NeurIPS 2025 Papers

Large Language Models (LLMs) offer substantial promise for clinical natural language processing (NLP); however, a lack of standardized benchmarking methodologies limits their objective evaluation and practical translation. To address this gap, we introduce ClinBench, an open-source, multi-model, multi-domain benchmarking framework. ClinBench is designed for the rigorous evaluation of LLMs on important structured information extraction tasks (e.g., tumor staging, histologic diagnoses, atrial fibrillation, and social determinants of health) from unstructured clinical notes. The framework standardizes the evaluation pipeline by: (i) operating on consistently structured input datasets; (ii) employing dynamic, YAML-based prompting for uniform task definition; and (iii) enforcing output validation via JSON schemas, supporting robust comparison across diverse LLM architectures. We demonstrate ClinBench through a large-scale study of 11 prominent LLMs (e.g., GPT-4o series, LLaMA3 variants, Mixtral) across three clinical domains using configurations of public datasets (TCGA for lung cancer, MIMIC-IV-ECG for atrial fibrillation, and MIMIC notes for SDOH). Our results reveal significant performance-efficiency trade-offs. For example, when averaged across the four benchmarked clinical extraction tasks, GPT-3.5-turbo achieved a mean F1 score of 0.83 with a mean runtime of 16.8 minutes. In comparison, LLaMA3.1-70b obtained a similar mean F1 of 0.82 but required a substantially longer mean runtime of 42.7 minutes. GPT-4o-mini also presented a favorable balance with a mean F1 of 0.81 and a mean runtime of 13.4 minutes. ClinBench provides a unified, extensible framework and empirical insights for reproducible, fair LLM benchmarking in clinical NLP. By enabling transparent and standardized evaluation, this work advances data-centric AI research, informs model selection based on performance, cost, and clinical priorities, and supports the effective integration of LLMs into healthcare. The framework and evaluation code are publicly available at https://github.com/ismaelvillanuevamiranda/ClinBench/.

Reinforcement fine-tuning (RFT) is essential for enhancing the reasoning and generalization capabilities of large language models, but its success heavily relies on the quality of the training data. While data selection has been extensively studied in supervised learning, its role in reinforcement learning, particularly during the RFT stage, remains largely underexplored. In this work, we introduce RFT-Inf, the first influence estimator designed for data in reinforcement learning. RFT-Inf quantifies the importance of each training example by measuring how its removal affects the final training reward, offering a direct estimate of its contribution to model learning. To ensure scalability, we propose a first-order approximation of the RFT-Inf score by backtracking through the optimization process and applying temporal differentiation to the sample-wise influence term, along with a first-order Taylor approximation to adjacent time steps. This yields a lightweight, gradient-based estimator that evaluates the alignment between an individual sample’s gradient and the average gradient direction of all training samples, where a higher degree of alignment implies greater training utility. Extensive experiments demonstrate that RFT-Inf consistently improves reward performance and accelerates convergence in reinforcement fine-tuning.

The practical use of reinforcement learning (RL) requires handling diverse settings, including online, offline, and offline-to-online learning. Instead of developing separate algorithms for each setting, we propose Uni-RL, a unified model-free RL framework that addresses all these scenarios within a single formulation. Uni-RL builds on the Implicit Value Regularization (IVR) framework and generalizes its dataset behavior constraint to the constraint w.r.t a reference policy, yielding an unified value learning objective for general settings. The reference policy is chosen to be the target policy in the online setting and the behavior policy in the offline setting. Using an iteratively refined behavior policy solves the over-constrained problem of directly applying IVR in the online setting, it provides an implicit trust-region style update through the value function while being off-policy. Uni-RL also introduces an unified policy extraction objective that estimates in-sample policy gradient using only actions from the reference policy. This supports various policy classes and theoretically guaranntees less value estimation error and larger performance improvement over the reference policy. We evaluate Uni-RL on a range of standard RL benchmarks across online, offline, and offline-to-online settings. In online RL, Uni-RL achieves higher sample efficiency than both off-policy methods without trust-region updates and on-policy methods with trust-region updates. In offline RL, Uni-RL retains the benefits of in-sample learning while outperforming IVR through better policy extraction. In offline-to-online RL, Uni-RL beats both constraint-based methods and unconstrained approaches by effectively balancing stability and adaptability.

While end-to-end video-to-audio generation has greatly improved, producing high-fidelity audio that authentically captures the nuances of visual content remains challenging. Like professionals in the creative industries, this generation requires sophisticated reasoning about items such as visual dynamics, acoustic environments, and temporal relationships. We present **ThinkSound**, a novel framework that leverages Chain-of-Thought (CoT) reasoning to enable stepwise, interactive audio generation and editing for videos. Our approach decomposes the process into three complementary stages: foundational foley generation that creates semantically coherent soundscapes, interactive object-centric refinement through precise user interactions, and targeted editing guided by natural language instructions. At each stage, a multimodal large language model generates contextually aligned CoT reasoning that guides a unified audio foundation model. Furthermore, we introduce **AudioCoT**, a comprehensive dataset with structured reasoning annotations that establishes connections between visual content, textual descriptions, and sound synthesis. Experiments demonstrate that ThinkSound achieves state-of-the-art performance in video-to-audio generation across both audio metrics and CoT metrics, and excels in the out-of-distribution Movie Gen Audio benchmark. The project page is available at https://ThinkSound-Project.github.io.

We study the effect of normalization schemes on token representations in deep transformers. Modeling their evolution as interacting particles on the sphere, we show that normalization acts as a form of speed regulation. This perspective enables a unified analysis of several schemes---including **Post-LN**, **Pre-LN**, **Mix-LN**, **Peri-LN**, **nGPT**, and **LN-scaling**---revealing how they influence clustering dynamics and representation collapse. Our framework clarifies how different schemes shape token representations across layers and provides a principled basis for comparing them, identifying **Peri-LN** as a particularly effective choice.

We introduce a novel method for low-rank personalization of a generic model for head avatar generation. Prior work proposes generic models that achieve high-quality face animation by leveraging large-scale datasets of multiple identities. However, such generic models usually fail to synthesize unique identity-specific details, since they learn a general domain prior. To adapt to specific subjects, we find that it is still challenging to capture high-frequency facial details via popular solutions like low-rank adaptation (LoRA). This motivates us to propose a specific architecture, a Register Module, that enhances the performance of LoRA, while requiring only a small number of parameters to adapt to an unseen identity. Our module is applied to intermediate features of a pre-trained model, storing and re-purposing information in a learnable 3D feature space. To demonstrate the efficacy of our personalization method, we collect a dataset of talking videos of individuals with distinctive facial details, such as wrinkles and tattoos. Our approach faithfully captures unseen faces, outperforming existing methods quantitatively and qualitatively.

With the emergence of new storage and communication methods, the insertion, deletion, and substitution (IDS) channel has attracted considerable attention. However, many topics on the IDS channel and the associated Levenshtein distance remain open, making the invention of a novel IDS-correcting code a hard task. Furthermore, current studies on single-IDS-correcting code misalign with the requirements of applications which necessitates the correcting of multiple errors. Compromise solutions have involved shortening codewords to reduce the chance of multiple errors. However, the code rates of existing codes are poor at short lengths, diminishing the overall storage density. In this study, a novel method is introduced for designing high-code-rate single-IDS-correcting codewords through deep Levenshtein distance embedding. A deep learning model is utilized to project the sequences into embedding vectors that preserve the Levenshtein distances between the original sequences. This embedding space serves as a proxy for the complex Levenshtein domain, within which algorithms for codeword search and segment correcting is developed. While the concept underpinning this approach is straightforward, it bypasses the mathematical challenges typically encountered in code design. The proposed method results in a code rate that outperforms existing combinatorial solutions, particularly for designing short-length codewords.

Despite its empirical success, the theoretical foundations of self-supervised contrastive learning (CL) are not yet fully established. In this work, we address this gap by showing that standard CL objectives implicitly approximate a supervised variant we call the negatives-only supervised contrastive loss (NSCL), which excludes same-class contrasts. We prove that the gap between the CL and NSCL losses vanishes as the number of semantic classes increases, under a bound that is both label-agnostic and architecture-independent. We characterize the geometric structure of the global minimizers of the NSCL loss: the learned representations exhibit augmentation collapse, within-class collapse, and class centers that form a simplex equiangular tight frame. We further introduce a new bound on the few-shot error of linear-probing. This bound depends on two measures of feature variability—within-class dispersion and variation along the line between class centers. We show that directional variation dominates the bound and that the within-class dispersion's effect diminishes as the number of labeled samples increases. These properties enable CL and NSCL-trained representations to support accurate few-shot label recovery using simple linear probes. Finally, we empirically validate our theoretical findings: the gap between CL and NSCL losses decays at a rate of $\mathcal{O}(\frac{1}{\#\text{classes}})$; the two losses are highly correlated; minimizing the CL loss implicitly brings the NSCL loss close to the value achieved by direct minimization; and the proposed few-shot error bound provides a tight estimate of probing performance in practice.

Preserving intellectual property (IP) within a pre-trained diffusion model is critical for protecting the model's copyright and preventing unauthorized model deployment. In this regard, model watermarking is a common practice for IP protection that embeds traceable information within models and allows for further verification. Nevertheless, existing watermarking schemes often face challenges due to their vulnerability to fine-tuning, limiting their practical application in general pre-training and fine-tuning paradigms. Inspired by using mode connectivity to analyze model performance between a pair of connected models, we investigate watermark vulnerability by leveraging Linear Mode Connectivity (LMC) as a proxy to analyze the fine-tuning dynamics of watermark performance. Our results show that existing watermarked models tend to converge to sharp minima in the loss landscape, thus making them vulnerable to fine-tuning. To tackle this challenge, we propose **RoMa**, a **Ro**bust **M**odel w**a**termarking scheme that improves the robustness of watermarks against fine-tuning. Specifically, RoMa decomposes watermarking into two components, including *Embedding Functionality*, which preserves reliable watermark detection capability, and *Path-specific Smoothness*, which enhances the smoothness along the watermark-connected path to improve robustness. Extensive experiments on benchmark datasets MS-COCO-2017 and CUB-200-2011 demonstrate that RoMa significantly improves watermark robustness against fine-tuning while maintaining generation quality, outperforming baselines. The code is available at [https://github.com/xiekks/RoMa](https://github.com/xiekks/RoMa).

We argue that the machine learning value chain is structurally unsustainable due to an economic data processing inequality: each state in the data cycle from inputs to model weights to synthetic outputs refines technical signal but strips economic equity from data generators. We show, by analyzing seventy-three public data deals, that the majority of value accrues to aggregators, with documented creator royalties rounding to zero and widespread opacity of deal terms. This is not just an economic welfare concern: as data and its derivatives become economic assets, the feedback loop that sustains current learning algorithms is at risk. We identify three structural faults - missing provenance, asymmetric bargaining power, and non-dynamic pricing - as the operational machinery of this inequality. In our analysis, we trace these problems along the machine learning value chain and propose an Equitable Data-Value Exchange (EDVEX) Framework to enable a minimal market that benefits all participants. Finally, we outline research directions where our community can make concrete contributions to data deals and contextualize our position with related and orthogonal viewpoints.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as `safety' and `robustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

We consider the problem of sampling distributions stemming from non-convex potentials with Unadjusted Langevin Algorithm (ULA). We prove the stability of the discrete-time ULA to drift approximations under the assumption that the potential is strongly convex at infinity. In many context, e.g. imaging inverse problems, potentials are non-convex and non-smooth. Proximal Stochastic Gradient Langevin Algorithm (PSGLA) is a popular algorithm to handle such potentials. It combines the forward-backward optimization algorithm with a ULA step. Our main stability result combined with properties of the Moreau envelope allows us to derive the first proof of convergence of the PSGLA for non-convex potentials. We empirically validate our methodology on synthetic data and in the context of imaging inverse problems. In particular, we observe that PSGLA exhibits faster convergence rates than Stochastic Gradient Langevin Algorithm for posterior sampling while preserving its restoration properties.

This paper proposes Virtual Fitting Room (VFR), a novel video generative model that produces arbitrarily long virtual try-on videos. Our VFR models long video generation tasks as an auto-regressive, segment-by-segment generation process, eliminating the need for resource-intensive generation and lengthy video data, while providing the flexibility to generate videos of arbitrary length. The key challenges of this task are twofold: ensuring local smoothness between adjacent segments and maintaining global temporal consistency across different segments. To address these challenges, we propose our VFR framework, which ensures smoothness through a prefix video condition and enforces consistency with the anchor video — a 360°-view video that comprehensively captures the human's whole-body appearance. Our VFR generates minute-scale virtual try-on videos with both local smoothness and global temporal consistency under various motions, making it a pioneering work in long virtual try-on video generation. Project Page: https://immortalco.github.io/VirtualFittingRoom/.

This position paper argues that post-deployment monitoring in clinical AI is underdeveloped and proposes statistically valid and label-efficient testing frameworks as a principled foundation for ensuring reliability and safety in real-world deployment. A recent review found that only 9\% of FDA-registered AI-based healthcare tools include a post-deployment surveillance plan. Existing monitoring approaches are often manual, sporadic, and reactive, making them ill-suited for the dynamic environments in which clinical models operate. We contend that post-deployment monitoring should be grounded in label-efficient and statistically valid testing frameworks, offering a principled alternative to current practices. We use the term "statistically valid" to refer to methods that provide explicit guarantees on error rates (e.g., Type I/II error), enable formal inference under pre-defined assumptions, and support reproducibility—features that align with regulatory requirements. Specifically, we propose that the detection of changes in the data and model performance degradation should be framed as distinct statistical hypothesis testing problems. Grounding monitoring in statistical rigor ensures a reproducible and scientifically sound basis for maintaining the reliability of clinical AI systems. Importantly, it also opens new research directions for the technical community---spanning theory, methods, and tools for statistically principled detection, attribution, and mitigation of post-deployment model failures in real-world settings.

Neural Architecture Search (NAS) aims to identify high-performance networks within a defined search space. Training-free metrics have been proposed to estimate network performance without actual training, reducing NAS deployment costs. However, individual training-free metrics often capture only partial architectural features, and their estimation capabilities are different in various tasks. Combining multiple training-free metrics has been explored to enhance scalability across tasks. Yet, these methods typically optimize global metric combinations over the entire search space, overlooking the varying sensitivities of different architectures to specific metrics, which may limit the final architectures' performance. To address these challenges, we propose the Per-Architecture Training-Free Metric Optimization NAS (PO-NAS) algorithm. This algorithm: (a) Integrates multiple training-free metrics as auxiliary scores, dynamically optimizing their combinations using limited real-time training data, without relying on benchmarks; (b) Individually optimizes metric combinations for each architecture; (c) Integrates an evolutionary algorithm that leverages efficient predictions from surrogate models, enhancing search efficiency in large search spaces. Notably, PO-NAS combines the efficiency of training-free search with the robust performance of training-based evaluations. Extensive experiments demonstrate the effectiveness of our approach. Our code has been made publicly available at https://anonymous.4open.science/r/PO-NAS-2953.

Mixture-of-Experts (MoE) models have become popular in machine learning, boosting performance by partitioning tasks across multiple experts. However, the need for several experts often results in high computational costs, limiting their application on resource-constrained devices with stringent real-time requirements, such as cochlear implants (CIs). We introduce the Omni-Expert (OE) – a simple and efficient solution that leverages feature transformations to achieve the 'divide-and-conquer' functionality of a full MoE ensemble in a single expert model. We demonstrate the effectiveness of the OE using phoneme-specific time-frequency masking for speech dereverberation in a CI. Empirical results show that the OE delivers statistically significant improvements in objective intelligibility measures of CI vocoded speech at different levels of reverberation across various speech datasets at a much reduced computational cost relative to a counterpart MoE.

Contrastive learning is a powerful technique for discovering meaningful data representations by optimizing objectives based on $\textit{contrastive information}$, often given as a set of weighted triplets $\{(x_i, y_i^+, z_{i}^-)\}_{i = 1}^m$ indicating that an "anchor" $x_i$ is more similar to a "positive" example $y_i$ than to a "negative" example $z_i$. The goal is to find representations (e.g., embeddings in $\mathbb{R}^d$ or a tree metric) where anchors are placed closer to positive than to negative examples. While finding $\textit{global}$ optima of contrastive objectives is $\mathsf{NP}$-hard, the complexity of finding $\text{\textit{local}}$ optima---representations that do not improve by local search algorithms such as gradient-based methods---remains open. Our work settles the complexity of finding local optima in various contrastive learning problems by proving $\mathsf{PLS}$-hardness in discrete settings (e.g., maximize satisfied triplets) and $\mathsf{CLS}$-hardness in continuous settings (e.g., minimize Triplet Loss), where $\mathsf{PLS}$ (Polynomial Local Search) and $\mathsf{CLS}$ (Continuous Local Search) are well-studied complexity classes capturing local search dynamics in discrete and continuous optimization, respectively. Our results imply that no polynomial time algorithm (local search or otherwise) can find a local optimum for various contrastive learning problems, unless $\mathsf{PLS}\subseteq\mathsf{P}$ (or $\mathsf{CLS}\subseteq \mathsf{P}$ for continuous problems). Even in the unlikely scenario that $\mathsf{PLS}\subseteq\mathsf{P}$ (or $\mathsf{CLS}\subseteq \mathsf{P}$), our reductions imply that there exist instances where local search algorithms need exponential time to reach a local optimum, even for $d=1$ (embeddings on a line).

Binary (0-1) integer programming (BIP) is pivotal in scientific domains requiring discrete decision-making. As the advance of AI computing, recent works explore neural network-based solver for integer linear programming (ILP) problems. Yet, they lack scalability for tackling nonlinear challenges. To handle nonlinearities, state-of-the-art Branch-and-Cut solvers employ linear relaxations, leading to exponential growth in auxiliary variables and severe computation limitations. To overcome these limitations, we propose BIPNN, an unsupervised learning framework to solve BIP problems via hypergraph neural networks (HyperGNN). Specifically, (i) BIPNN reformulates BIPs-constrained, discrete, and nonlinear (sin, log, exp) optimization problems-into unconstrained, differentiable, and polynomial loss functions. The reformulation stems from the observation of a precise one-to-one mapping between polynomial BIP objectives and hypergraph structures, enabling the unsupervised training of HyperGNN to optimize BIP problems in an end-to-end manner. On this basis, (ii) we propose a GPU-accelerated and continuous-annealing-enhanced training pipeline for BIPNN. The pipeline enables BIPNN to optimize large-scale nonlinear terms in BIPs fully in parallel via straightforward gradient descent, thus significantly reducing the training cost while ensuring the generation of discrete, high-quality solutions. Extensive experiments on synthetic and real-world datasets highlight the superiority of our approach.

We study a sequential decision-making problem on a $n$-node graph $\mathcal{G}$ where each node has an unknown label from a finite set $\mathbf{\Omega}$, drawn from a joint distribution $\mathcal{P}$ that is Markov with respect to $\mathcal{G}$. At each step, selecting a node reveals its label and yields a label-dependent reward. The goal is to adaptively choose nodes to maximize expected accumulated discounted rewards. We impose a frontier exploration constraint, where actions are limited to neighbors of previously selected nodes, reflecting practical constraints in settings such as contact tracing and robotic exploration. We design a Gittins index-based policy that applies to general graphs and is provably optimal when $\mathcal{G}$ is a forest. Our implementation runs in $\mathcal{O}(n^2 \cdot |\mathbf{\Omega}|^2)$ time while using $\mathcal{O}(n \cdot |\mathbf{\Omega}|^2)$ oracle calls to $\mathcal{P}$ and $\mathcal{O}(n^2 \cdot |\mathbf{\Omega}|)$ space. Experiments on synthetic and real-world graphs show that our method consistently outperforms natural baselines, including in non-tree, budget-limited, and undiscounted settings. For example, in HIV testing simulations on real-world sexual interaction networks, our policy detects nearly all positive cases with only half the population tested, substantially outperforming other baselines.