ICLR 2026 论文列表

Multimodal large language models (MLLMs) represent images and video frames as visual tokens. Scaling from single images to hour-long videos, however, inflates the token budget far beyond practical limits. Popular pipelines therefore either uniformly subsample or apply keyframe selection with retrieval-style scoring using smaller vision-language models. However, these keyframe selection methods still rely on pre-filtering before selection to reduce the inference cost and can miss the most informative moments. We propose FOCUS, Frame-Optimistic Confidence Upper-bound Selection, a training-free, model-agnostic keyframe selection module that selects query-relevant frames under a strict token budget. FOCUS formulates keyframe selection as a combinatorial pure-exploration (CPE) problem in multi-armed bandits: it treats short temporal clips as arms, and uses empirical means and Bernstein confidence radius to identify informative regions while preserving exploration of uncertain areas. The resulting two-stage exploration-exploitation procedure reduces from a sequential policy with theoretical guarantees, first identifying high-value temporal regions, then selecting top-scoring frames within each region. Extensive experiments across four long-video question-answering benchmarks and four popular MLLMs demonstrate that FOCUS delivers substantial accuracy improvements while processing less than 2% of video frames. For videos longer than 20 minutes, it achieves an 11.9% gain in accuracy on LongVideoBench, demonstrating its effectiveness as a keyframe selection method and providing a simple and general solution for scalable long-video understanding with MLLMs.

Training on Chain-of-Thought (CoT) traces has empirically shown to dramatically improve the capabilities of Large Language Models (LLMs), yet a formal understanding of its power remains limited. In this work, we investigate the role of training on such computational traces from the perspective of language learnability. We introduce a new learning model, identification in the limit with trace, which augments Gold's classic paradigm [Gold'67] by providing the learner not only with examples from a target language but also with computational traces from the machine that accepts them. Our results reveal that access to these traces dramatically enhances the power of the learner. We first prove that with perfect computational traces, the class of all computable languages (those recognizable by Turing Machines) becomes identifiable in the limit. This stands in sharp contrast to Gold's famous impossibility result, which holds even for the simple class of languages that are recognizable by deterministic finite automata. We then analyze the more challenging scenario where the learner has only partial information regarding the computational traces, which are also subject to adversarial corruptions. In this setting, we establish a set of trichotomic results on the amount of error that can be tolerated for the successful identification of language classes across the Chomsky hierarchy.

Autoregressive (AR) large language models (LLMs) have achieved remarkable performance across a wide range of natural language tasks, yet their inherent sequential decoding limits inference efficiency. In this work, we propose Fast-dLLM v2, a carefully designed block diffusion language model (dLLM) that efficiently adapts pretrained AR models into dLLMs for parallel text generation—requiring only ∼1B tokens of fine-tuning. This represents a 500× reduction in training data compared to full-attention diffusion LLMs such as Dream (580B tokens), while preserving the original model’s performance. Our approach introduces a novel training recipe that combines a block diffusion mechanism with a complementary attention mask, enabling blockwise bidirectional context modeling without sacrificing AR training objectives. To further accelerate decoding, we design a hierarchical caching mechanism: a block-level cache that stores historical context representations across blocks, and a sub-block cache that enables efficient parallel generation within partially decoded blocks. Coupled with our parallel decoding pipeline, Fast-dLLM v2 achieves up to 2.5× speedup over standard AR decoding without compromising generation quality. Extensive experiments across diverse benchmarks demonstrate that Fast-dLLM v2 matches or surpasses AR baselines in accuracy, while delivering state-of-the-art efficiency among dLLMs—marking a significant step toward the practical deployment of fast and accurate LLMs. Code and model will be publicly released.

In information retrieval, training reranking models mainly focuses on two types of objectives: metric learning (e.g. contrastive loss to increase the predicted scores on relevant query-document pairs) and classification (binary label prediction of relevance vs. irrelevance). For BERT-style encoders, various studies have shown that contrastive learning (CL) can be more effective than discriminative (classification) learning. However, for large language models (LLMs), classification via supervised fine-tuning (SFT), which predicts ''yes'' (resp. ''no'') token for relevant (resp. irrelevant) pairs, appears more promising as it aligns well with the generative nature of LLMs. This divergence raises a central question: which objective is intrinsically better suited to LLM-based reranking, and what mechanism underlies the difference? In this work, we conduct a comprehensive comparison and analysis between CL and SFT for reranking, taking the universal multimodal retrieval (UMR) as the experimental playground. We first decompose the objectives into two components: weight, which controls the magnitude of those updates, and direction, which guides the model updates, then present a unified framework for understanding their interactions. Through probing experiments, we find that SFT provides a substantially stronger weighting scheme than CL, whereas the preferred scoring direction shows no clear winner. Taken together, these results point to a consistent advantage of SFT over CL for LLM reranking. To further validate our findings, we conduct large-scale training with SFT and present new state-of-the-art rerankers on the MRB benchmark. We also provide ablations on SFT settings and expect our findings to benefit future research and applications in this area.

Recent studies have shown that long chain-of-thought (CoT) reasoning can significantly enhance the performance of large language models (LLMs) on complex tasks. However, this benefit is yet to be demonstrated in the domain of video understanding, since most existing benchmarks lack the reasoning depth required to demonstrate the advantages of extended CoT chains. While recent efforts have proposed benchmarks aimed at video reasoning, the tasks are often knowledge-driven and do not rely heavily on visual content. To bridge this gap, we introduce **VideoReasonBench**, a benchmark designed to evaluate **vision-centric, complex video reasoning**. To ensure visual richness and high reasoning complexity, each video in VideoReasonBench depicts a sequence of fine-grained operations on a latent state that is only visible in part of the video. The questions evaluate three escalating levels of video reasoning skills: recalling observed visual information, inferring the content of latent states, and predicting information beyond the video. Under such task setting, models have to precisely recall multiple operations in the video, and perform step-by-step reasoning to get correct final answers for these questions. Using VideoReasonBench, we comprehensively evaluate 18 state-of-the-art multimodal LLMs (MLLMs), finding that most perform poorly on complex video reasoning—e.g., GPT-4o achieves only 6.9\% accuracy—while the thinking-enhanced Gemini-2.5-Pro significantly outperforms others with 56.0% accuracy. Our investigations on "test-time scaling" further reveal that extended thinking budget, while offering none or minimal benefits on existing video benchmarks, is essential for improving the performance on VideoReasonBench.

Real-world temporal data often consists of multiple signal types recorded at irregular, asynchronous intervals. For instance, in the medical domain, different types of blood tests can be measured at different times and frequencies, resulting in fragmented and unevenly scattered temporal data. Similar issues of irregular sampling occur in other domains, such as the monitoring of large systems using event log files. Effectively learning from such data requires handling sets of temporally sparse and heterogeneous signals. In this work, we propose Graph Mixing Additive Networks (GMAN), a novel and interpretable-by-design framework for learning directly from sets of graphs that represent such signals. GMAN provides diverse interpretability capabilities, including node-level, graph-level, and subset-level importance, and enables practitioners to trade finer-grained interpretability for greater expressivity when domain priors are available. GMAN achieves state-of-the-art performance in real-world high-stakes tasks, including predicting Crohn’s disease onset and hospital length of stay from routine blood test measurements and detecting fake news. Furthermore, we demonstrate how GMAN’s interpretability properties assist in revealing disease development phase transitions and provide crucial insights in the healthcare domain.

Solving math problems through verifiable languages such as Lean has significantly impacted both the mathematics and computer science communities. Current state-of-the-art models are often trained with expensive online Reinforcement Learning (RL) or expert iteration. However, these approaches rely on fixed problem sets, which causes inefficient training and limits the model to tackle complex problems. To overcome these limitations, we propose **GAR**: *Generative Adversarial Reinforcement learning*, a comprehensive RL training framework that jointly trains the problem composer and solver in an adversarial loop. **GAR** introduces an implicit curriculum learning mechanism, which aligns task difficulty with the prover's evolving capability. It thereby improves the training efficiency and enables stronger performance of proving advanced theorems. Experiments show that with **GAR** training, Goedel-Prover-V2-8B and DeepSeek-Prover-V2-7B achieve an average relative improvement in pass@32 of **4.20%** on MiniF2F-Test benchmark, while DeepSeek-Prover-V2's pass@32 on ProofNet-Test increases from 22.58% to **25.81%**. Beyond formal proving, **GAR** establishes a general RL paradigm for co-evolution of problem generation and solving under verifiable environments.

Current unlearning methods for LLMs optimize on the private information they seek to remove by incorporating it into their fine-tuning data. We argue this not only risks reinforcing exposure to sensitive data, it also fundamentally contradicts the principle of minimizing its use. As a remedy, we propose a novel unlearning method—Partial Model Collapse (PMC), which does not require unlearning targets in the unlearning objective. Our approach is inspired by recent observations that training generative models on their own generations leads to distribution collapse, effectively removing information from model outputs. Our central insight is that model collapse can be leveraged for machine unlearning by deliberately triggering it for data we aim to remove. We theoretically analyze that our approach converges to the desired outcome, i.e. the model unlearns the data targeted for removal. We empirically demonstrate that PMC overcomes three key limitations of existing unlearning methods that explicitly optimize on unlearning targets, and more effectively removes private information from model outputs while preserving general model utility. Overall, our contributions represent an important step toward more comprehensive unlearning that aligns with real-world privacy constraints.

Rotary Position Embeddings (RoPE) are widely used in Transformers to encode positional information in token representations, yet the internal frequency structure of RoPE remains poorly understood. Previous studies have reported conflicting findings on the roles of high- and low-frequency dimensions, offering empirical observations but no unifying explanation. In this paper, we present a systematic framework that bridges these disparate results. We introduce Frequency Entropy (FE), a metric that quantifies the effective utilization of each RoPE frequency dimension, and we provide an analysis of how RoPE’s sinusoidal components contribute to model representations on a per-dimension basis. Based on an analysis of the Llama-4 model, which incorporates both RoPE and NoPE layers, we find that the periodicity captured by FE appears in RoPE layers but not in NoPE layers. Furthermore, FE identifies dimensions in which energy concentrates under RoPE. These characteristics are observed across the spectrum rather than being confined to specific dimensions. Moreover, attenuating extreme-entropy dimensions at inference yields downstream accuracy that is statistically indistinguishable from the baseline, with modest perplexity improvements on average, suggesting that such dimensions are often redundant. Overall, FE provides a simple, general diagnostic for RoPE with implications for analysis and design.

As large language models grow in capability and agency, identifying vulnerabilities through red-teaming becomes vital for safe deployment. However, traditional prompt-engineering approaches may prove ineffective once red-teaming turns into a \emph{weak-to-strong} problem, where target models surpass red-teamers in capabilities. To study this shift, we frame red-teaming through the lens of the \emph{capability gap} between attacker and target. We evaluate more than 600 attacker-target pairs using LLM-based jailbreak attacks that mimic human red-teamers across diverse families, sizes, and capability levels. Three strong trends emerge: (i) more capable models are better attackers, (ii) attack success drops sharply once the target’s capability exceeds the attacker's, and (iii) attack success rates correlate with high performance on social science splits of the MMLU-Pro benchmark. From these observations, we derive a \emph{jailbreaking scaling curve} that predicts attack success for a fixed target based on attacker-target capability gap. These findings suggest that fixed-capability attackers (e.g., humans) may become ineffective against future models, increasingly capable open-source models amplify risks for existing systems, and model providers must accurately measure and control models' persuasive and manipulative abilities to limit their effectiveness as attackers.

We study logical reasoning in language models by asking whether their errors follow established human fallacy patterns. Using the Erotetic Theory of Reasoning (ETR) and its open‑source implementation, PyETR, we programmatically generate 383 formally specified reasoning problems and evaluate 38 models. For each response, we judge logical correctness and, when incorrect, whether it matches an ETR‑predicted fallacy. Two results stand out: (i) as a capability proxy (Chatbot Arena Elo) increases, a larger share of a model’s incorrect answers are ETR‑predicted fallacies ($\rho=0.360, p=0.0265$), while overall correctness on this dataset shows no correlation with capability; (ii) reversing premise order significantly reduces fallacy production for many models, mirroring human order effects. Methodologically, PyETR provides an open‑source pipeline for unbounded, synthetic, contamination‑resistant reasoning tests linked to a cognitive theory, enabling analyses that focus on error composition rather than error rate.

Linear dynamical systems (LDSs) have been powerful tools for modeling high-dimensional time-series data across many domains, including neuroscience. However, a single LDS often struggles to capture the heterogeneity of neural data, where trajectories recorded under different conditions can have variations in their dynamics. Mixtures of linear dynamical systems (MoLDS) provide a path to model these variations in temporal dynamics for different observed trajectories. However, MoLDS remains difficult to apply in complex and noisy settings, limiting its practical use in neural data analysis. Tensor-based moment methods can provide global identifiability guarantees for MoLDS, but their performance degrades under realistic noise and complexity. Commonly used expectation-maximization (EM) methods offer flexibility in fitting latent models but are highly sensitive to initialization and prone to poor local minima. Here, we propose a tensor-based moment method that provides identifiability guarantees for learning MoLDS, which can be followed by EM updates to combine the strengths of both approaches. The novelty in our approach lies in the construction of moment tensors using the input–output data, on which we then apply Simultaneous Matrix Diagonalization (SMD) to recover globally consistent estimates of mixture weights and system parameters. These estimates can then be refined through a full Kalman filter-smoother EM algorithm, with closed-form updates for all LDS parameters. We validate our framework on synthetic benchmarks and real-world datasets. On synthetic data, the proposed Tensor-EM method achieves more reliable recovery and improved robustness compared to either pure tensor or randomly initialized EM methods. We then analyze neural recordings from the primate somatosensory cortex while a non-human primate performs reaches in different directions. In this setting, our method successfully models and clusters different conditions as separate subsystems, which is consistent with supervised single-LDS fits for each condition. Finally, we apply this approach to another neural dataset where monkeys perform a sequential reaching task. These results demonstrate that MoLDS provides an effective framework for modeling complex neural data in different brain regions, and that Tensor-EM is a principled and reliable approach to MoLDS learning for these applications.

Federated Learning (FL) was initially proposed as a privacy-preserving machine learning paradigm. However, FL has been shown to be susceptible to a series of privacy attacks. Recently, there has been concern about the Source Inference Attack (SIA), where an honest-but-curious central server attempts to identify exactly which client owns a given data point which was used in the training phase. Alarmingly, standard gradient obfuscation techniques with Differential Privacy have been shown to be ineffective against SIAs, at least without severely diminishing the accuracy. In this work, we propose a defense against SIAs within the widely studied shuffle model of FL, where an honest shuffler acts as an intermediary between the clients and the server. First, we demonstrate that standard naive shuffling alone is insufficient to prevent SIAs. To effectively defend against SIAs, shuffling needs to be applied at a more granular level; we propose a novel combination of parameter-level shuffling with the residue number system (RNS). Our approach provides robust protection against SIAs without affecting the accuracy of the joint model and can be seamlessly integrated into other privacy protection mechanisms. We conduct experiments on a series of models and datasets, confirming that standard shuffling approaches fail to prevent SIAs and that, in contrast, our proposed method reduce the attack’s accuracy to the level of random guessing.

Machine learning (ML)-based malware detection systems often fail to account for the dynamic nature of real-world training and test data distributions. In practice, these distributions evolve due to frequent changes in the Android ecosystem, adversarial development of new malware families, and the continuous emergence of both benign and malicious applications. Prior studies have shown that such concept drift—distributional shifts in benign and malicious samples, leads to significant degradation in detection performance over time. Despite the practical importance of this issue, existing datasets are often outdated and limited in temporal scope, diversity of malware families, and sample scale, making them insufficient for the systematic evaluation of concept drift in malware detection. To address this gap, we present LAMDA, the largest and most temporally diverse Android malware benchmark to date, designed specifically for concept drift analysis. LAMDA spans 12 years (2013–2025, excluding 2015), includes over 1 million samples (approximately 37\% labeled as malware), and covers 1,380 malware families and 150,000 singleton samples, reflecting the natural distribution and evolution of real-world Android applications. We empirically demonstrate LAMDA's utility by quantifying the performance degradation of standard ML models over time and analyzing feature stability across years. As the most comprehensive Android malware dataset to date, LAMDA enables in-depth research into temporal drift, generalization, explainability, and evolving detection challenges.

Modifications to test-time sampling have emerged as an important extension to diffusion algorithms, with the goal of biasing the generative process to achieve a given objective without having to retrain the entire diffusion model. However, generating jointly correlated samples from multiple pre-trained diffusion models while simultaneously enforcing task-specific constraints without costly retraining has remained challenging. To this end, we propose Projected Coupled Diffusion (PCD), a novel test-time framework for constrained joint generation. PCD introduces a coupled guidance term into the generative dynamics to encourage coordination between diffusion models and incorporates a projection step at each diffusion step to enforce hard constraints. Empirically, we demonstrate the effectiveness of PCD in application scenarios of image-pair generation, object manipulation, and multi-robot motion planning. Our results show improved coupling effects and guaranteed constraint satisfaction without incurring excessive computational costs.

Tensors are a fundamental data structure for many scientific contexts, such as time series analysis, materials science, and physics, among many others. Improving our ability to produce and handle tensors is essential to efficiently address problems in these domains. In this paper, we show how to exploit the underlying symmetries of functions that map tensors to tensors. More concretely, we develop universally expressive equivariant machine learning architectures on tensors that exploit that, in many cases, these tensor functions are equivariant with respect to the diagonal action of the orthogonal, Lorentz, and/or symplectic groups. We showcase our results on three problems coming from material science, theoretical computer science, and time series analysis. For time series, we combine our method with the increasingly popular path signatures approach, which is also invariant with respect to reparameterizations. Our numerical experiments show that our equivariant models perform better than corresponding non-equivariant baselines.

Dynamic Path Planning (DPP) in urban road networks faces fundamental challenges, as traffic conditions change rapidly over time and often render planned routes ineffective. Reinforcement Learning (RL) provides an effective way to adaptively handle such uncertainties by incorporating traffic dynamics into state, but its performance crucially depends on how these dynamics are represented. Existing approaches either rely on global traffic information, which ensures decision completeness but suffers from redundancy and high computational cost, or oversimplified local features, which are efficient but often omit critical dynamics and lead to suboptimal paths. To address this, we propose a Dynamics Feature Representation (DFR) framework that progressively refines global traffic dynamics into compact features for RL-based DPP. Specifically, we introduce a policy attention mechanism that identifies a core subset of global dynamics by extracting the top-k shortest paths, and further constructs node-related local features by intersecting with n-hop neighborhoods, enabling near-optimal policy learning. Theoretical analysis demonstrates that DFR guarantees state completeness, while empirical results confirm that, compared to classical baselines and standard RL methods, DFR significantly improves path planning performance and accelerates convergence. This work highlights the central role of feature representation in RL-based DPP and proposes a general framework that balances information sufficiency with computational efficiency, paving the way for scalable dynamic decision-making in real-world transportation systems.

We introduce a new, flexible, and theoretically justified framework for dimension reduction in high-dimensional regression, based on an ensemble of random projections. Specifically, we consider disjoint groups of independent random projections, retain the best projection in each group according to the empirical regression performance on the projected covariates, and then aggregate the selected projections via singular value decomposition. The singular values quantify the relative importance of corresponding projection directions and guide the dimension selection process. We investigate various aspects of our framework, including the choice of projection distribution and the number of projections used. Our theoretical results show that the expected estimation error decreases as the number of groups of projections increases. Finally, we demonstrate that our proposal consistently matches or outperforms state-of-the-art methods through extensive numerical studies on simulated and real data.

Large Reasoning Models (LRMs) with long chain-of-thought capabilities, optimized via reinforcement learning with verifiable rewards (RLVR), excel at **objective reasoning** tasks like mathematical problem solving and code generation. However, RLVR is known for degrading generation diversity, which causes LRMs to fall short on **subjective reasoning** that has multiple answers depending on different role perspectives. While recent studies recognize the importance of diversity-enhanced training in objective reasoning, limited attention has been given to subjective tasks. In this paper, we find that subjective reasoning can be improved by introducing perspective diversity and token-level diversity, with the former one providing a coherent scaffolding anchored to a real-world stakeholder group and the latter one broadening the answer search space. We propose MultiRole-R1, a diversity-enhanced training framework featuring an unsupervised data construction pipeline that synthesizes reasoning chains incorporating various role perspectives. It also employs reinforcement learning via Group Relative Policy Optimization with reward shaping, taking diversity as a reward signal in addition to verifiable reward. Training on subjective tasks solely, MultiRole-R1 increases the in-domain and out-of-domain accuracy by 14.1% and 7.64%, and even enhances the performance on advanced math reasoning such as AIME 2024. We further show that diversity is a more consistent indicator of accuracy than reasoning length.