NeurIPS 2025 Papers

It is difficult to evaluate machine learning classifiers without large labeled datasets, which are often unavailable. In contrast, unlabeled data is plentiful, but not easily used for evaluation. Here, we introduce Semi-Supervised Model Evaluation (SSME), a method that uses both labeled and unlabeled data to evaluate machine learning classifiers. The key idea is to estimate the joint distribution of ground truth labels and classifier scores using a semi-supervised mixture model. The semi-supervised mixture model allows SSME to learn from three sources of information: unlabeled data, multiple classifiers, and probabilistic classifier scores. Once fit, the mixture model enables estimation of any metric that is a function of classifier scores and ground truth labels (e.g., accuracy or AUC). We derive theoretical bounds on the error of these estimates, showing that estimation error decreases with the number of classifiers and the amount of unlabeled data. We present experiments in four domains where obtaining large labeled datasets is often impractical: healthcare, content moderation, molecular property prediction, and text classification. Our results demonstrate that SSME estimates performance more accurately than do competing methods, reducing error by 5.1x relative to using labeled data alone and 2.4x relative to the next best method.

Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment and confirmation bias, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for *belief entrenchment* in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates are unpredictable from the current belief. We propose the unsupervised, regression-based *Martingale Score* to measure violations of this property, which signal deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains including event forecasting, value-laden questions, and academic paper review, we find such violations to be widespread across models and setups, where the current belief positively predicts future belief updates, a phenomenon which we term *belief entrenchment*. We identify the models, reasoning techniques, and domains more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of a reasoning process.

The growing scale of evaluation tasks has led to the widespread adoption of automated evaluation using LLMs, a paradigm known as “LLM-as-a-judge”. However, improving its alignment with human preferences without complex prompts or fine-tuning remains challenging. Previous studies mainly optimize based on shallow outputs, overlooking rich cross-layer representations. In this work, motivated by preliminary findings that middle-to-upper layers encode semantically and task-relevant representations that are often more aligned with human judgments than the final layer, we propose LAGER, a post-hoc, plug-and-play framework for improving the alignment of LLM-as-a-Judge point-wise evaluations with human scores by leveraging internal representations. LAGER produces fine-grained judgment scores by aggregating cross-layer score-token logits and computing the expected score from a softmax-based distribution, while keeping the LLM backbone frozen and ensuring no impact on the inference process. LAGER fully leverages the complementary information across different layers, overcoming the limitations of relying solely on the final layer. We evaluate our method on the standard alignment benchmarks Flask, HelpSteer, and BIGGen using Spearman correlation, and find that LAGER achieves improvements of up to 7.5% over the best baseline across these benchmarks. Without reasoning steps, LAGER matches or outperforms reasoning-based methods. Experiments on downstream applications, such as data selection and emotional understanding, further show the generalization of LAGER.

Modern language model (LM) training has been divided into multiple stages, making it difficult for downstream developers to evaluate the impact of design choices made at each stage. We present EvoLM, a model suite that enables systematic and transparent analysis of LMs' training dynamics across pre-training, continued pre-training, supervised fine-tuning, and reinforcement learning. By training over 100 LMs with 1B and 4B parameters from scratch, we rigorously evaluate both upstream (language modeling) and downstream (problem-solving) reasoning capabilities, including considerations of both in-domain and out-of-domain generalization. Key insights highlight the diminishing returns from excessive pre-training and post-training, the importance and practices of mitigating forgetting during domain-specific continued pre-training, the crucial role of continued pre-training in bridging pre-training and post-training phases, and various intricate trade-offs when configuring supervised fine-tuning and reinforcement learning. To facilitate open research and reproducibility, we release all pre-trained and post-trained models, training datasets for all stages, and our entire training and evaluation pipeline.

Large language models (LLMs) have shown impressive capabilities across tasks such as mathematics, coding, and reasoning, yet their learning ability, which is crucial for adapting to dynamic environments and acquiring new knowledge, remains underexplored. In this work, we address this gap by introducing a framework inspired by cognitive psychology and education. Specifically, we decompose general learning ability into three distinct, complementary dimensions: *Learning from Instructor* (acquiring knowledge via explicit guidance), *Learning from Concept* (internalizing abstract structures and generalizing to new contexts), and *Learning from Experience* (adapting through accumulated exploration and feedback). We conduct a comprehensive empirical study across the three learning dimensions and identify several insightful findings, such as (i) interaction improves learning; (ii) conceptual understanding is scale-emergent and benefits larger models; and (iii) LLMs are effective few-shot learners but not many-shot learners. Based on our framework and empirical findings, we introduce a benchmark that provides a unified and realistic evaluation of LLMs' general learning abilities across three learning cognition dimensions. It enables diagnostic insights and supports evaluation and development of more adaptive and human-like models.

Polyak–Ruppert averaging is a widely used technique to achieve the optimal asymptotic variance of stochastic approximation (SA) algorithms, yet its high-probability performance guarantees remain underexplored in general settings. In this paper, we present a general framework for establishing non-asymptotic concentration bounds for the error of averaged SA iterates. Our approach assumes access to individual concentration bounds for the unaveraged iterates and yields a sharp bound on the averaged iterates. We also construct an example, showing the tightness of our result up to constant multiplicative factors. As direct applications, we derive tight concentration bounds for contractive SA algorithms and for algorithms such as temporal difference learning and $Q$-learning with averaging, obtaining new bounds in settings where traditional analysis is challenging.

While machine learning models become more capable in discriminative tasks at scale, their ability to overcome biases introduced by training data has come under increasing scrutiny. Previous results suggest that there are two extremes of parameterization with very different behaviors: the population (underparameterized) setting where loss weighting is optimal and the separable overparameterized setting where loss weighting is ineffective at ensuring equal performance across classes. This work explores the regime of last layer retraining (LLR) in which the unseen limited (retraining) data is frequently inseparable and the model proportionately sized, falling between the two aforementioned extremes. We show, in theory and practice, that loss weighting is still effective in this regime, but that these weights _must_ take into account the relative overparameterization of the model.

How should we evaluate the quality of generative models? Many existing metrics focus on a model's producibility, i.e. the quality and breadth of outputs it can generate. However, the actual value from using a generative model stems not just from what it can produce but whether a user with a specific goal can produce an output that satisfies that goal. We refer to this property as steerability. In this paper, we first introduce a mathematical decomposition for quantifying steerability independently from producibility. Steerability is more challenging to evaluate than producibility because it requires knowing a user's goals. We address this issue by creating a benchmark task that relies on one key idea: sample an output from a generative model and ask users to reproduce it. We implement this benchmark in user studies of text-to-image and large language models. Despite the ability of these models to produce high-quality outputs, they all perform poorly on steerability. These results suggest that we need to focus on improving the steerability of generative models. We show such improvements are indeed possible: simple image-based steering mechanisms achieve more than 2x improvement on this benchmark.

Scaling laws for large language models (LLMs) predict model performance based on parameters like size and training data. However, differences in training configurations and data processing across model families lead to significant variations in benchmark performance, making it difficult for a single scaling law to generalize across all LLMs. On the other hand, training family-specific scaling laws requires training models of varying sizes for every family. In this work, we propose Skills Scaling Laws (SSLaws, pronounced as Sloth), a novel scaling law that leverages publicly available benchmark data and assumes LLM performance is driven by low-dimensional latent skills, such as reasoning and instruction following. These latent skills are influenced by computational resources like model size and training tokens, but with varying efficiencies across model families. Sloth exploits correlations across benchmarks to provide more accurate and interpretable predictions while alleviating the need to train multiple LLMs per family. We present both theoretical results on parameter identification and empirical evaluations on 12 prominent benchmarks, from Open LLM Leaderboard v1/v2, demonstrating that Sloth predicts LLM performance accurately and offers insights into scaling behaviors for complex downstream tasks, increased test-time compute, and compute-optimal scaling of skills.

Humans and traditional computer vision methods rely on a diverse set of monocular cues to infer 3D structure from a single image, such as shading, texture, silhouette, etc. While recent deep generative models have dramatically advanced single-image 3D generation, it remains unclear which image cues these methods actually exploit. We introduce Cue3D, the first comprehensive, model-agnostic framework for quantifying the influence of individual image cues in single-image 3D generation. Our unified benchmark evaluates seven state-of-the-art methods, spanning regression-based, multi-view, and native 3D generative paradigms. By systematically perturbing cues such as shading, texture, silhouette, perspective, edges, and local continuity, we measure their impact on 3D output quality. Our analysis reveals that shape meaningfulness, not texture, dictates generalization. Geometric cues, particularly shading, are crucial for 3D generation. We further identify over-reliance on provided silhouettes and diverse sensitivities to cues such as perspective and local continuity across model families. By dissecting these dependencies, Cue3D advances our understanding of how modern 3D networks leverage classical vision cues, and offers directions for developing more transparent, robust, and controllable single-image 3D generation models.

Graph Neural Networks operate through bottom-up message-passing, fundamentally differing from human visual perception, which intuitively captures global structures first. We investigate the underappreciated potential of vision models for graph understanding, finding they achieve performance comparable to GNNs on established benchmarks while exhibiting distinctly different learning patterns. These divergent behaviors, combined with limitations of existing benchmarks that conflate domain features with topological understanding, motivate our introduction of GraphAbstract. This benchmark evaluates models' ability to perceive global graph properties as humans do: recognizing organizational archetypes, detecting symmetry, sensing connectivity strength, and identifying critical elements. Our results reveal that vision models significantly outperform GNNs on tasks requiring holistic structural understanding and maintain generalizability across varying graph scales, while GNNs struggle with global pattern abstraction and degrade with increasing graph size. This work demonstrates that vision models possess remarkable yet underutilized capabilities for graph structural understanding, particularly for problems requiring global topological awareness and scale-invariant reasoning. These findings open new avenues to leverage this underappreciated potential for developing more effective graph foundation models for tasks dominated by holistic pattern recognition.

Large language models (LLMs) have demonstrated remarkable proficiency in mainstream academic disciplines such as mathematics, physics, and computer science. However, human knowledge encompasses over 200 specialized disciplines, far exceeding the scope of existing benchmarks. The capabilities of LLMs in many of these specialized fields-particularly in light industry, agriculture, and service-oriented disciplines-remain inadequately evaluated. To address this gap, we present SuperGPQA, a comprehensive benchmark that evaluates graduate-level knowledge and reasoning capabilities across 285 disciplines. Our benchmark employs a novel Human-LLM collaborative filtering mechanism to eliminate trivial or ambiguous questions through iterative refinement based on both LLM responses and expert feedback. Our experimental results reveal significant room for improvement in the performance of current state-of-the-art LLMs across diverse knowledge domains (e.g., the reasoning-focused model Gemini-2.5-Pro achieved the highest accuracy of 63.56% on SuperGPQA), highlighting the considerable gap between current model capabilities and artificial general intelligence. Additionally, we present comprehensive insights from our management of a large-scale annotation process, involving over 80 expert annotators and an interactive Human-LLM collaborative system, offering valuable methodological guidance for future research initiatives of comparable scope.

Differentially private (DP) synthetic data generation is a promising technique for utilizing private datasets that otherwise cannot be exposed for model training or other analytics. While much research literature has focused on generating private unstructured text and image data, in enterprise settings, structured data (e.g., tabular) is more common, often including natural language fields or components. Existing synthetic data evaluation techniques (e.g., FID) struggle to capture the structural properties and correlations of such datasets. In this work, we propose Struct-Bench, a framework and benchmark for evaluating synthetic datasets derived from structured datasets that contain natural language data. The Struct-Bench framework requires users to provide a representation of their dataset structure as a Context-Free Grammar (CFG). Our benchmark comprises 5 real-world and 2 synthetically generated datasets. We show that these datasets demonstrably present a great challenge even for state-of-the-art DP synthetic data generation methods. Struct-Bench provides reference implementations of different metrics and a leaderboard, offering a standardized platform to benchmark and investigate privacy-preserving synthetic data methods. We also present a case study showing how Struct-Bench improves the synthetic data quality of Private Evolution (PE) on structured data. The benchmark and the leaderboard have been publicly made available at https://struct-bench.github.io.

Recent advancements in video generation have significantly improved the ability to synthesize videos from text instructions. However, existing models still struggle with key challenges such as instruction misalignment, content hallucination, safety concerns, and generation bias. To address these limitations, we introduce MJ-BENCH-VIDEO, a large-scale video preference benchmark designed to evaluate video generation across five critical aspects: Alignment, Safety, Fineness, Coherence & Consistency, and Bias & Fairness. This benchmark further incorporates 28 fine-grained criteria to provide a comprehensive evaluation of video preference. Building upon this dataset, we propose MJ-VIDEO, a Mixture-of-Experts (MoE)-based video reward model designed to deliver fine-grained reward. MJ-VIDEO can dynamically select relevant experts to accurately judge the preference based on the input text-video pair. This architecture enables more precise and adaptable preference judgments. Through extensive benchmarking on MJ-BENCH-VIDEO, we analyze the limitations of existing video reward models and demonstrate the superior performance of MJ-VIDEO in video preference assessment, achieving 17.58% and 15.87% improvements in overall and fine-grained preference judgments, respectively. Additionally, MJ-VIDEO is able to improve the alignment performance in video generation via preference fine-tuning.

Vertical Federated Learning (VFL) has emerged as a critical privacy-preserving learning paradigm, enabling collaborative model training by leveraging distributed features across clients. However, due to privacy concerns, there are few publicly available real-world datasets for evaluating VFL methods, which poses significant challenges to related research. To bridge this gap, we propose MARS-VFL, a unified benchmark for realistic VFL evaluation. It integrates data from practical applications involving collaboration across different features, maintaining compatibility with the VFL setting. Based on this, we standardize the evaluation of VFL methods from the mainstream aspects of efficiency, robustness, and security. We conduct comprehensive experiments to assess different VFL approaches, providing references for unified evaluation. Furthermore, we are the first to unify the evaluation of robustness challenges in VFL and introduce a new method for addressing robustness challenges, establishing standard baselines for future research.

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we found one provider testing 27 private variants before making one model public at the second position on the leaderboard. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. The top two providers have individually received an estimated 19.2% and 20.4% of all data on the arena. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. With conservative estimates, we show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on ArenaHard, a test set from the arena distribution. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field.

Optimal Transport (OT) has emerged as a fundamental tool in machine learning for comparing probability distributions in a geometrically meaningful manner. However, a key limitation of classical OT is its requirement that the source and target distributions have equal total mass, limiting its use in real-world settings involving imbalanced data, noise, outliers, or structural inconsistencies. Partial Transport (PT) addresses this limitation by allowing only a fraction of the mass to be transported, offering greater flexibility and robustness. Nonetheless, similar to OT, PT remains computationally expensive, as it typically involves solving large-scale linear programs—especially in high-dimensional spaces. To alleviate this computational burden, several emerging works have introduced the Tree-Sliced Wasserstein (TSW) distance, which projects distributions onto tree-metric spaces where OT problems admit closed-form solutions. Building on this line of research, we propose a novel framework that extends the tree-sliced approach to the PT setting, introducing the Partial Tree-Sliced Wasserstein (PartialTSW) distance. Our method is based on the key observation that, within tree-metric space, the PT problem can be equivalently reformulated as a standard balanced OT problem between suitably modified measures. This reformulation enables efficient computation while preserving the adaptability and robustness of partial transport. Our method proves effective across challenging tasks such as outlier removal and addressing class imbalance in image-to-image translation. Our code is publicly available at [this https URL](https://github.com/thanhqt2002/PartialTSW).

Large Vision-Language Models (VLMs) have achieved remarkable success in understanding complex real-world scenarios and supporting data-driven decision-making processes. However, VLMs exhibit significant vulnerability against adversarial examples, either text or image, which can lead to various adversarial outcomes, e.g., jailbreaking, hijacking, and hallucination, etc. In this work, we empirically and theoretically demonstrate that VLMs are particularly susceptible to image-based adversarial examples, where imperceptible perturbations can precisely manipulate each output token. To this end, we propose a novel attack called Vision-language model Manipulation Attack (VMA), which integrates first-order and second-order momentum optimization techniques with a differentiable transformation mechanism to effectively optimize the adversarial perturbation. Notably, VMA can be a double-edged sword: it can be leveraged to implement various attacks, such as jailbreaking, hijacking, privacy breaches, Denial-of-Service, and the generation of sponge examples, etc, while simultaneously enabling the injection of watermarks for copyright protection. Extensive empirical evaluations substantiate the efficacy and generalizability of VMA across diverse scenarios and datasets. Code is available at https://github.com/Trustworthy-AI-Group/VMA.

In this paper, we study the finite-sample statistical rates of distributional temporal difference (TD) learning with linear function approximation. The aim of distributional TD learning is to estimate the return distribution of a discounted Markov decision process for a given policy $\pi$. Previous works on statistical analysis of distributional TD learning mainly focus on the tabular case. In contrast, we first consider the linear function approximation setting and derive sharp finite-sample rates. Our theoretical results demonstrate that the sample complexity of linear distributional TD learning matches that of classic linear TD learning. This implies that, with linear function approximation, learning the full distribution of the return from streaming data is no more difficult than learning its expectation (value function). To derive tight sample complexity bounds, we conduct a fine-grained analysis of the linear-categorical Bellman equation and employ the exponential stability arguments for products of random matrices. Our results provide new insights into the statistical efficiency of distributional reinforcement learning algorithms.