ICLR 2026 Papers

Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet optimally integrating Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through a comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from an entropy-based perspective, we reveal key differences between these paradigms: SFT induces coarse-grained, global shifts to policy distributions, while RL performs fine-grained, selective optimizations. Our analysis further establishes entropy as a critical indicator of training efficacy. Building on these observations, we introduce **S**upervised **R**einforcement **F**ine-**T**uning (**SRFT**), a single-stage framework that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. SRFT simultaneously applies SFT and RL to directly optimize LLMs using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT outperforms zero-RL baselines by **9.0%** on five mathematical reasoning benchmarks and by **10.9%** on three out-of-distribution benchmarks. Moreover, by leveraging demonstration data, SRFT maintains a more stable policy entropy, facilitating sustained policy improvement.

Noise is ubiquitous in data used to train large language models, but it is not well understood whether these models are able to correctly generalize to inputs generated without noise. Here, we study noise-robust learning: are transformers trained on data with noisy features able to find a target function that correctly predicts labels for noiseless features? We show that transformers succeed at noise-robust learning for a selection of $k$-sparse parity and majority functions, compared to LSTMs which fail at this task for even modest feature noise. However, we find that transformers typically fail at noise-robust learning of random $k$-juntas, especially when the boolean sensitivity of the optimal solution is smaller than that of the target function. We argue that this failure is due to a combination of two factors: transformers' bias toward simpler functions, combined with an observation that the empirically optimal function for noise-robust learning has lower sensitivity than the target function. We test this hypothesis by exploiting transformers' simplicity bias to trap them in an incorrect solution, but show that transformers can escape this trap by training with an additional loss term penalizing high-sensitivity solutions. Overall, we find that transformers are particularly ineffective for learning boolean functions in the presence of feature noise.

Semivalue–based data valuation uses cooperative‐game theory intuitions to assign each data point a value reflecting its contribution to a downstream task. Still, those values depend on the practitioner’s choice of utility, raising the question: *How robust is semivalue-based data valuation to changes in the utility?* This issue is critical when the utility is set as a trade‐off between several criteria and when practitioners must select among multiple equally valid utilities. We address it by introducing the notion of a dataset’s *spatial signature*: given a semivalue, we embed each data point into a lower-dimensional space where any utility becomes a linear functional, making the data valuation framework amenable to a simpler geometric picture. Building on this, we propose a practical methodology centered on an explicit robustness metric that informs practitioners whether and by how much their data valuation results will shift as the utility changes. We validate this approach across diverse datasets and semivalues, demonstrating strong agreement with rank‐correlation analyses and offering analytical insight into how choosing a semivalue can amplify or diminish robustness.

Mobile GUI agents are becoming critical tools to improve user experience on smart devices, with multimodal large language models (MLLMs) emerging as the dominant paradigms in this domain. Current agents, however, rely on explicit human instructions, overlooking the potential to leverage the contextual information (like location, time, user profile) and historical data for proactive task suggestions. Besides, previous works focus on optimizing the success rate during task execution, but pay less attention to the personalized execution trajectory, thereby neglecting potentially vast differences in user preferences. To address these challenges, we introduce the FingerTip 20K benchmark. We collected 20K unique human demonstrations of multi-step Android device interactions across a variety of everyday apps. These demonstrations are not isolated but are continuously acquired from the users' long-term usage in their real lives, and encompass essential user-related contextual information. The benchmark contains two new tracks: proactive task suggestions by analyzing environment observation and users' previous intents, and personalized task execution by catering to users' action preferences. Our experiments reveal that the tracks we propose pose significant challenges for leveraging user-related information in GUI tasks. We also performed a human study to show that there exists a huge gap between existing agents and humans. The model fine-tuned with the data we collected effectively utilized user information and achieved good results, highlighting the potential of our approach in building more user-oriented mobile LLM agents. Our code is open-source at \url{https://anonymous.4open.science/r/FingerTip-57B8} for reproducibility.

Graph-based Retrieval-Augmented Generation (RAG) systems leverage interconnected knowledge structures to capture complex relationships that flat retrieval struggles with, enabling multi-hop reasoning. Yet most existing graph-based methods suffer from (i) heuristic designs lacking theoretical guarantees for subgraph quality or relevance and/or (ii) the use of static exploration strategies that ignore the query's holistic meaning, retrieving neighborhoods or communities regardless of intent. We propose *Query-Aware Flow Diffusion RAG* (QAFD-RAG), a training-free framework that dynamically adapts graph traversal to each query's holistic semantics. The central innovation is *query-aware traversal*: during graph exploration, edges are dynamically weighted by how well their endpoints align with the query's embedding, guiding flow along semantically relevant paths while avoiding structurally connected but irrelevant regions. These query-specific reasoning subgraphs enable the first statistical guarantees for query-aware graph retrieval, showing that QAFD-RAG recovers relevant subgraphs with high probability under mild signal-to-noise conditions. The algorithm converges exponentially fast, with complexity scaling with the retrieved subgraph size rather than the full graph. Experiments on question answering and text-to-SQL tasks demonstrate consistent improvements over state-of-the-art graph-based RAG methods.

Current language models rely on static vocabularies determined at pretraining time, which can lead to decreased performance and increased computational cost for domains underrepresented in the original vocabulary. New tokens can be added to solve this problem, when coupled with a good initialization for their new embeddings. However, existing embedding initialization methods require expensive further training or pretraining of additional modules. In this paper, we propose Token Distillation and show that by distilling representations obtained using the original tokenization, we can quickly learn high-quality input embeddings for new tokens. Experimental results with a wide range of open-weight models show that Token Distillation outperforms even strong baselines.

Federated Prompt Learning has emerged as a communication-efficient and privacy-preserving paradigm for adapting large vision-language models like CLIP across decentralized clients. However, the security implications of this setup remain underexplored. In this work, we present the first study of backdoor attacks in Federated Prompt Learning. We show that when malicious clients inject visually imperceptible, learnable noise triggers into input images, the global prompt learner becomes vulnerable to targeted misclassification while still maintaining high accuracy on clean inputs. Motivated by this vulnerability, we propose SABRE-FL, a lightweight, modular defense that filters poisoned prompt updates using an embedding-space anomaly detector trained offline on out-of-distribution data. SABRE-FL requires no access to raw client data or labels and generalizes across diverse datasets. We show, both theoretically and empirically, that malicious clients can be reliably identified and filtered using an embedding-based detector. Across five diverse datasets and four baseline defenses, SABRE-FL outperforms all baselines by significantly reducing backdoor accuracy while preserving clean accuracy, demonstrating strong empirical performance and underscoring the need for robust prompt learning in future federated systems.

Agents in the real world must make not only logical but also *timely* judgments. This requires continuous awareness of the dynamic environment: hazards emerge, opportunities arise, and other agents act, while the agent's reasoning is still unfolding. Despite advances in language model reasoning, existing approaches fail to account for this dynamic nature. We introduce *real-time reasoning* as a new problem formulation for agents in evolving environments and build **Real-Time Reasoning Gym** to demonstrate it. We study two paradigms for deploying language models in agents: (1) **Reactive agents**, which employ language models with *bounded reasoning computation for rapid responses*. (2) **Planning agents**, which allow *extended reasoning computation for complex problems*. Our experiments show that even state-of-the-art models struggle with making logical and timely judgments in either paradigm. To address this limitation, we propose **AgileThinker**, which simultaneously engages *both reasoning paradigms*. **AgileThinker** consistently outperforms agents engaging only one reasoning paradigm as task difficulty and time pressure rise, effectively balancing reasoning depth and response latency. Our work establishes real-time reasoning as a critical testbed for developing practical agents and provides a foundation for research in temporally constrained AI systems, highlighting a path toward real-time capable agents.

Instruction-guided video editing has emerged as a rapidly advancing research direction, offering new opportunities for intuitive content transformation while also posing significant challenges for systematic evaluation. Existing video editing benchmarks fail to support the evaluation of instruction-guided video editing adequately and further suffer from limited source diversity, narrow task coverage and incomplete evaluation metrics. To address above limitations, we introduce IVEBench, a modern benchmark suite specifically designed for instruction-guided video editing assessment. IVEBench comprises a diverse database of 600 high-quality source videos, spanning seven semantic dimensions, and covering video lengths ranging from 32 to 1,024 frames. It further includes 8 categories of editing tasks with 35 subcategories, whose prompts are generated and refined through large language models and expert review. Crucially, IVEBench establishes a three-dimensional evaluation protocol encompassing video quality, instruction compliance and video fidelity, integrating both traditional metrics and multimodal large language model-based assessments. Extensive experiments demonstrate the effectiveness of IVEBench in benchmarking state-of-the-art instruction-guided video editing methods, showing its ability to provide comprehensive and human-aligned evaluation outcomes. All data and code will be made publicly available.

Interactive segmentation uses real-time user inputs, such as mouse clicks, to iteratively refine model predictions. Although not originally designed to address distribution shifts, this paradigm naturally lends itself to such challenges. In medical imaging, where distribution shifts are common, interactive methods can use user inputs to guide models towards improved predictions. Moreover, once a model is deployed, user corrections can be used to adapt the network parameters to the new data distribution, mitigating distribution shift. Based on these insights, we aim to develop a practical, effective method for improving the adaptive capabilities of interactive segmentation models to new data distributions in medical imaging. Firstly, we found that strengthening the model's responsiveness to clicks is important for the initial training process. Moreover, we show that by treating the post-interaction user-refined model output as pseudo-ground-truth, we can design a lean, practical online adaptation method that enables a model to learn effectively across sequential test images. The framework includes two components: (i) a Post-Interaction adaptation process, updating the model after the user has completed interactive refinement of an image, and (ii) a Mid-Interaction adaptation process, updating incrementally after each click. Both processes include a Click-Centered Gaussian loss that strengthens the model's reaction to clicks and enhances focus on user-guided, clinically relevant regions. Experiments on 5 fundus and 4 brain‑MRI databases show that our approach consistently outperforms existing methods under diverse distribution shifts, including unseen imaging modalities and pathologies. Code and pretrained models will be released upon publication.

Model-based reinforcement learning promises strong sample efficiency but often underperforms in practice due to compounding model error, unimodal world models that average over multi-modal dynamics, and overconfident predictions that bias learning. We introduce WIMLE, a model-based method that extends Implicit Maximum Likelihood Estimation (IMLE) to the model-based RL framework to learn stochastic, multi-modal world models without iterative sampling and to estimate predictive uncertainty via ensembles and latent sampling. During training, WIMLE weights each synthetic transition by its predicted confidence, preserving useful model rollouts while attenuating bias from uncertain predictions and enabling stable learning. Across 40 continuous-control tasks spanning DeepMind Control, MyoSuite, and HumanoidBench, WIMLE achieves superior sample efficiency and competitive or better asymptotic performance than strong model-free and model-based baselines. Notably, on the challenging Humanoid-run task, WIMLE improves sample efficiency by over 50\% relative to the strongest competitor, and on HumanoidBench it solves 8 of 14 tasks (versus 4 for BRO and 5 for SimbaV2). These results highlight the value of IMLE-based multi-modality and uncertainty-aware weighting for stable model-based RL.

As one of the fundamental intermediate modules in autonomous driving, online high-definition (HD) maps have attracted significant attention due to their cost-effectiveness and real-time capabilities. Since vehicles always cruise in highly dynamic environments, spatial displacement of onboard sensors inevitably causes shifts in real-time HD mapping results, and such instability poses fundamental challenges for downstream tasks. However, existing online map construction models tend to prioritize improving each frame's mapping accuracy, while the mapping stability has not yet been systematically studied. To fill this gap, this paper presents the first comprehensive benchmark for evaluating the temporal stability of online HD mapping models. We propose a multi-dimensional stability evaluation framework with novel metrics for Presence, Localization, and Shape Stability, integrated into a unified mean Average Stability (mAS) score. Extensive experiments on 42 models and variants show that accuracy (mAP) and stability (mAS) represent largely independent performance dimensions. We further analyze the impact of key model design choices on both criteria, identifying architectural and training factors that contribute to high accuracy, high stability, or both. To encourage broader focus on stability, we will release a public benchmark. Our work highlights the importance of treating temporal stability as a core evaluation criterion alongside accuracy, advancing the development of more reliable autonomous driving systems. The benchmark toolkit, code, and models will be available at \hyperlink{https://stablehdmap.github.io/}{https://stablehdmap.github.io/}.

In this work, we take a first step toward elucidating the mechanisms behind emergent exploration in unsupervised reinforcement learning. We study Single-Goal Contrastive Reinforcement Learning (SGCRL) (Liu et al., 2025), a self-supervised algorithm capable of solving challenging long-horizon goal-reaching tasks without external rewards or curricula. We combine theoretical analysis of the algorithm’s objective function with controlled experiments to understand what drives its exploration. We show that SGCRL maximizes implicit rewards shaped by its learned representations. These representations automatically modify the reward landscape to promote exploration before reaching the goal and exploitation thereafter. Our experiments also demonstrate that these exploration dynamics arise from learning low-rank representations of the state space rather than from neural network function approximation. Our improved understanding enables us to adapt SGCRL to perform safety-aware exploration.

Reinforcement post training (RPT) has recently shown promise in improving the reasoning abilities of large language models (LLMs). However, it remains unclear how well these improvements generalize to new domains, as prior work evaluates RPT models on data from the same domains used for fine-tuning. To understand the generalizability of RPT, we conduct two studies. (1) Observational: we compare a wide range of open-weight RPT models against their corresponding base models across multiple domains, including both seen and unseen domains in their fine-tuning data. (2) Interventional: we fine-tune LLMs with RPT on single domains and evaluate their performance across multiple domains. Both studies converge on the same conclusion that, although RPT brings substantial gains on tasks similar to the fine-tuning data, the gains generalize inconsistently and can vanish on domains with different reasoning patterns.

Articulated objects are ubiquitous in daily life. Our goal is to achieve a high-quality reconstruction, segmentation of independent moving parts, and analysis of articulation. Recent methods analyse two different articulation states and perform per-point part segmentation, optimising per-part articulation using cross-state correspondences, given a priori knowledge of the number of parts. Such assumptions greatly limit their applications and performance. Their robustness is reduced when objects cannot be clearly visible in both states. To address these issues, in this paper, we present a new framework, *Articulation in Motion (AiM)*. We infer part-level decomposition, articulation kinematics, and reconstruct an interactive 3D digital replica from a user–object interaction video and a start-state scan. We propose a dual-Gaussian scene representation that is learned from an initial 3DGS scan of the object and a video that shows the movement of separate parts. It uses motion cues to segment the object into parts and assign articulation joints. Subsequently, a robust, sequential RANSAC is employed to achieve part mobility analysis \textit{without any part-level structural priors}, which clusters moving primitives into rigid parts and estimates kinematics while automatically determining the number of parts. The proposed approach separates the object into parts, each represented as a 3D Gaussian set, enabling high-quality rendering. Our approach yields higher quality part segmentation than previous methods, without prior knowledge. Extensive experimental analysis on both simple and complex objects validates the effectiveness and strong generalisation ability of our approach. Project page: https://haoai-1997.github.io/AiM/.

Recent advances in video generation have enabled thrilling experiences in producing realistic videos driven by scalable diffusion transformers. However, they usually fail to produce satisfactory outputs that are aligned to users' authentic demands and preferences. In this work, we introduce Dual-Iterative Optimization (Dual-IPO), an iterative paradigm that sequentially optimizes both the reward model and the video generation model for improved synthesis quality and human preference alignment. For the reward model, our framework ensures reliable and robust reward signals via CoT-guided reasoning, voting-based self-consistency, and preference certainty estimation. Given this, we optimize video foundation models with guidance of signals from reward model's feedback, thus improving the synthesis quality in subject consistency, motion smoothness and aesthetic quality, etc. The reward model and video generation model complement each other and are progressively improved in the multi-round iteration, without requiring tediously manual preference annotations. Comprehensive experiments demonstrate that the proposed Dual-IPO can effectively and consistently improve the video generation quality of base model with various architectures and sizes, even help a model with only 2B parameters surpass a 5B one. Moreover, our analysis experiments and ablation studies identify the rational of our systematic design and the efficacy of each component.

Reward modeling is crucial for aligning large language models with human preferences, yet current approaches lack a principled mathematical framework for leveraging ordinal preference data. When human annotators provide graded preferences on a Likert scale (e.g., significantly better, better, slightly better, negligibly better), existing methods typically apply ad-hoc heuristics, such as margin terms or scaling factors, to loss functions derived from binary preference models like Bradley-Terry. These approaches lack an underlying mathematical model for how ordinal preference data is generated. We present a theoretically grounded framework that formulates reward modeling with Likert scale preferences as a discrete ordinal regression problem. We derive two loss functions from this formulation: a negative log-likelihood loss and an all-threshold loss, both of which learn threshold parameters that naturally capture the ordinal structure of preferences. Unlike existing heuristic methods that manually specify fixed margins or scaling weights, our approach learns these parameters directly from data within a coherent probabilistic framework. Experimental results on multiple benchmarks demonstrate that our ordinal regression approach consistently achieves competitive or superior performance compared to existing heuristic methods across diverse evaluation categories including chat, reasoning, and safety tasks. Our work provides the first principled mathematical framework for incorporating Likert scale preferences into reward model training, moving beyond ad-hoc modifications of binary preference models to enable more effective utilization of fine-grained human feedback.

Fast and high-quality language generation is the holy grail that people pursue in the age of AI. In this work, we introduce **Di**screte **Di**ffusion Divergence **Instruct** (**DiDi-Instruct**), a training-based method that initializes from a pre-trained diffusion large language model (dLLM) and distills a few-step student for fast generation. The model distilled with DiDi-Instruct matches or surpasses its dLLM teacher and the GPT-2 baseline while providing up to **64$\times$ acceleration**. The theoretical foundation of DiDi-Instruct is a novel framework based on integral KL-divergence minimization, which leads to a practical training algorithm. We further introduce *grouped reward normalization, intermediate-state matching, and the reward-guided ancestral sampler* to improve *training stability, model coverage, and inference quality*. On the OpenWebText benchmark, DiDi-Instruct achieves perplexity ranging from 62.2 (8 NFEs) to 18.4 (128 NFEs), outperforming prior accelerated dLLMs and the GPT-2 baseline. These gains incur a negligible entropy loss (around $1$%) and reduce additional training wall-clock time by **more than $20\times$** compared to competing dLLM distillation methods. We further validate the robustness and effectiveness of DiDi-Instruct through extensive ablation studies, model scaling, downstream task evaluations, and unconditional protein sequence generation. In conclusion, DiDi-Instruct enables efficient and effective distillation for language generation in the blink of an eye.

Synthetic image generators evolve rapidly, challenging detectors to generalize across current methods and adapt to new ones. We study domain-incremental synthetic image detection with a two-phase evaluation. Phase I trains on either diffusion- or GAN-based data and tests on the combined group to quantify bidirectional cross-generator transfer. Phase II sequentially introduces renders from 3D Gaussian Splatting (3DGS) head avatar pipelines, requiring adaptation while preserving earlier performance. We observe that CLIP-based detectors inherit text-image alignment semantics that are irrelevant to authenticity and hinder generalization. We introduce a Hilbert-Schmidt Independence Criterion (HSIC) bottleneck loss on intermediate CLIP ViT features, encouraging representations predictive of real versus synthetic while independent of generator identity and caption alignment. For domain-incremental learning, we propose HSIC-Guided Replay (HGR), which selects per-class exemplars via a hybrid score combining HSIC relevance with k-center coverage, yielding compact memories that mitigate forgetting. Empirically, the HSIC bottleneck improves transfer between diffusion and GAN families, and HGR sustains prior accuracy while adapting to 3DGS renders. These results underscore the value of information-theoretic feature shaping and principled replay for resilient detection under shifting generative regimes.