this finding, we develop Bayesian Grounding with Large Multimodal Models (BaGLM), further injecting knowledge of past frames into the LMM-based predictions. BaGLM exploits Bayesian filtering principles, modeling step transitions via (i) a dependency matrix extracted through large language models and (ii) an estimation of step progress. Experiments on three datasets show superior performance of BaGLM over state-of-the-art training-based offline methods.

Recent advances in diffusion models have revolutionized text-guided image editing, yet existing editing methods face critical challenges in hyperparameter identification. To get the reasonable editing performance, these methods often require the user to brute-force tune multiple interdependent hyperparameters, such as inversion timesteps and attention modification, \textit{etc.} This process incurs high computational costs due to the huge hyperparameter search space. We consider searching optimal editing's hyperparameters as a sequential decision-making task within the diffusion denoising process. Specifically, we propose a reinforcement learning framework, which establishes a Markov Decision Process that dynamically adjusts hyperparameters across denoising steps, integrating editing objectives into a reward function. The method achieves time efficiency through proximal policy optimization while maintaining optimal hyperparameter configurations. Experiments demonstrate significant reduction in search time and computational overhead compared to existing brute-force approaches, advancing the practical deployment of a diffusion-based image editing framework in the real world.

Machine learning models have achieved widespread success but often inherit and amplify historical biases, resulting in unfair outcomes. Traditional fairness methods typically impose constraints at the prediction level, without addressing underlying biases in data representations. In this work, we propose a principled framework that adjusts data representations to balance predictive utility and fairness. Using sufficient dimension reduction, we decompose the feature space into target-relevant, sensitive, and shared components, and control the fairness–utility trade-off by selectively removing sensitive information. We provide a theoretical analysis of how prediction error and fairness gaps evolve as shared subspaces are added, and employ influence functions to quantify their effects on the asymptotic behavior of parameter estimates. Experiments on both synthetic and real-world datasets validate our theoretical insights and show that the proposed method effectively improves fairness while preserving predictive performance.

Vision-language models heavily rely on visual representations, yet ensuring its efficiency remains a critical challenge. Most existing approaches focus on reducing visual tokens either at the visual encoder phase or during the LLM decoder stage. Inspired by human visual cognition, where an initial global glance precedes focused attention on semantically salient regions, we introduce Glance2Gaze, a cognitively inspired framework that mimics the human two-stage attention process. The framework consists of two key components: the Glance Fusion module, which integrates multi-layer vision transformer features with text-aware attention to generate a semantically enriched global representation, and the Gaze Compression module, which utilizes a novel query-guided mechanism to selectively compress visual tokens based on their semantic relevance. Experimental results on widely adopted benchmarks demonstrate that Glance2Gaze outperforms existing methods, achieving superior performance with equal or lower computational cost. Furthermore, it generalizes well to high-resolution and video scenarios, showcasing robust and scalable efficiency improvements in VLMs.

Although transformer-based models have shown exceptional empirical performance, the fundamental principles governing their training dynamics are inadequately characterized beyond configuration-specific studies. Inspired by empirical evidence showing improved reasoning capabilities under small initialization scales in language models, we employ the gradient flow analytical framework established in \cite{zhou2022towards} to systematically investigate linearized Transformer training dynamics. Our theoretical analysis dissects the dynamics of attention modules into two distinct stages. In the first stage, asymmetric weight perturbations from random initialization sustain non-degenerate gradient dynamics in parameter matrices, facilitating systematic escape from small initialization regimes. Subsequently, these matrices undergo condensation, progressively aligning toward the target orientation. In the second stage, the previously static key-query matrices actively participate in training, driving the normalized matrices toward asymptotic rank collapse. This two-stage framework generalizes classical directional convergence results.

Recent advances in multimodal models have raised questions about whether vision-and-language models (VLMs) integrate cross-modal information in ways that reflect human cognition. One well-studied test case in this domain is the bouba-kiki effect, where humans reliably associate pseudowords like ‘bouba’ with round shapes and ‘kiki’ with jagged ones. Given the mixed evidence found in prior studies for this effect in VLMs, we present a comprehensive re-evaluation focused on two variants of CLIP, ResNet and Vision Transformer (ViT), given their centrality in many state-of-the-art VLMs. We apply two complementary methods closely modelled after human experiments: a prompt-based evaluation that uses probabilities as a measure of model preference, and we use Grad-CAM as a novel approach to interpret visual attention in shape-word matching tasks. Our findings show that these model variants do not consistently exhibit the bouba-kiki effect. While ResNet shows a preference for round shapes, overall performance across both model variants lacks the expected associations. Moreover, direct comparison with prior human data on the same task shows that the models’ responses fall markedly short of the robust, modality-integrated behaviour characteristic of human cognition. These results contribute to the ongoing debate about the extent to which VLMs truly understand cross-modal concepts, highlighting limitations in their internal representations and alignment with human intuitions.