NeurIPS 2025 论文列表

We consider the problem of minimizing a $d$-dimensional Lipschitz convex function using a stochastic gradient oracle. We introduce and motivate a setting where the noise of the stochastic gradient is isotropic in that it is bounded in every direction with high probability. We then develop an algorithm for this setting which improves upon prior results by a factor of $d$ in certain regimes, and as a corollary, achieves a new state-of-the-art complexity for sub-exponential noise. We give matching lower bounds (up to polylogarithmic factors) for both results. Additionally, we develop an efficient quantum isotropifier, a quantum algorithm which converts a variance-bounded quantum sampling oracle into one that outputs an unbiased estimate with isotropic error. Combining our results, we obtain improved dimension-dependent rates for quantum stochastic convex optimization.

Tensor decomposition is a canonical non-convex optimization problem that is computationally challenging, and yet important due to applications in factor analysis and parameter estimation of latent variable models. In practice, scalable iterative methods, particularly Alternating Least Squares (ALS), remain the workhorse for tensor decomposition despite the lack of global convergence guarantees. A popular approach to tackle challenging non-convex optimization problems is overparameterization--- on input an $n \times n \times n$ tensor of rank $r$, the algorithm can output a decomposition of potentially rank $k$ (potentially larger than $r$). On the theoretical side, overparameterization for iterative methods is challenging to reason about and requires new techniques. The work of Wang et al., (NeurIPS 2020) makes progress by showing that a variant of gradient descent globally converges when overparameterized to $k=O(r^{7.5} \log n)$. Our main result shows that overparameterization provably enables global convergence of ALS: on input a third order $n \times n \times n$ tensor with a decomposition of rank $r \ll n$, ALS overparameterized with rank $k=O(r^2)$ achieves global convergence with high probability under random initialization. Moreover our analysis also gives guarantees for the more general low-rank approximation problem. The analysis introduces new techniques for understanding iterative methods in the overparameterized regime based on new matrix anticoncentration arguments.

In-Context Learning (ICL) is a technique where large language models (LLMs) leverage multiple demonstrations (i.e., examples) to perform tasks. With the recent expansion of LLM context windows, many-shot ICL (generally with more than 50 demonstrations) can lead to significant performance improvements on a variety of language tasks such as text classification and question answering. Nevertheless, ICL faces the issue of demonstration order instability (ICL-DOI), which means that performance varies significantly depending on the order of demonstrations. Moreover, ICL-DOI persists in many-shot ICL, validated by our thorough experimental investigation. Current strategies for handling ICL-DOI are not applicable to many-shot ICL due to two critical challenges: (1) Most existing methods assess demonstration order quality by first prompting the LLM, then using heuristic metrics based on the LLM's predictions. In the many-shot scenarios, these metrics without theoretical grounding become unreliable, where the LLMs struggle to effectively utilize information from long input contexts, making order distinctions less clear. The requirement to examine all orders for the large number of demonstrations is computationally infeasible due to the super-exponential complexity of the order space in many-shot ICL. To tackle the first challenge, we design a demonstration order evaluation metric based on information theory for measuring order quality, which effectively quantifies the usable information gain of a given demonstration order. To address the second challenge, we propose a hierarchical demonstration order optimization method named \texttt{HIDO} that enables a more refined exploration of the order space, achieving high ICL performance without the need to evaluate all possible orders. Extensive experiments on multiple LLMs and real-world datasets demonstrate that our \texttt{HIDO} method consistently and efficiently outperforms other baselines. Our code project can be found at https://github.com/YinhanHe123/HIDO/.

Combinatorial optimization problems involving multiple agents are notoriously challenging due to their NP-hard nature and the necessity for effective agent coordination. Despite advancements in learning-based methods, existing approaches often face critical limitations, including suboptimal agent coordination, poor generalization, and high computational latency. To address these issues, we propose PARCO (Parallel AutoRegressive Combinatorial Optimization), a general reinforcement learning framework designed to construct high-quality solutions for multi-agent combinatorial tasks efficiently. To this end, PARCO integrates three key novel components: (1) transformer-based communication layers to enable effective agent collaboration during parallel solution construction, (2) a multiple pointer mechanism for low-latency, parallel agent decision-making, and (3) priority-based conflict handlers to resolve decision conflicts via learned priorities. We evaluate PARCO in multi-agent vehicle routing and scheduling problems, where our approach outperforms state-of-the-art learning methods, demonstrating strong generalization ability and remarkable computational efficiency. We make our source code publicly available to foster future research: https://github.com/ai4co/parco.

Invariant object recognition-the ability to identify objects despite changes in appearance-is a hallmark of visual processing in the brain, yet its understanding remains a central challenge in systems neuroscience. Artificial neural networks trained to predict neural responses to visual stimuli (“digital twins”) could provide a powerful framework for studying such complex computations in silico. However, while current models accurately capture single-neuron responses within individual visual areas, their ability to reproduce how populations of neurons represent object identity, and how these representations transform across the cortical hierarchy, remains largely unexplored. Here we examine key functional signatures observed experimentally and find that current models account for hierarchical changes in basic single-neuron properties, such as receptive field size, but fail to capture more complex population-level phenomena, particularly invariant object representations. To address this gap, we introduce a biologically inspired hierarchical readout scheme that mirrors cortical anatomy, modeling each visual area as a projection from a distinct depth within a shared core network. This approach significantly improves the prediction of population-level representational transformations, outperforming standard models that use only the final layer, as well as alternatives with modified architecture, regularization, and loss function. Our results suggest that incorporating anatomical information provides a strong inductive bias in digital twin models, enabling them to better capture general principles of brain function.

The go-to strategy to apply deep networks in settings where uncertainty informs decisions—ensembling multiple training runs with random initializations—is ill-suited for the extremely large-scale models and practical fine-tuning workflows of today. We introduce a new cost-effective strategy for improving the uncertainty quantification and downstream decisions of a large model (e.g. a fine-tuned ViT-B): coupling it with a less accurate but much smaller "sidekick" (e.g. a fine-tuned ResNet-34) with a fraction of the computational cost. We propose aggregating the predictions of this *Asymmetric Duo* by simple learned weighted averaging. Surprisingly, despite their inherent asymmetry, the sidekick model almost never harms the performance of the larger model. In fact, across five image classification benchmarks, and a variety of model architectures and training schemes (including soups), Asymmetric Duos significantly improve accuracy, uncertainty quantification, and selective classification metrics with only ${\sim}10-20$% more computation.

Collaborative perception expands the perception range by sharing information among agents, effectively improving task performance. Immutable heterogeneity poses a significant challenge in collaborative perception, as participating agents may employ different and fixed perception models. This leads to domain gaps in the intermediate features shared among agents, consequently degrading collaborative performance. Aligning the features of all agents to a common representation can eliminate domain gaps with low training cost. However, in existing methods, the common representation is designated as the representation of a specific agent, making it difficult for agents with significant domain discrepancies from this specific agent to achieve proper alignment. This paper proposes NegoCollab, a heterogeneous collaboration method based on negotiated common representation. It achieves bidirectional transformation of each modality's features between local representation space and common representation space through paired sender-receiver, thereby eliminating domain gaps. The common representation in NegoCollab is negotiated from local representations of each modality's agent via a negotiator introduced during training, effectively reducing inherent domain discrepancies with each local representation. Furthermore, to better align local representations with the multimodal common representation, we introduce both structural alignment loss and pragmatic alignment loss alongside the conventional distribution alignment loss during supervised training, enabling comprehensive knowledge distillation from the common representation to the senders. The experimental results demonstrate that NegoCollab significantly outperforms existing methods in common representation-based collaboration approaches. The negotiation-based mechanism for acquiring common representations provides more diverse and reliable alternatives for establishing common representations required in heterogeneous collaboration perception.

Non-conservative uncertainty bounds are key for both assessing an estimation algorithm’s accuracy and in view of downstream tasks, such as its deployment in safety-critical contexts. In this paper, we derive a tight, non-asymptotic uncertainty bound for kernel-based estimation, which can also handle correlated noise sequences. Its computation relies on a mild norm-boundedness assumption on the unknown function and the noise, returning the worst-case function realization within the hypothesis class at an arbitrary query input location. The value of this function is shown to be given in terms of the posterior mean and covariance of a Gaussian process for an optimal choice of the measurement noise covariance. By rigorously analyzing the proposed approach and comparing it with other results in the literature, we show its effectiveness in returning tight and easy-to-compute bounds for kernel-based estimates.

Geometric diffusion models have shown remarkable success in molecular dynamics and structure generation. However, efficiently fine-tuning them for downstream tasks with varying geometric controls remains underexplored. In this work, we propose an SE(3)-equivariant adapter framework (GeoAda) that enables flexible and parameter-efficient fine-tuning for controlled generative tasks without modifying the original model architecture. GeoAda introduces a structured adapter design: control signals are first encoded through coupling operators, then processed by a trainable copy of selected base model layers, and finally projected back via decoupling operators followed by an equivariant zero-initialized convolution. By fine-tuning only these lightweight adapter modules, GeoAda preserves the model’s geometric consistency while mitigating overfitting and catastrophic forgetting. We theoretically prove that the proposed adapters maintain SE(3)-equivariance, ensuring that the geometric inductive biases of the pretrained diffusion model remain intact during adaptation. We demonstrate the wide applicability of \method across diverse geometric control types, including frame control, global control, subgraph control, and a broad range of application domains such as particle dynamics, molecular dynamics, human motion prediction, and molecule generation. Empirical results show that GeoAda achieves state-of-the-art fine-tuning performance while preserving original task accuracy, whereas other baselines experience significant performance degradation due to overfitting and catastrophic forgetting.

Instruction tuning improves the ability of large language models (LLMs) to follow diverse human instructions, but achieving strong performance on specific target tasks remains challenging. A critical bottleneck is selecting the most relevant data to maximize task-specific performance. Existing data selection approaches include unstable influence-based methods and more stable distribution alignment methods, the latter of which critically rely on the underlying sample representation. In practice, most distribution alignment methods, from shallow features (e.g., BM25) to neural embeddings (e.g., BGE, LLM2Vec), may fail to capture how the model internally processes samples. To bridge this gap, we adopt a model-centric strategy in which each sample is represented by its neuronal activation pattern in the model, directly reflecting internal computation. However, directly using raw neuron activations leads to spurious similarity between unrelated samples due to neuron polysemanticity, where a single neuron may respond to multiple, unrelated concepts. To address this, we employ sparse autoencoders to disentangle polysemantic activations into sparse, monosemantic representations, and introduce a dedicated similarity metric for this space to better identify task-relevant data. Comprehensive experiments across multiple instruction datasets, models, tasks, and selection ratios show that our approach consistently outperforms existing data selection baselines in both stability and task-specific performance.

Current text-to-image (T2I) generation models struggle to align spatial composition with the input text, especially in complex scenes. Even layout-based approaches yield suboptimal spatial control, as their generation process is decoupled from layout planning, making it difficult to refine the layout during synthesis. We present CoT-Diff, a framework that brings step-by-step CoT-style reasoning into T2I generation by tightly integrating Multimodal Large Language Model (MLLM)-driven 3D layout planning with the diffusion process. CoT-Diff enables layout-aware reasoning inline within a single diffusion round: at each denoising step, the MLLM evaluates intermediate predictions, dynamically updates the 3D scene layout, and continuously guides the generation process. The updated layout is converted into semantic conditions and depth maps, which are fused into the diffusion model via a condition-aware attention mechanism, enabling precise spatial control and semantic injection. Experiments on 3D Scene benchmarks show that CoT-Diff significantly improves spatial alignment and compositional fidelity, and outperforms the state-of-the-art method by 34.7% in complex scene spatial accuracy, thereby validating the effectiveness of this entangled generation paradigm.

Memorization in language models is a critical yet poorly understood phenomenon. In this work, we investigate memorization in transformer-based language models by analyzing their memorization dynamics during training over multiple epochs. We find that memorization is neither a constant accumulation of sequences nor simply dictated by the recency of exposure to these sequences. Instead, much like generalization, memorization appears to be driven by pattern recognition. Tracking memorization dynamics in mixed datasets, we observe that models memorize different sub-datasets in distinct bursts, suggesting that each subset is associated with unique underlying patterns, and that the model prefers to learn these patterns in a consistent order. We also find that easily learnable patterns tend to support generalization on unseen data, while more complex patterns do not. Furthermore, in datasets with weak or absent patterns, larger models may delay memorization relative to smaller ones, a behavior we term $\textit{overthinking}$. Our results show that the subset of sequences memorized by a model over time is not arbitrary, and give insights into the internal processes a model goes through during training. Our code is available at: https://github.com/mdrpanwar/memorization-patterns.

Recent advances in large-scale neural recordings have enabled accurate decoding of behavior and cognitive states, yet decoding anatomical regions remains underexplored, despite being crucial for consistent targeting in multiday recordings and effective deep brain stimulation. Current approaches typically rely on external anatomical information, from atlas-based planning to post hoc histology, which are limited in precision, longitudinal applicability, and real-time feedback. In this work, we develop a self-supervised learning framework, Lfp2vec, to infer anatomical regions directly from the neural signal in vivo. We adapt an audio-pretrained transformer model by continuing self-supervised training on a large corpus of unlabeled local-field-potential (LFP) data, then fine-tuning for anatomical region decoding. Ablations show that combining out-of-domain initialization with in-domain self-supervision outperforms training from scratch. We demonstrate that our method achieves strong zero-shot generalization across different labs and probe geometries, and outperforming state-of-the-art self-supervised models on electrophysiology data. The learned embeddings form anatomically coherent clusters and transfer effectively to downstream tasks like disease classification with minimal fine-tuning. Altogether, our approach enables zero-shot prediction of brain regions in novel subjects, demonstrates that LFP signals encode rich anatomical information, and establishes self-supervised learning on raw LFP as a foundation to learn representations that can be tuned for diverse neural decoding tasks. Code to reproduce our results is found in the github repository at https://github.com/tianxiao18/lfp2vec.

Process Reward Models (PRMs) have shown promise in enhancing the mathematical reasoning capabilities of Large Language Models (LLMs) through Test-Time Scaling (TTS). However, their integration into multimodal reasoning remains largely unexplored. In this work, we take the first step toward unlocking the potential of PRMs in multimodal mathematical reasoning. We identify three key challenges: (i) the scarcity of high-quality reasoning data constrains the capabilities of foundation Multimodal Large Language Models (MLLMs), which imposes further limitations on the upper bounds of TTS and reinforcement learning (RL); (ii) a lack of automated methods for process labeling within multimodal contexts persists; (iii) the employment of process rewards in unimodal RL faces issues like reward hacking, which may extend to multimodal scenarios. To address these issues, we introduce URSA, a three-stage Unfolding multimodal pRocess-Supervision Aided training framework. We first construct MMathCoT-1M, a high-quality large-scale multimodal Chain-of-Thought (CoT) reasoning dataset, to build a stronger math reasoning foundation MLLM, URSA-8B. Subsequently, we go through an automatic process to synthesize process supervision data, which emphasizes both logical correctness and perceptual consistency. We introduce DualMath-1.1M to facilitate the training of URSA-8B-RM. Finally, we propose Process-Supervised Group-Relative-Policy-Optimization (PS-GRPO), pioneering a multimodal PRM-aided online RL method that outperforms vanilla GRPO. With PS-GRPO application, URSA-8B-PS-GRPO outperforms Gemma3-12B and GPT-4o by 8.4% and 2.7% on average across 6 benchmarks.

We study fundamental connections between coalition formation games and clustering, illustrating the cross-disciplinary relevance of these concepts. We focus on graphical hedonic games where agents' preferences are compactly represented by a friendship graph and an enemy graph. In the context of clustering, friendship relations naturally align with data point similarities, whereas enmity corresponds to dissimilarities. We consider two stability notions based on single-agent deviations: local popularity and local stability. Exploring these concepts from an algorithmic viewpoint, we design efficient mechanisms for finding locally stable or locally popular partitions. Besides gaining theoretical insight into the computational complexity of these problems, we perform simulations that demonstrate how our algorithms can be successfully applied in clustering and community detection. Our findings highlight the interplay between coalition formation games and data-driven clustering techniques, offering fresh perspectives and applications in both areas.

Understanding causal relationships between variables is fundamental across scientific disciplines. Most causal discovery algorithms rely on two key assumptions: (i) all variables are observed, and (ii) the underlying causal graph is acyclic. While these assumptions simplify theoretical analysis, they are often violated in real-world systems, such as biological networks. Existing methods that account for confounders either assume linearity or struggle with scalability. To address these limitations, we propose DCCD-CONF, a novel framework for differentiable learning of nonlinear cyclic causal graphs in the presence of unmeasured confounders using interventional data. Our approach alternates between optimizing the graph structure and estimating the confounder distribution by maximizing the log-likelihood of the data. Through experiments on synthetic data and real-world gene perturbation datasets, we show that DCCD-CONF outperforms state-of-the-art methods in both causal graph recovery and confounder identification. Additionally, we provide consistency guarantees for our framework, reinforcing its theoretical soundness.

Reinforcement Learning (RL) traditionally relies on scalar reward signals, limiting its ability to leverage the rich semantic knowledge often available in real-world tasks. In contrast, humans learn efficiently by combining numerical feedback with language, prior knowledge, and common sense. We introduce Prompted Policy Search (ProPS), a novel RL method that unifies numerical and linguistic reasoning within a single framework. Unlike prior work that augments existing RL components with language, ProPS places a large language model (LLM) at the center of the policy optimization loop—directly proposing policy updates based on both reward feedback and natural language input. We show that LLMs can perform numerical optimization in-context, and that incorporating semantic signals, such as goals, constraints, and strategy hints can lead to more informed exploration and sample-efficient learning. ProPS is evaluated across 15 Gymnasium tasks, spanning classic control, Atari games, and MuJoCo environments, and compared to seven widely-adopted RL algorithms (e.g., PPO, SAC, TRPO). It outperforms all baselines on 8 out of 15 tasks and demonstrates substantial gains when provided with domain knowledge. These results highlight the potential of unifying semantics and numerics for transparent, generalizable, and human-aligned reinforcement learning.

Self-correction of large language models (LLMs) emerges as a critical component for enhancing their reasoning performance. Although various self-correction methods have been proposed, a comprehensive evaluation of these methods remains largely unexplored, and the question of whether LLMs can truly correct themselves is a matter of significant interest and concern. In this study, we introduce **CorrectBench**, a benchmark developed to evaluate the effectiveness of self-correction strategies, including intrinsic, external, and fine-tuned approaches, across three tasks: commonsense reasoning, mathematical reasoning, and code generation. Our findings reveal that: 1) Self-correction methods can improve accuracy, especially for complex reasoning tasks; 2) Mixing different self-correction strategies yields further improvements, though it reduces efficiency; 3) Reasoning LLMs (e.g., DeepSeek-V3) have limited optimization under additional self-correction methods and have high time costs. Interestingly, a comparatively simple chain-of-thought (CoT) baseline demonstrates competitive accuracy and efficiency. These results underscore the potential of self-correction to enhance LLM's reasoning performance while highlighting the ongoing challenge of improving their efficiency. Consequently, we advocate for further research focused on optimizing the balance between reasoning capabilities and operational efficiency.

AI-powered attacks on Learning with Errors (LWE)—an important hard math problem in post-quantum cryptography—rival or outperform "classical" attacks on LWE under certain parameter settings. Despite the promise of this approach, a dearth of accessible data limits AI practitioners' ability to study and improve these attacks. Creating LWE data for AI model training is time- and compute-intensive and requires significant domain expertise. To fill this gap and accelerate AI research on LWE attacks, we propose the TAPAS datasets, a ${\bf t}$oolkit for ${\bf a}$nalysis of ${\bf p}$ost-quantum cryptography using ${\bf A}$I ${\bf s}$ystems. These datasets cover several LWE settings and can be used off-the-shelf by AI practitioners to prototype new approaches to cracking LWE. This work documents TAPAS dataset creation, establishes attack performance baselines, and lays out directions for future work.