NeurIPS 2025 Papers

Large Language Models (LLMs) have shown remarkable progress across domains, yet their ability to perform inductive reasoning—inferring latent rules from sparse examples—remains limited. It is often assumed that chain-of-thought (CoT) prompting, as used in Large Reasoning Models (LRMs), enhances such reasoning. We investigate this assumption with creating four controlled, diagnostic game-based tasks—chess, Texas Hold’em, dice games, and blackjack—with hidden human-defined rules. We find that CoT reasoning can degrade inductive performance, with LRMs often underperforming their non-reasoning counterparts. To explain this, we present a theoretical framework that reveals how reasoning steps can amplify error through three failure modes: incorrect sub-task decomposition, incorrect sub-task solving, and incorrect final answer summarization. Based on our theoretical and empirical analysis, we introduce structured interventions that adapt CoT generation according to our identified failure types. These interventions improve inductive accuracy without retraining. Our findings suggest that effective (CoT) reasoning depends not only on taking more steps but also on ensuring those steps are well-structured.

Large Language Model (LLM) agents have demonstrated impressive capabilities for social interaction and are increasingly being deployed in situations where they might engage with both human and artificial agents. These interactions represent a critical frontier for LLM-based agents, yet existing evaluation methods fail to measure how well these capabilities generalize to novel social situations. In this paper, we introduce a method for evaluating the ability of LLM-based agents to cooperate in zero-shot, mixed-motive environments using Concordia, a natural language multi-agent simulation environment. Our method measures general cooperative intelligence by testing an agent's ability to identify and exploit opportunities for mutual gain across diverse partners and contexts. We present empirical results from the NeurIPS 2024 Concordia Contest, where agents were evaluated on their ability to achieve mutual gains across a suite of diverse scenarios ranging from negotiation to collective action problems. Our findings reveal significant gaps between current agent capabilities and the robust generalization required for reliable cooperation, particularly in scenarios demanding persuasion and norm enforcement.

Spatial reasoning is a key capability in the field of artificial intelligence, especially crucial in areas such as robotics, computer vision, and natural language understanding. However, evaluating the ability of multimodal large language models (MLLMs) in complex spatial reasoning still faces challenges, particularly in scenarios requiring multi-step reasoning and precise mathematical constraints. This paper introduces ORIGAMISPACE, a new dataset and benchmark designed to evaluate the multi-step spatial reasoning ability and the capacity to handle mathematical constraints of MLLMs through origami tasks. The dataset contains 350 data instances, each comprising a strictly formatted crease pattern (CP diagram), the Compiled Flat Pattern, the complete Folding Process, and the final Folded Shape Image. We propose four evaluation tasks: Pattern Prediction, Multi-step Spatial Reasoning, Spatial Relationship Prediction, and End-to-End CP Code Generation. For the CP code generation task, we design an interactive environment and explore the possibility of using reinforcement learning methods to train MLLMs. Through experiments on existing MLLMs, we initially reveal the strengths and weaknesses of these models in handling complex spatial reasoning tasks.

Evolutionary multi-view classification (EMVC) methods have gained wide recognition due to their adaptive mechanisms. Fitness evaluation (FE), which aims to calculate the classification performance of each individual in the population and provide reliable performance ranking for subsequent operations, is a core step in such methods. Its accuracy directly determines the correctness of the evolutionary direction. However, when FE fails to correctly reflect the superiority-inferiority relationship among individuals, it will lead to confusion in individual performance ranking, which in turn misleads the evolutionary direction and results in trapping into local optima. This paper is the first to identify the aforementioned issue in the field of EMVC and call it as fitness evaluation bias (FEB). FEB may be caused by a variety of factors, and this paper approaches the issue from the perspective of view information content: existing methods generally adopt joint training strategies, which restrict the exploration of key information in views with low information content. This makes it difficult for multi-view model (MVM) to achieve optimal performance during convergence, which in turn leads to FE failing to accurately reflect individual performance rankings and ultimately triggering FEB. To address this issue, we propose an evolutionary multi-view classification via eliminating individual fitness bias (EFB-EMVC) method, which alleviates the FEB issue by introducing evolutionary navigators for each MVM, thereby providing more accurate individual ranking. Experimental results fully verify the effectiveness of the proposed method in alleviating the FEB problem, and the EMVC method equipped with this strategy exhibits more superior performance compared with the original EMVC method. (The code is available at https://github.com/LiShuailzn/Neurips-2025-EFB-EMVC)

Detecting AI-involved text is essential for combating misinformation, plagiarism, and academic misconduct. However, AI text generation includes diverse collaborative processes (AI-written text edited by humans, human-written text edited by AI, and AI-generated text refined by other AI), where various or even new LLMs could be involved. Texts generated through these varied processes exhibit complex characteristics, presenting significant challenges for detection. Current methods model these processes rather crudely, primarily employing binary classification (purely human vs. AI-involved) or multi-classification (treating human-AI collaboration as a new class). We observe that representations of texts generated through different processes exhibit inherent clustering relationships. Therefore, we propose DETree, a novel approach that models the relationships among different processes as a Hierarchical Affinity Tree structure, and introduces a specialized loss function that aligns text representations with this tree. To facilitate this learning, we developed RealBench, a comprehensive benchmark dataset that automatically incorporates a wide spectrum of hybrid texts produced through various human-AI collaboration processes. Our method improves performance in hybrid text detection tasks and significantly enhances robustness and generalization in out-of-distribution scenarios, particularly in few-shot learning conditions, further demonstrating the promise of training-based approaches in OOD settings. Our code and dataset are available at https://github.com/heyongxin233/DETree.

Reinforcement learning (RL) has become the dominant paradigm for improving the performance of language models on complex reasoning tasks. Despite the substantial empirical gains demonstrated by RL-based training methods like GRPO, a granular understanding of why and how RL enhances performance is still lacking. To bridge this gap, we introduce SPARKLE, a fine-grained analytic framework to dissect the effects of RL across three key dimensions: (1) plan following and execution, (2) knowledge integration, and (3) chain of subproblems. Using this framework, we gain insights beyond mere accuracy. For instance, providing models with explicit human-crafted, step-by-step plans can surprisingly degrade performance on the most challenging benchmarks, yet RL-tuned models exhibit greater robustness, experiencing markedly smaller performance drops than base or SFT models. This suggests that RL may not primarily enhance the execution of external plans but rather empower models to formulate and follow internal strategies better suited to their reasoning processes. Conversely, we observe that RL enhances models' ability to integrate provided knowledge into their reasoning process, yielding consistent gains across diverse tasks. Finally, we study whether difficult problems---those yielding no RL signals and mixed-quality reasoning traces---can still be effectively used for training. We introduce SparkleRL-PSS, a multi-stage RL pipeline that reuses hard problems with partial step scaffolding, guiding exploration effectively without additional data generation. Together, our findings provide a principled foundation for understanding how RL shapes model behavior, offering practical insights for building more adaptive, data-efficient, and interpretable RL pipelines for reasoning tasks. Our code, data, and checkpoints are available at: https://sparkle-reasoning.github.io/.

LLM-based agents have been widely applied as personal assistants, capable of memorizing information from user messages and responding to personal queries. However, there still lacks an objective and automatic evaluation on their memory capability, largely due to the challenges in constructing reliable questions and answers (QAs) according to user messages. In this paper, we propose MemSim, a Bayesian simulator designed to automatically construct reliable QAs from generated user messages, simultaneously keeping their diversity and scalability. Specifically, we introduce the Bayesian Relation Network (BRNet) and a causal generation mechanism to mitigate the impact of LLM hallucinations on factual information, facilitating the automatic creation of an evaluation dataset. Based on MemSim, we generate a dataset in the daily-life scenario, named MemDaily, and conduct extensive experiments to assess the effectiveness of our approach. We also provide a benchmark for evaluating different memory mechanisms in LLM-based agents with the MemDaily dataset.

Label Distribution Learning (LDL) has emerged as a powerful framework for estimating complete conditional label distributions, providing crucial reliability for risk-sensitive decision-making tasks. While existing LDL algorithms exhibit competent performance under the conventional LDL performance evaluation methods, two key limitations remain: (1) current algorithms systematically underperform on the samples with low-entropy label distributions, which can be particularly valuable for decision making, and (2) the conventional performance evaluation methods are inherently biased due to the numerical imbalance of samples. In this paper, through empirical and theoretical analyses, we find that excessive cohesion between anchor vectors contributes significantly to the observed entropy bias phenomenon in LDL algorithms. Accordingly, we propose an inter-anchor angular regularization term that mitigates cohesion among anchor vectors by penalizing over-small angles. Besides, to alleviate the numerical imbalance of high-entropy samples in test set, we propose an entropy-calibrated aggregation strategy that obtains the overall model performance by evaluating performance on the low-entropy and high-entropy subsets of the overall test set separately. Finally, we conduct extensive experiments on various real-world datasets to demonstrate the effectiveness of our proposal.

Recent advancements in large language models (LLMs) underscore the need for more comprehensive evaluation methods to accurately assess their reasoning capabilities. Existing benchmarks are often domain-specific and thus cannot fully capture an LLM’s general reasoning potential. To address this limitation, we introduce the **Knowledge Orthogonal Reasoning Gymnasium (KORGym)**, a dynamic evaluation platform inspired by KOR-Bench and Gymnasium. KORGym offers over fifty games in either textual or visual formats and supports interactive, multi-turn assessments with reinforcement learning scenarios. Using KORGym, we conduct extensive experiments on 19 LLMs and 8 VLMs, revealing consistent reasoning patterns within model families and demonstrating the superior performance of closed-source models. Further analysis examines the effects of modality, reasoning strategies, reinforcement learning techniques, and response length on model performance. We expect KORGym to become a valuable resource for advancing LLM reasoning research and developing evaluation methodologies suited to complex, interactive environments.

Coreset selection methods have shown promise in reducing the training data size while maintaining model performance for data-efficient machine learning. However, many large real-world datasets suffer from unknown spurious correlations and hidden biases. Therefore, it is crucial to understand how such biases would affect downstream tasks via the selected coresets. In this work, we conduct the first comprehensive analysis of the implications of data selection on the bias levels of the selected coresets and the robustness of downstream models trained on them. We use an extensive experimental setting spanning ten different spurious correlations benchmarks, five score metrics to characterize sample importance/ difficulty, and five data selection policies across a broad range of coreset sizes to identify important patterns and derive insights. Thereby, we unravel a series of nontrivial nuances in well-known interactions between sample difficulty and bias alignment, as well as dataset bias and resultant model robustness. For example, we show that embedding-based sample characterizations run a comparatively lower risk of inadvertently exacerbating bias when used for selecting coresets compared to characterizations based on learning dynamics. Our analysis also reveals that lower bias levels achieved by coresets of difficult samples do not reliably guarantee downstream robustness. Most importantly, we show that special considerations need to be made when the coreset size is very small, since there is a unique risk of highly prototypical coresets reaching high average performance while obscuring their low group-robustness.

We consider off-policy evaluation (OPE) in Partially Observable Markov Decision Processes (POMDPs) with unobserved confounding. Recent advances have introduced bridge-function to circumvent unmeasured confounding and develop estimators for the policy value, yet the statistical error bounds of them related to the length of horizon $T$ and the size of the state-action space $|\mathcal{O}||\mathcal{A}|$ remain largely unexplored. In this paper, we systematically investigate the finite-sample error bounds of OPE estimators in finite-horizon tabular confounded POMDPs. Specifically, we show that under certain rank conditions, the estimation error for policy value can achieve a rate of $\mathcal{O}(T^{1.5}/\sqrt{n})$, excluding the cardinality of the observation space $|\mathcal{O}|$ and the action space $|\mathcal{A}|$. With an additional mild condition on the concentrability coefficients in confounded POMDPs, the rate of estimation error can be improved to $\mathcal{O}(T/\sqrt{n})$. We also show that for a fully history-dependent policy, the estimation error scales as $\mathcal{O}\big(T/\sqrt{n}(|\mathcal{O}| |\mathcal{A}|)^{\frac{T}{2}}\big)$, highlighting the exponential error dependence introduced by history-based proxies to infer hidden states. Furthermore, when the target policy is memoryless policy, the error bound improves to $\mathcal{O}\big(T/\sqrt{n}\sqrt{|\mathcal{O}| |\mathcal{A}|}\big)$, which matches the optimal rate known for tabular MDPs. To the best of our knowledge, this is the first work to provide a comprehensive finite-sample analysis of OPE in confounded POMDPs.

We study the task of agnostically learning general (as opposed to homogeneous) ReLUs under the Gaussian distribution with respect to the squared loss. In the passive learning setting, recent work gave a computationally efficient algorithm that uses $poly(d,1/\epsilon)$ labeled examples and outputs a hypothesis with error $O(opt)+\epsilon$, where $opt$ is the squared loss of the best fit ReLU. Here we focus on the interactive setting, where the learner has some form of query access to the labels of unlabeled examples. Our main result is the first computationally efficient learner that uses $d polylog(1/\epsilon)+\tilde{O}(\min\{1/p, 1/\epsilon\})$ black-box label queries, where $p$ is the bias of the target function, and achieves error $O(opt)+\epsilon$. We complement our algorithmic result by showing that its query complexity bound is qualitatively near-optimal, even ignoring computational constraints. Finally, we establish that query access is essentially necessary to improve on the label complexity of passive learning. Specifically, for pool-based active learning, any active learner requires $\tilde{\Omega}(d/\epsilon)$ labels, unless it draws a super-polynomial number of unlabeled examples.

Large Language Models (LLMs) often struggle to maintain their original performance when faced with semantically coherent but task-irrelevant contextual information. Although prior studies have explored this issue using fixed-template or retrieval-based distractions, such static methods show limited effectiveness against contemporary models. To address this problem, we propose a dynamic distraction generation framework based on tree search, where the generation process is guided by model behavior. Without modifying the original question or answer, the method efficiently produces challenging adaptive distractions across multiple datasets, enabling systematic stress testing of LLMs’ contextual robustness. Experiments on four benchmarks demonstrate that the generated distractions lead to an average performance drop of over 45\% for mainstream models. Further comparisons of mitigation strategies show that prompt-based optimization methods yield limited gains, whereas post-training approaches (e.g., DPO) significantly enhance the model's contextual robustness. The results indicate that these issues do not stem from knowledge deficits in LLMs, but from a fundamental inability to maintain consistent reasoning under contextual distraction, posing a major challenge to the reliability of LLMs in real-world applications.

Developing large language models is expensive and often involves making decisions with small experiments, typically by evaluating on large, multi-task evaluation suites. In this work, we analyze specific properties which make a benchmark more reliable and useful for such decisions, and interventions to design higher-quality evaluation benchmarks. We introduce two key metrics that show differences in current benchmarks: signal, a benchmark’s ability to separate better models from worse models, and noise, a benchmark’s sensitivity to random variability between training steps. We demonstrate that benchmarks with a better signal-to-noise ratio are more reliable when making decisions at small scale, and those with less noise have lower scaling law prediction error. These results suggest that improving signal or noise will lead to more useful benchmarks, so we introduce four interventions designed to directly affect signal or noise. For example, we propose that switching to a metric that has better signal and noise (e.g., perplexity rather than accuracy) leads to better reliability and scaling law error. We also find that filtering noisy benchmarks such that they have better signal-to-noise ratio leads to more reliable evaluations. We also find that averaging the output of a model's checkpoints to reduce noise leads to consistent improvements. We conclude by recommending that those creating new benchmarks, or selecting which existing benchmarks to use, aim for high signal and low noise. We use 30 benchmarks for these experiments, and 465 open-weight language models from 60M to 32B parameters, resulting in a new, publicly available dataset of 50K evaluation benchmark results, totaling 200M instances.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available in the supplementary materials.

Recent advances in deep learning have driven rapid progress in time series forecasting, yet many state-of-the-art models continue to struggle with robust performance in real-world applications, even when they achieve strong results on standard benchmark datasets. This persistent gap can be attributed to the black-box nature of deep learning architectures and the inherent limitations of current evaluation frameworks, which frequently lack the capacity to provide clear, quantitative insights into the specific strengths and weaknesses of different models, thereby complicating the selection of appropriate models for particular forecasting scenarios. To address these issues, we propose a synthetic data-driven evaluation paradigm, SynTSBench, that systematically assesses fundamental modeling capabilities of time series forecasting models through programmable feature configuration. Our framework isolates confounding factors and establishes an interpretable evaluation system with three core analytical dimensions: (1) temporal feature decomposition and capability mapping, which enables systematic evaluation of model capacities to learn specific pattern types; (2) robustness analysis under data irregularities, which quantifies noise tolerance thresholds and anomaly recovery capabilities; and (3) theoretical optimum benchmarking, which establishes performance boundaries for each pattern type—enabling direct comparison between model predictions and mathematical optima.Our experiments show that current deep learning models do not universally approach optimal baselines across all types of temporal features.

Large Language Models (LLMs) have demonstrated remarkable instruction-following capabilities across various applications. However, their performance in multilingual settings lacks systematic investigation, with existing evaluations lacking fine-grained constraint analysis across diverse linguistic contexts. We introduce **XIFBench**, a comprehensive constraint-based benchmark for evaluating multilingual instruction-following abilities of LLMs, comprising 558 instructions with 0-5 additional constraints across five categories (*Content*, *Style*, *Situation*, *Format*, and *Numerical*) in six languages spanning different resource levels. To support reliable and consistent cross-lingual evaluation, we implement three methodological innovations: cultural accessibility annotation, constraint-level translation validation, and requirement-based evaluation using English requirements as semantic anchors across languages. Extensive experiments with various LLMs not only quantify performance disparities across resource levels but also provide detailed insights into how language resources, constraint categories, instruction complexity, and cultural specificity influence multilingual instruction-following. Our code and data are available at https://github.com/zhenyuli801/XIFBench.

Training effective AI agents for multi-turn interactions requires high-quality data that captures realistic human-agent dynamics, yet such data is scarce and expensive to collect manually. We introduce APIGen-MT, a two-phase framework that generates verifiable and diverse multi-turn agent data. In the first phase, our agentic pipeline produces detailed task blueprints with ground-truth actions, leveraging a committee of LLM reviewers and iterative feedback loops. These blueprints are then transformed into complete interaction trajectories through simulated human-agent interplay. We train a family of models---the xLAM-2-fc-r series with sizes ranging from 1B to 70B parameters. Our models outperform frontier models such as GPT-4o and Claude 3.5 on $\tau$-bench and BFCL benchmarks, with the smaller models surpassing their larger counterparts, particularly in multi-turn settings, while maintaining superior consistency across multiple trials. Comprehensive experiments demonstrate that our verified blueprint-to-details approach yields high-quality training data, enabling the development of more reliable, efficient, and capable agents. We open-source both the synthetic data collected and the trained xLAM-2-fc-r models to advance research in AI agents. Dataset: https://huggingface.co/datasets/Salesforce/APIGen-MT-5k & Models: https://huggingface.co/collections/Salesforce/xlam-2-67ef5be12949d8dcdae354c4

This paper studies the problem of class-incremental learning (CIL), a core setting within continual learning where a model learns a sequence of tasks, each containing a distinct set of classes. Traditional CIL methods, which do not leverage pre-trained models (PTMs), suffer from catastrophic forgetting (CF) due to the need to incrementally learn both feature representations and the classifier. The integration of PTMs into CIL has recently led to efficient approaches that treat the PTM as a fixed feature extractor combined with analytic classifiers, achieving state-of-the-art performance. However, they still face a major limitation: the inability to continually adapt feature representations to best suit the CIL tasks, leading to suboptimal performance. To address this, we propose AnaCP (Analytic Contrastive Projection), a novel method that preserves the efficiency of analytic classifiers while enabling incremental feature adaptation without gradient-based training, thereby eliminating the CF caused by gradient updates. Our experiments show that AnaCP not only outperforms existing baselines but also achieves the accuracy level of joint training, which is regarded as the upper bound of CIL.