ICLR 2026 Papers

Offline reinforcement learning (RL) holds the promise of training effective policies from static datasets without the need for costly online interactions. However, offline RL faces key limitations, most notably the challenge of generalizing to unseen or infrequently encountered state-action pairs. When a value function is learned from limited data in sparse-reward environments, it can become overly optimistic about parts of the space that are poorly represented, leading to unreliable value estimates and degraded policy quality. To address these challenges, we introduce a novel approach based on contrastive preference learning that bypasses direct value function estimation. Our method trains policies by contrasting successful demonstrations with failure behaviors present in the dataset, as well as synthetic behaviors generated outside the support of the dataset distribution. This contrastive formulation mitigates overestimation bias and improves robustness in offline learning. Empirical results on challenging sparse-reward offline RL benchmarks show that our method substantially outperforms existing state-of-the-art baselines in both learning efficiency and final performance.

Single-cell RNA sequencing (scRNA-seq) enables the study of cellular heterogeneity. Yet, clustering accuracy, and with it downstream analyses based on cell labels, remain challenging due to measurement noise and biological variability. In standard latent spaces (e.g., obtained through PCA), data from different cell types can be projected close together, making accurate clustering difficult. We introduce a latent plug-and-play diffusion framework that separates the observation and denoising space. This separation is operationalized through a novel Gibbs sampling procedure: the learned diffusion prior is applied in a low-dimensional latent space to perform denoising, while to steer this process, noise is reintroduced into the original high-dimensional observation space. This unique ``input-space steering'' ensures the denoising trajectory remains faithful to the original data structure. Our approach offers three key advantages: (1) adaptive noise handling via a tunable balance between prior and observed data; (2) uncertainty quantification through principled uncertainty estimates for downstream analysis; and (3) generalizable denoising by leveraging clean reference data to denoise noisier datasets, and via averaging, improve quality beyond the training set. We evaluate robustness on both synthetic and real single-cell genomics data. Our method improves clustering accuracy on synthetic data across varied noise levels and dataset shifts. On real-world single-cell data, our method demonstrates improved biological coherence in the resulting cell clusters, with cluster boundaries that better align with known cell type markers and developmental trajectories.

Models like OpenAI-o3 pioneer visual grounded reasoning by dynamically ref- erencing visual regions, just like human “thinking with images”. However, no benchmark exists to evaluate these capabilities holistically. To bridge this gap, we propose TreeBench (Traceable Evidence Evaluation Benchmark), a diagnostic benchmark built on three principles: (1) focused visual perception of subtle targets in complex scenes, (2) traceable evidence via bounding box evaluation, and (3) second-order reasoning to test object interactions and spatial hierarchies beyond simple object localization. Prioritizing images with dense objects, we initially sample 1K high-quality images from SA-1B, and incorporate eight LMM experts to manually annotate questions, candidate options, and answers for each image. After three stages of quality control, TreeBench consists of 405 challenging vi- sual question-answering pairs, even the most advanced models struggle with this benchmark, where none of them reach 60% accuracy, e.g., OpenAI-o3 scores only 54.87. Furthermore, we introduce TreeVGR (Traceable Evidence Enhanced Visual Grounded Reasoning), a training paradigm to supervise localization and reasoning jointly with reinforcement learning, enabling accurate localizations and explainable reasoning pathways. Initialized from Qwen2.5-VL-7B, it improves V* Bench (+16.8), MME-RealWorld (+12.6), and TreeBench (+13.4), proving traceability is key to advancing vision-grounded reasoning. The code and data will be released.

Given a training dataset, the goal of dataset distillation is to derive a synthetic dataset such that models trained on the latter perform as well as those trained on the training dataset. In this work, we develop and analyze an efficient dataset distillation algorithm for supervised learning, specifically regression in $\mathbb{R}^d$, based on matching the losses on the training and synthetic datasets with respect to a fixed set of randomly sampled regressors without any model training. Our first key contribution is a novel performance guarantee proving that our algorithm needs only $\tilde{O}(d^2)$ sampled regressors to derive a synthetic dataset on which the MSE loss of any bounded linear model is approximately the same as its MSE loss on the given training data. In particular, the model optimized on the synthetic data has close to minimum loss on the training data, thus performing nearly as well as the model optimized on the latter. Complementing this, we also prove a matching lower bound of $\Omega(d^2)$ for the number of sampled regressors showing the tightness of our analysis. Our second contribution is to extend our algorithm to offline RL dataset distillation by matching the Bellman loss, unlike previous works which used a behavioral cloning objective. This is the first such method which leverages both, the rewards and the next state information, available in offline RL datasets, without any policy model optimization. We show similar guarantees: our algorithm generates a synthetic dataset whose Bellman loss with respect to any linear action-value predictor is close to the latter’s Bellman loss on the offline RL training dataset. Therefore, a policy associated with an action-value predictor optimized on the synthetic dataset performs nearly as well as that derived from the one optimized on the training data. We conduct extensive experiments to validate our theoretical guarantees and observe performance gains on real-world RL environments with offline training datasets and supervised regression datasets.

We introduce a framework for robust uncertainty quantification in situations where labeled training data are corrupted, through noisy or missing labels. We build on conformal prediction, a statistical tool for generating prediction sets that cover the test label with a pre-specified probability. The validity of conformal prediction, however, holds under the i.i.d assumption, which does not hold in our setting due to the corruptions in the data. To account for this distribution shift, the privileged conformal prediction (PCP) method proposed leveraging privileged information (PI)---additional features available only during training---to re-weight the data distribution, yielding valid prediction sets under the assumption that the weights are accurate. In this work, we analyze the robustness of PCP to inaccuracies in the weights. Our analysis indicates that PCP can still yield valid uncertainty estimates even when the weights are poorly estimated. Furthermore, we introduce uncertain imputation (UI), a new conformal method that does not rely on weight estimation. Instead, we impute corrupted labels in a way that preserves their uncertainty. Our approach is supported by theoretical guarantees and validated empirically on both synthetic and real benchmarks. Finally, we show that these techniques can be integrated into a triply robust framework, ensuring statistically valid predictions as long as at least one underlying method is valid.

Massive training datasets are fueling the astounding progress in molecular machine learning. Since these datasets are typically generated with computational chemistry codes which do not randomize pose, the resulting geometries are usually not randomly oriented. While cheminformaticians are well aware of this fact, it can be a real pitfall for machine learners entering the burgeoning field of molecular machine learning. We demonstrate that molecular poses in the popular datasets QM9, QMugs and OMol25 are indeed biased. While the fact can easily be overseen by visual inspection alone, we show that a simple classifier can separate original data samples from randomly rotated ones with high accuracy. Second, we validate empirically that neural networks can and do exploit the orientedness in these datasets by successfully training a model on chemical property regression using the molecular orientation as _sole_ input. Third, we present visualizations of all molecular orientations and confirm that chemically similar molecules tend to have similar canonical poses. In summary, we recall and document orientational bias in the prevalent datasets that machine learners should be aware of.

Knowledge editing and machine unlearning are two popular approaches for large language models (LLMs) to stay up-to-date. However, the knowledge updating mechanism of LLMs remains largely unexplored due to insufficient, isolated, and small-scale evaluation. For instance, are LLMs similar to humans in modifying certain knowledge? What differs editing and unlearning as training data increases? This paper proposes KnowledgeSmith, a unified framework to systematically understand the updating mechanism of LLMs. We first cast editing and unlearning as instances of one constrained optimization problem. Then, we propose an automatic dataset generator that provides structured interventions across multiple graph levels and data scales, enabling controlled studies of how different modification strategies propagate through model knowledge. Extensive experiments demonstrate nuanced insights over knowledge propagation, plasticity scaling, consistency, and robustness. For instance, our results show that LLMs do not exhibit similar updating as humans for different levels of knowledge, and there exists consistency-capacity trade-off. We hope our findings can offer suggestions to the design of more reliable and scalable strategies.

Automatically detecting machine-generated text (MGT) is critical to maintaining the knowledge integrity of user-generated content (UGC) platforms such as Wikipedia. Existing detection benchmarks primarily focus on \textit{generic} text generation tasks (e.g., ``Write an article about machine learning.''). However, editors frequently employ LLMs for specific writing tasks (e.g., summarisation). These \textit{task-specific} MGT instances tend to resemble human-written text more closely due to their constrained task formulation and contextual conditioning. In this work, we show that a range of MGT detectors struggle to identify task-specific MGT reflecting real-world editing on Wikipedia. We introduce \textsc{TSM-Bench}, a multilingual, multi-generator, and multi-task benchmark for evaluating MGT detectors on common, real-world Wikipedia editing tasks. Our findings demonstrate that (\textit{i}) average detection accuracy drops by 10--40\% compared to prior benchmarks, and (\textit{ii}) a generalisation asymmetry exists: fine-tuning on task-specific data enables generalisation to generic data---even across domains---but not vice versa. We demonstrate that models fine-tuned exclusively on generic MGT overfit to superficial artefacts of machine generation. Our results suggest that, in contrast to prior benchmarks, most detectors remain unreliable for automated detection in real-world contexts such as UGC platforms. \textsc{TSM-Bench} therefore provides a crucial foundation for developing and evaluating future models.

We present a dynamical-systems based model for resting-state functional magnetic resonance imaging (rs-fMRI), trained on a dataset of roughly $40$K rs-fMRI sequences covering a wide variety of public and available-by-permission datasets. While most existing proposals use transformer backbones, we utilize multi-resolution temporal modeling of the dynamics across parcellated brain regions. We show that MnemoDyn is compute efficient and generalizes very well across diverse populations and scanning protocols. When benchmarked against current state-of-the-art transformer-based approaches, MnemoDyn consistently delivers superior reconstruction quality. Overall, we find that with such large-scale pre-training on (non-proprietary) rs-fMRI datasets, we get a highly performant model for various downstream tasks. Our results also provide evidence of the efficacy of the model on small sample size studies which has implications for neuroimaging studies at large where resting state fMRI is a commonly acquired imaging modality.

Deep learning models have driven significant progress in predicting protein function and interactions at the protein level. While these advancements have been invaluable for many biological applications such as enzyme engineering and function annotation, a more detailed perspective is essential for understanding protein functional mechanisms and evaluating the biological knowledge captured by models. This study introduces VenusX, the first benchmark designed to assess protein representation learning with a focus on fine-grained intra-protein functional understanding. VenusX comprises three major task categories across six types of annotations, including residue-level binary classification, fragment-level multi-class classification, and pairwise functional similarity scoring for identifying critical active sites, binding sites, conserved sites, motifs, domains, and epitopes. The benchmark features over 878,000 samples curated from major open-source databases such as InterPro, BioLiP, and SAbDab. By providing mixed-family and cross-family splits at three sequence identity thresholds, our benchmark enables a comprehensive assessment of model performance on both in-distribution and out-of-distribution scenarios. For baseline evaluation, we assess a diverse set of popular and open-source models, including pre-trained protein language models, sequence-structure hybrids, structure-based methods, and alignment-based techniques. Their performance is reported across all benchmark datasets and evaluation settings using multiple metrics, offering a thorough comparison and a strong foundation for future research. Our code (https://anonymous.4open.science/r/VenusX-4674), data (https://huggingface.co/collections/anonymous-researcher-123/venusx-68cc5163ade527b0974bab29), and a leaderboard (https://anonymous-researcher-816.github.io/) are provided as open-source resources.

We introduce SealQA, a challenge benchmark for evaluating SEarch-Augmented Language models on fact-seeking questions where web search yields conflicting, noisy, or unhelpful results. SealQA comes in three flavors: (1) SEAL-0 (main) and (2) SEAL-HARD, both of which assess factual accuracy and reasoning capabilities, where SEAL-0 targets the most challenging questions that frontier non-reasoning models (e.g., GPT-4.1) answer with near-zero accuracy; and (3) LongSeal, which extends SealQA to test long-context, multi-document reasoning in "needle-in-a-haystack" settings. Our evaluation reveals critical limitations in current models. Even frontier reasoning models face significant challenges across SealQA flavors. On SEAL-0, GPT-5 with tools achieves only 43.2% accuracy at its best reasoning effort. We also find that even advanced reasoning models (e.g., DeepSeek-R1) can be vulnerable to noisy search results. Notably, increasing test-time compute does not yield reliable gains across GPT-5 and the o-series of models, with performance often plateauing or even declining early. Finally, while current models are less affected by the "lost-in-the-middle" issue, they still fail to reliably identify relevant documents in LongSeal when faced with numerous distractors. To facilitate future work, we release SealQA at https://anonymous.4open.science/r/SealQA.

Modern neural networks are expected to simultaneously satisfy a host of desirable properties: accurate fitting to training data, generalization to unseen inputs, parameter and computational efficiency, and robustness to adversarial perturbations. While compressibility and robustness have each been studied extensively, a unified understanding of their interaction still remains elusive. In this work, we develop a principled framework to analyze how different forms of structured compressibility - such as neuron-level sparsity and spectral compressibility - affect adversarial robustness. We show that these forms of compression can induce a small number of highly sensitive directions in the representation space, which adversaries can exploit to construct effective perturbations. Our analysis yields a robustness bound that reveals how neuron and spectral compressibility impact $\ell_\infty$ and $\ell_2$ robustness via their effects on the learned representations. Crucially, the vulnerabilities we identify arise irrespective of how compressibility is achieved - whether via regularization, architectural bias, or implicit learning dynamics. Through empirical evaluations across synthetic and realistic tasks, we confirm our theoretical predictions, and further demonstrate that these vulnerabilities persist under adversarial training and transfer learning, and contribute to the emergence of universal adversarial perturbations. Our findings show a fundamental tension between structured compressibility and robustness and highlight new pathways for designing models that are both efficient and safe.

Multimodal Large Language Models (MLLMs) have achieved significant advances in integrating visual and linguistic information, yet their ability to reason about complex and real-world scenarios remains limited. Existing benchmarks are usually constructed in a task-oriented manner, without a guarantee that different task samples come from the same data distribution. Therefore, they often fall short in evaluating the synergistic effects of lower-level perceptual capabilities on higher-order reasoning. To lift this limitation, we contribute Lens, a multi-level evaluation benchmark of multimodal reasoning with with 3.4K contemporary images and 60K+ human-authored questions covering eight tasks and 12 daily scenarios, forming three progressive task tiers, i.e., perception, understanding, and reasoning. One feature is that each image is equipped with rich annotations for all tasks. Thus, this data set intrinsically supports evaluating MLLMs to handle image-invariable prompts, from basic perception to compositional reasoning. In addition, our images have been collected manually from social media, with $53$% published after Jan. 2025. We evaluate 15+ frontier MLLMs such as Qwen2.5-VL, InternVL3, GPT-4o and two reasoning models QVQ-Max and Kimi-VL. Most models were released in 2025, and none of them achieve an accuracy beyond $60$% in the reasoning tasks. Furthermore, we propose the Self-Driven Multi-Expert Collaborative Framework (SMEC), a framework designed for MLLMs that simulates a panel of experts discussing and exchanging viewpoints via self-generated role-specific prompts. The experimental results confirm the existence of synergistic effects in a hierarchical task structure, where low-level tasks facilitate the reasoning of MLLMs on more complex, high-level tasks. Statistical analysis and ablation studies further demonstrate the comprehensiveness of our dataset and the superiority of our methodology.

Semantic caches return cached responses for semantically similar prompts to reduce LLM inference latency and cost. They embed cached prompts and store them alongside their response in a vector database. Embedding similarity metrics assign a numerical score to quantify the similarity between a request and its nearest neighbor prompt from the cache. Existing systems use the same static similarity threshold across all requests to determine whether two prompts can share similar responses. However, we observe that static thresholds do not give formal correctness guarantees, result in unexpected error rates, and lead to suboptimal cache hit rates. This paper proposes vCache, the first verified semantic cache with user-defined error rate guarantees for predictable performance. It employs an online learning algorithm to estimate an optimal threshold for each cached prompt, enabling reliable cache responses without additional training. Our experiments show that vCache consistently meets the specified error bounds while outperforming state-of-the-art static-threshold and fine-tuned embedding baselines with up to 12.5$\times$ higher cache hit and 26$\times$ lower error rates. We release the vCache implementation and four benchmarks to support future research.

We propose a statistical benchmark for diffusion posterior sampling (DPS) algorithms in linear inverse problems. Our test signals are discretized Lévy processes whose posteriors admit efficient Gibbs methods. These Gibbs methods provide gold-standard posterior samples for direct, distribution-level comparisons with DPS algorithms. They can also sample the denoising posteriors in the reverse diffusion, which enables the arbitrary-precision Monte Carlo estimation of various objects that may be needed in the DPS algorithms, such as the expectation or the covariance of the denoising posteriors. In turn, this can be used to isolate algorithmic errors from the errors due to learned components. We instantiate the benchmark with the minimum-mean-squared-error optimality gap and posterior-coverage tests and evaluate popular algorithms on the inverse problems of denoising, deconvolution, imputation, and reconstruction from partial Fourier measurements. We release the benchmark code at https://github.com/emblem-saying/dps-benchmark and invite the community to contribute and report results.

The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and language, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

The recent hardware-accelerated microscaling 4-bit floating-point formats such as MXFP4 and NVFP4, supported on NVIDIA and AMD GPUs, promise to revolutionize large language model (LLM) inference. Yet, their practical benefits remain unproven. We present the first comprehensive study of MXFP4 and NVFP4 for post-training quantization, revealing gaps between their promise and real-world performance. Our analysis shows that state-of-the-art methods struggle with FP4, due to two key issues: (1) NVFP4's small group size \emph{provably} neutralizes traditional outlier mitigation techniques; (2) MXFP4's power-of-two scale quantization severely degrades accuracy due to high induced error. To bridge this gap, we introduce Micro-Rotated-GPTQ (MR-GPTQ), a variant of the classic GPTQ quantization algorithm that tailors the quantization process to FP4's unique properties, by using block-wise Hadamard transforms and format-specific optimizations. We support our proposal with a set of high-performance GPU kernels that enable the MR-GPTQ format with negligible overhead, by rotation fusion into the weights, and fast online computation of the activations. This leads to speedups vs. FP16 of up to 3.6x layer-wise, and 2.2x end-to-end on NVIDIA B200, and of 6x layer-wise and 4x end-to-end on RTX5090. Our extensive empirical evaluation demonstrates that MR-GPTQ matches or outperforms state-of-the-art accuracy, significantly boosting MXFP4, to the point where it nears that of NVFP4. We conclude that, while FP4 is not an automatic upgrade over INT4, format-specialized methods like MR-GPTQ can unlock a new frontier of accuracy-performance trade-offs.

Large language models (LLMs) have recently been applied to forecasting tasks, with some works claiming these systems match or exceed human performance. In this paper, we argue that, as a community, we should be careful about such conclusions as evaluating LLM forecasters presents unique challenges. We identify two broad categories of issues: (1) difficulty in trusting evaluation results due to many forms of temporal leakage, and (2) difficulty in extrapolating from evaluation performance to real-world forecasting. Through systematic analysis and concrete examples from prior work, we demonstrate how evaluation flaws can raise concerns about current and future performance claims. We argue that more rigorous evaluation methodologies are needed to confidently assess the forecasting abilities of LLMs.

Real-world language agents must handle complex, multi-step workflows across diverse applications. For instance, an agent may manage emails by coordinating with calendars and file systems, or monitor a production database like BigQuery to detect anomalies and generate reports following a standard operating manual. However, existing language agent benchmarks often focus on narrow domains or simplified tasks that lack the diversity, realism, and long-horizon complexity required to evaluate agents' real-world performance. To address this gap, we introduce the Tool Decathlon (dubbed as Toolathlon), a benchmark for language agents offering diverse applications and tools, realistic environment setup, and reliable execution-based evaluation. Toolathlon spans 32 software applications and 604 tools, ranging from everyday platforms such as Google Calendar and Notion to professional applications like WooCommerce, Kubernetes, and BigQuery. Most of the tools are based on a high-quality set of Model Context Protocol (MCP) servers that we may have revised or implemented ourselves. Unlike prior works, which primarily ensure functional realism but offer limited environment state diversity, we provide realistic initial environment states from real software, such as multiple Canvas courses each with dozens of students or real-world financial spreadsheets. The Toolathlon benchmark includes 108 manually sourced or crafted tasks in total, requiring interacting with multiple applications over ~20 interaction turns on average to complete. Each task is strictly verifiable through dedicated evaluation scripts. Comprehensive evaluation of state-of-the-art models highlights their significant shortcomings in performing real-world, long-horizon tasks: the best-performing model, Claude-4-Sonnet, achieves only a 29.9% success rate with 28 tool calling turns on average, while the top open-weights model DeepSeek-V3.1 reaches 13.9%. We expect Toolathlon to drive the development of more capable language agents for real-world, long-horizon task execution.